diff --git a/tutorials/27_First_RAG_Pipeline.ipynb b/tutorials/27_First_RAG_Pipeline.ipynb index d9467ff3..d8a4fc49 100644 --- a/tutorials/27_First_RAG_Pipeline.ipynb +++ b/tutorials/27_First_RAG_Pipeline.ipynb @@ -1,1446 +1,1339 @@ { - "cells": [ - { - "cell_type": "markdown", - "metadata": { - "id": "2OvkPji9O-qX" - }, - "source": [ - "# Tutorial: Creating Your First QA Pipeline with Retrieval-Augmentation\n", - "\n", - "- **Level**: Beginner\n", - "- **Time to complete**: 10 minutes\n", - "- **Components Used**: [`InMemoryDocumentStore`](https://docs.haystack.deepset.ai/docs/inmemorydocumentstore), [`SentenceTransformersDocumentEmbedder`](https://docs.haystack.deepset.ai/docs/sentencetransformersdocumentembedder), [`SentenceTransformersTextEmbedder`](https://docs.haystack.deepset.ai/docs/sentencetransformerstextembedder), [`InMemoryEmbeddingRetriever`](https://docs.haystack.deepset.ai/docs/inmemoryembeddingretriever), [`PromptBuilder`](https://docs.haystack.deepset.ai/docs/promptbuilder), [`OpenAIGenerator`](https://docs.haystack.deepset.ai/docs/openaigenerator)\n", - "- **Prerequisites**: You must have an [OpenAI API Key](https://platform.openai.com/api-keys).\n", - "- **Goal**: After completing this tutorial, you'll have learned the new prompt syntax and how to use PromptBuilder and OpenAIGenerator to build a generative question-answering pipeline with retrieval-augmentation.\n", - "\n", - "> This tutorial uses Haystack 2.0. To learn more, read the [Haystack 2.0 announcement](https://haystack.deepset.ai/blog/haystack-2-release) or visit the [Haystack 2.0 Documentation](https://docs.haystack.deepset.ai/docs/intro)." - ] - }, - { - "cell_type": "markdown", - "metadata": { - "id": "LFqHcXYPO-qZ" - }, - "source": [ - "## Overview\n", - "\n", - "This tutorial shows you how to create a generative question-answering pipeline using the retrieval-augmentation ([RAG](https://www.deepset.ai/blog/llms-retrieval-augmentation)) approach with Haystack 2.0. The process involves four main components: [SentenceTransformersTextEmbedder](https://docs.haystack.deepset.ai/docs/sentencetransformerstextembedder) for creating an embedding for the user query, [InMemoryBM25Retriever](https://docs.haystack.deepset.ai/docs/inmemorybm25retriever) for fetching relevant documents, [PromptBuilder](https://docs.haystack.deepset.ai/docs/promptbuilder) for creating a template prompt, and [OpenAIGenerator](https://docs.haystack.deepset.ai/docs/openaigenerator) for generating responses.\n", - "\n", - "For this tutorial, you'll use the Wikipedia pages of [Seven Wonders of the Ancient World](https://en.wikipedia.org/wiki/Wonders_of_the_World) as Documents, but you can replace them with any text you want.\n" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "id": "QXjVlbPiO-qZ" - }, - "source": [ - "## Preparing the Colab Environment\n", - "\n", - "- [Enable GPU Runtime in Colab](https://docs.haystack.deepset.ai/docs/enabling-gpu-acceleration)\n", - "- [Set logging level to INFO](https://docs.haystack.deepset.ai/docs/logging)" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "id": "Kww5B_vXO-qZ" - }, - "source": [ - "## Installing Haystack\n", - "\n", - "Install Haystack 2.0 and other required packages with `pip`:" - ] - }, - { - "cell_type": "code", - "execution_count": 1, - "metadata": { - "colab": { - "base_uri": "https://localhost:8080/" - }, - "id": "UQbU8GUfO-qZ", - "outputId": "c33579e9-5557-43bd-a3c5-63b8373770c7" - }, - "outputs": [ - { - "name": "stdout", - "output_type": "stream", - "text": [ - "Requirement already satisfied: haystack-ai in /usr/local/lib/python3.10/dist-packages (2.0.0b8)\n", - "Requirement already satisfied: boilerpy3 in /usr/local/lib/python3.10/dist-packages (from haystack-ai) (1.0.7)\n", - "Requirement already satisfied: haystack-bm25 in /usr/local/lib/python3.10/dist-packages (from haystack-ai) (1.0.2)\n", - "Requirement already satisfied: jinja2 in /usr/local/lib/python3.10/dist-packages (from haystack-ai) (3.1.3)\n", - "Requirement already satisfied: lazy-imports in /usr/local/lib/python3.10/dist-packages (from haystack-ai) (0.3.1)\n", - "Requirement already satisfied: more-itertools in /usr/local/lib/python3.10/dist-packages (from haystack-ai) (10.1.0)\n", - "Requirement already satisfied: networkx in /usr/local/lib/python3.10/dist-packages (from haystack-ai) (3.2.1)\n", - "Requirement already satisfied: openai>=1.1.0 in /usr/local/lib/python3.10/dist-packages (from haystack-ai) (1.13.3)\n", - "Requirement already satisfied: pandas in /usr/local/lib/python3.10/dist-packages (from haystack-ai) (1.5.3)\n", - "Requirement already satisfied: posthog in /usr/local/lib/python3.10/dist-packages (from haystack-ai) (3.5.0)\n", - "Requirement already satisfied: pyyaml in /usr/local/lib/python3.10/dist-packages (from haystack-ai) (6.0.1)\n", - "Requirement already satisfied: tenacity in /usr/local/lib/python3.10/dist-packages (from haystack-ai) (8.2.3)\n", - "Requirement already satisfied: tqdm in /usr/local/lib/python3.10/dist-packages (from haystack-ai) (4.66.2)\n", - "Requirement already satisfied: typing-extensions in /usr/local/lib/python3.10/dist-packages (from haystack-ai) (4.10.0)\n", - "Requirement already satisfied: anyio<5,>=3.5.0 in /usr/local/lib/python3.10/dist-packages (from openai>=1.1.0->haystack-ai) (3.7.1)\n", - "Requirement already satisfied: distro<2,>=1.7.0 in /usr/lib/python3/dist-packages (from openai>=1.1.0->haystack-ai) (1.7.0)\n", - "Requirement already satisfied: httpx<1,>=0.23.0 in /usr/local/lib/python3.10/dist-packages (from openai>=1.1.0->haystack-ai) (0.27.0)\n", - "Requirement already satisfied: pydantic<3,>=1.9.0 in /usr/local/lib/python3.10/dist-packages (from openai>=1.1.0->haystack-ai) (2.6.3)\n", - "Requirement already satisfied: sniffio in /usr/local/lib/python3.10/dist-packages (from openai>=1.1.0->haystack-ai) (1.3.1)\n", - "Requirement already satisfied: numpy in /usr/local/lib/python3.10/dist-packages (from haystack-bm25->haystack-ai) (1.25.2)\n", - "Requirement already satisfied: MarkupSafe>=2.0 in /usr/local/lib/python3.10/dist-packages (from jinja2->haystack-ai) (2.1.5)\n", - "Requirement already satisfied: python-dateutil>=2.8.1 in /usr/local/lib/python3.10/dist-packages (from pandas->haystack-ai) (2.8.2)\n", - "Requirement already satisfied: pytz>=2020.1 in /usr/local/lib/python3.10/dist-packages (from pandas->haystack-ai) (2023.4)\n", - "Requirement already satisfied: requests<3.0,>=2.7 in /usr/local/lib/python3.10/dist-packages (from posthog->haystack-ai) (2.31.0)\n", - "Requirement already satisfied: six>=1.5 in /usr/local/lib/python3.10/dist-packages (from posthog->haystack-ai) (1.16.0)\n", - "Requirement already satisfied: monotonic>=1.5 in /usr/local/lib/python3.10/dist-packages (from posthog->haystack-ai) (1.6)\n", - "Requirement already satisfied: backoff>=1.10.0 in /usr/local/lib/python3.10/dist-packages (from posthog->haystack-ai) (2.2.1)\n", - "Requirement already satisfied: idna>=2.8 in /usr/local/lib/python3.10/dist-packages (from anyio<5,>=3.5.0->openai>=1.1.0->haystack-ai) (3.6)\n", - "Requirement already satisfied: exceptiongroup in /usr/local/lib/python3.10/dist-packages (from anyio<5,>=3.5.0->openai>=1.1.0->haystack-ai) (1.2.0)\n", - "Requirement already satisfied: certifi in /usr/local/lib/python3.10/dist-packages (from httpx<1,>=0.23.0->openai>=1.1.0->haystack-ai) (2024.2.2)\n", - "Requirement already satisfied: httpcore==1.* in /usr/local/lib/python3.10/dist-packages (from httpx<1,>=0.23.0->openai>=1.1.0->haystack-ai) (1.0.4)\n", - "Requirement already satisfied: h11<0.15,>=0.13 in /usr/local/lib/python3.10/dist-packages (from httpcore==1.*->httpx<1,>=0.23.0->openai>=1.1.0->haystack-ai) (0.14.0)\n", - "Requirement already satisfied: annotated-types>=0.4.0 in /usr/local/lib/python3.10/dist-packages (from pydantic<3,>=1.9.0->openai>=1.1.0->haystack-ai) (0.6.0)\n", - "Requirement already satisfied: pydantic-core==2.16.3 in /usr/local/lib/python3.10/dist-packages (from pydantic<3,>=1.9.0->openai>=1.1.0->haystack-ai) (2.16.3)\n", - "Requirement already satisfied: charset-normalizer<4,>=2 in /usr/local/lib/python3.10/dist-packages (from requests<3.0,>=2.7->posthog->haystack-ai) (3.3.2)\n", - "Requirement already satisfied: urllib3<3,>=1.21.1 in /usr/local/lib/python3.10/dist-packages (from requests<3.0,>=2.7->posthog->haystack-ai) (2.0.7)\n", - "Requirement already satisfied: datasets>=2.6.1 in /usr/local/lib/python3.10/dist-packages (2.18.0)\n", - "Requirement already satisfied: filelock in /usr/local/lib/python3.10/dist-packages (from datasets>=2.6.1) (3.13.1)\n", - "Requirement already satisfied: numpy>=1.17 in /usr/local/lib/python3.10/dist-packages (from datasets>=2.6.1) (1.25.2)\n", - "Requirement already satisfied: pyarrow>=12.0.0 in /usr/local/lib/python3.10/dist-packages (from datasets>=2.6.1) (14.0.2)\n", - "Requirement already satisfied: pyarrow-hotfix in /usr/local/lib/python3.10/dist-packages (from datasets>=2.6.1) (0.6)\n", - "Requirement already satisfied: dill<0.3.9,>=0.3.0 in /usr/local/lib/python3.10/dist-packages (from datasets>=2.6.1) (0.3.8)\n", - "Requirement already satisfied: pandas in /usr/local/lib/python3.10/dist-packages (from datasets>=2.6.1) (1.5.3)\n", - "Requirement already satisfied: requests>=2.19.0 in /usr/local/lib/python3.10/dist-packages (from datasets>=2.6.1) (2.31.0)\n", - "Requirement already satisfied: tqdm>=4.62.1 in /usr/local/lib/python3.10/dist-packages (from datasets>=2.6.1) (4.66.2)\n", - "Requirement already satisfied: xxhash in /usr/local/lib/python3.10/dist-packages (from datasets>=2.6.1) (3.4.1)\n", - "Requirement already satisfied: multiprocess in /usr/local/lib/python3.10/dist-packages (from datasets>=2.6.1) (0.70.16)\n", - "Requirement already satisfied: fsspec[http]<=2024.2.0,>=2023.1.0 in /usr/local/lib/python3.10/dist-packages (from datasets>=2.6.1) (2023.6.0)\n", - "Requirement already satisfied: aiohttp in /usr/local/lib/python3.10/dist-packages (from datasets>=2.6.1) (3.9.3)\n", - "Requirement already satisfied: huggingface-hub>=0.19.4 in /usr/local/lib/python3.10/dist-packages (from datasets>=2.6.1) (0.20.3)\n", - "Requirement already satisfied: packaging in /usr/local/lib/python3.10/dist-packages (from datasets>=2.6.1) (23.2)\n", - "Requirement already satisfied: pyyaml>=5.1 in /usr/local/lib/python3.10/dist-packages (from datasets>=2.6.1) (6.0.1)\n", - "Requirement already satisfied: aiosignal>=1.1.2 in /usr/local/lib/python3.10/dist-packages (from aiohttp->datasets>=2.6.1) (1.3.1)\n", - "Requirement already satisfied: attrs>=17.3.0 in /usr/local/lib/python3.10/dist-packages (from aiohttp->datasets>=2.6.1) (23.2.0)\n", - "Requirement already satisfied: frozenlist>=1.1.1 in /usr/local/lib/python3.10/dist-packages (from aiohttp->datasets>=2.6.1) (1.4.1)\n", - "Requirement already satisfied: multidict<7.0,>=4.5 in /usr/local/lib/python3.10/dist-packages (from aiohttp->datasets>=2.6.1) (6.0.5)\n", - "Requirement already satisfied: yarl<2.0,>=1.0 in /usr/local/lib/python3.10/dist-packages (from aiohttp->datasets>=2.6.1) (1.9.4)\n", - "Requirement already satisfied: async-timeout<5.0,>=4.0 in /usr/local/lib/python3.10/dist-packages (from aiohttp->datasets>=2.6.1) (4.0.3)\n", - "Requirement already satisfied: typing-extensions>=3.7.4.3 in /usr/local/lib/python3.10/dist-packages (from huggingface-hub>=0.19.4->datasets>=2.6.1) (4.10.0)\n", - "Requirement already satisfied: charset-normalizer<4,>=2 in /usr/local/lib/python3.10/dist-packages (from requests>=2.19.0->datasets>=2.6.1) (3.3.2)\n", - "Requirement already satisfied: idna<4,>=2.5 in /usr/local/lib/python3.10/dist-packages (from requests>=2.19.0->datasets>=2.6.1) (3.6)\n", - "Requirement already satisfied: urllib3<3,>=1.21.1 in /usr/local/lib/python3.10/dist-packages (from requests>=2.19.0->datasets>=2.6.1) (2.0.7)\n", - "Requirement already satisfied: certifi>=2017.4.17 in /usr/local/lib/python3.10/dist-packages (from requests>=2.19.0->datasets>=2.6.1) (2024.2.2)\n", - "Requirement already satisfied: python-dateutil>=2.8.1 in /usr/local/lib/python3.10/dist-packages (from pandas->datasets>=2.6.1) (2.8.2)\n", - "Requirement already satisfied: pytz>=2020.1 in /usr/local/lib/python3.10/dist-packages (from pandas->datasets>=2.6.1) (2023.4)\n", - "Requirement already satisfied: six>=1.5 in /usr/local/lib/python3.10/dist-packages (from python-dateutil>=2.8.1->pandas->datasets>=2.6.1) (1.16.0)\n", - "Requirement already satisfied: sentence-transformers>=2.2.0 in /usr/local/lib/python3.10/dist-packages (2.5.1)\n", - "Requirement already satisfied: transformers<5.0.0,>=4.32.0 in /usr/local/lib/python3.10/dist-packages (from sentence-transformers>=2.2.0) (4.38.2)\n", - "Requirement already satisfied: tqdm in /usr/local/lib/python3.10/dist-packages (from sentence-transformers>=2.2.0) (4.66.2)\n", - "Requirement already satisfied: torch>=1.11.0 in /usr/local/lib/python3.10/dist-packages (from sentence-transformers>=2.2.0) (2.1.0+cu121)\n", - "Requirement already satisfied: numpy in /usr/local/lib/python3.10/dist-packages (from sentence-transformers>=2.2.0) (1.25.2)\n", - "Requirement already satisfied: scikit-learn in /usr/local/lib/python3.10/dist-packages (from sentence-transformers>=2.2.0) (1.2.2)\n", - "Requirement already satisfied: scipy in /usr/local/lib/python3.10/dist-packages (from sentence-transformers>=2.2.0) (1.11.4)\n", - "Requirement already satisfied: huggingface-hub>=0.15.1 in /usr/local/lib/python3.10/dist-packages (from sentence-transformers>=2.2.0) (0.20.3)\n", - "Requirement already satisfied: Pillow in /usr/local/lib/python3.10/dist-packages (from sentence-transformers>=2.2.0) (9.4.0)\n", - "Requirement already satisfied: filelock in /usr/local/lib/python3.10/dist-packages (from huggingface-hub>=0.15.1->sentence-transformers>=2.2.0) (3.13.1)\n", - "Requirement already satisfied: fsspec>=2023.5.0 in /usr/local/lib/python3.10/dist-packages (from huggingface-hub>=0.15.1->sentence-transformers>=2.2.0) (2023.6.0)\n", - "Requirement already satisfied: requests in /usr/local/lib/python3.10/dist-packages (from huggingface-hub>=0.15.1->sentence-transformers>=2.2.0) (2.31.0)\n", - "Requirement already satisfied: pyyaml>=5.1 in /usr/local/lib/python3.10/dist-packages (from huggingface-hub>=0.15.1->sentence-transformers>=2.2.0) (6.0.1)\n", - "Requirement already satisfied: typing-extensions>=3.7.4.3 in /usr/local/lib/python3.10/dist-packages (from huggingface-hub>=0.15.1->sentence-transformers>=2.2.0) (4.10.0)\n", - "Requirement already satisfied: packaging>=20.9 in /usr/local/lib/python3.10/dist-packages (from huggingface-hub>=0.15.1->sentence-transformers>=2.2.0) (23.2)\n", - "Requirement already satisfied: sympy in /usr/local/lib/python3.10/dist-packages (from torch>=1.11.0->sentence-transformers>=2.2.0) (1.12)\n", - "Requirement already satisfied: networkx in /usr/local/lib/python3.10/dist-packages (from torch>=1.11.0->sentence-transformers>=2.2.0) (3.2.1)\n", - "Requirement already satisfied: jinja2 in /usr/local/lib/python3.10/dist-packages (from torch>=1.11.0->sentence-transformers>=2.2.0) (3.1.3)\n", - "Requirement already satisfied: triton==2.1.0 in /usr/local/lib/python3.10/dist-packages (from torch>=1.11.0->sentence-transformers>=2.2.0) (2.1.0)\n", - "Requirement already satisfied: regex!=2019.12.17 in /usr/local/lib/python3.10/dist-packages (from transformers<5.0.0,>=4.32.0->sentence-transformers>=2.2.0) (2023.12.25)\n", - "Requirement already satisfied: tokenizers<0.19,>=0.14 in /usr/local/lib/python3.10/dist-packages (from transformers<5.0.0,>=4.32.0->sentence-transformers>=2.2.0) (0.15.2)\n", - "Requirement already satisfied: safetensors>=0.4.1 in /usr/local/lib/python3.10/dist-packages (from transformers<5.0.0,>=4.32.0->sentence-transformers>=2.2.0) (0.4.2)\n", - "Requirement already satisfied: joblib>=1.1.1 in /usr/local/lib/python3.10/dist-packages (from scikit-learn->sentence-transformers>=2.2.0) (1.3.2)\n", - "Requirement already satisfied: threadpoolctl>=2.0.0 in /usr/local/lib/python3.10/dist-packages (from scikit-learn->sentence-transformers>=2.2.0) (3.3.0)\n", - "Requirement already satisfied: MarkupSafe>=2.0 in /usr/local/lib/python3.10/dist-packages (from jinja2->torch>=1.11.0->sentence-transformers>=2.2.0) (2.1.5)\n", - "Requirement already satisfied: charset-normalizer<4,>=2 in /usr/local/lib/python3.10/dist-packages (from requests->huggingface-hub>=0.15.1->sentence-transformers>=2.2.0) (3.3.2)\n", - "Requirement already satisfied: idna<4,>=2.5 in /usr/local/lib/python3.10/dist-packages (from requests->huggingface-hub>=0.15.1->sentence-transformers>=2.2.0) (3.6)\n", - "Requirement already satisfied: urllib3<3,>=1.21.1 in /usr/local/lib/python3.10/dist-packages (from requests->huggingface-hub>=0.15.1->sentence-transformers>=2.2.0) (2.0.7)\n", - "Requirement already satisfied: certifi>=2017.4.17 in /usr/local/lib/python3.10/dist-packages (from requests->huggingface-hub>=0.15.1->sentence-transformers>=2.2.0) (2024.2.2)\n", - "Requirement already satisfied: mpmath>=0.19 in /usr/local/lib/python3.10/dist-packages (from sympy->torch>=1.11.0->sentence-transformers>=2.2.0) (1.3.0)\n" - ] - } - ], - "source": [ - "%%bash\n", - "\n", - "pip install haystack-ai\n", - "pip install \"datasets>=2.6.1\"\n", - "pip install \"sentence-transformers>=3.0.0\"" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "id": "Wl_jYERtO-qa" - }, - "source": [ - "### Enabling Telemetry\n", - "\n", - "Knowing you're using this tutorial helps us decide where to invest our efforts to build a better product but you can always opt out by commenting the following line. See [Telemetry](https://docs.haystack.deepset.ai/docs/enabling-telemetry) for more details." - ] - }, - { - "cell_type": "code", - "execution_count": 2, - "metadata": { - "id": "A76B4S49O-qa" - }, - "outputs": [], - "source": [ - "from haystack.telemetry import tutorial_running\n", - "\n", - "tutorial_running(27)" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "id": "_lvfew16O-qa" - }, - "source": [ - "## Fetching and Indexing Documents\n", - "\n", - "You'll start creating your question answering system by downloading the data and indexing the data with its embeddings to a DocumentStore. \n", - "\n", - "In this tutorial, you will take a simple approach to writing documents and their embeddings into the DocumentStore. For a full indexing pipeline with preprocessing, cleaning and splitting, check out our tutorial on [Preprocessing Different File Types](https://haystack.deepset.ai/tutorials/30_file_type_preprocessing_index_pipeline).\n", - "\n", - "\n", - "### Initializing the DocumentStore\n", - "\n", - "Initialize a DocumentStore to index your documents. A DocumentStore stores the Documents that the question answering system uses to find answers to your questions. In this tutorial, you'll be using the `InMemoryDocumentStore`." - ] - }, - { - "cell_type": "code", - "execution_count": 3, - "metadata": { - "id": "CbVN-s5LO-qa" - }, - "outputs": [], - "source": [ - "from haystack.document_stores.in_memory import InMemoryDocumentStore\n", - "\n", - "document_store = InMemoryDocumentStore()" - ] + "cells": [ + { + "cell_type": "markdown", + "metadata": { + "id": "2OvkPji9O-qX" + }, + "source": [ + "# Tutorial: Creating Your First QA Pipeline with Retrieval-Augmentation\n", + "\n", + "- **Level**: Beginner\n", + "- **Time to complete**: 10 minutes\n", + "- **Components Used**: [`InMemoryDocumentStore`](https://docs.haystack.deepset.ai/docs/inmemorydocumentstore), [`SentenceTransformersDocumentEmbedder`](https://docs.haystack.deepset.ai/docs/sentencetransformersdocumentembedder), [`SentenceTransformersTextEmbedder`](https://docs.haystack.deepset.ai/docs/sentencetransformerstextembedder), [`InMemoryEmbeddingRetriever`](https://docs.haystack.deepset.ai/docs/inmemoryembeddingretriever), [`PromptBuilder`](https://docs.haystack.deepset.ai/docs/promptbuilder), [`OpenAIChatGenerator`](https://docs.haystack.deepset.ai/docs/openaichatgenerator)\n", + "- **Prerequisites**: You must have an [OpenAI API Key](https://platform.openai.com/api-keys).\n", + "- **Goal**: After completing this tutorial, you'll have learned the new prompt syntax and how to use PromptBuilder and OpenAIChatGenerator to build a generative question-answering pipeline with retrieval-augmentation.\n", + "\n", + "> This tutorial uses Haystack 2.0. To learn more, read the [Haystack 2.0 announcement](https://haystack.deepset.ai/blog/haystack-2-release) or visit the [Haystack 2.0 Documentation](https://docs.haystack.deepset.ai/docs/intro)." + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "LFqHcXYPO-qZ" + }, + "source": [ + "## Overview\n", + "\n", + "This tutorial shows you how to create a generative question-answering pipeline using the retrieval-augmentation ([RAG](https://www.deepset.ai/blog/llms-retrieval-augmentation)) approach with Haystack 2.0. The process involves four main components: [SentenceTransformersTextEmbedder](https://docs.haystack.deepset.ai/docs/sentencetransformerstextembedder) for creating an embedding for the user query, [InMemoryBM25Retriever](https://docs.haystack.deepset.ai/docs/inmemorybm25retriever) for fetching relevant documents, [PromptBuilder](https://docs.haystack.deepset.ai/docs/promptbuilder) for creating a template prompt, and [OpenAIChatGenerator](https://docs.haystack.deepset.ai/docs/openaichatgenerator) for generating responses.\n", + "\n", + "For this tutorial, you'll use the Wikipedia pages of [Seven Wonders of the Ancient World](https://en.wikipedia.org/wiki/Wonders_of_the_World) as Documents, but you can replace them with any text you want.\n" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "QXjVlbPiO-qZ" + }, + "source": [ + "## Preparing the Colab Environment\n", + "\n", + "- [Enable GPU Runtime in Colab](https://docs.haystack.deepset.ai/docs/enabling-gpu-acceleration)\n", + "- [Set logging level to INFO](https://docs.haystack.deepset.ai/docs/logging)" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "Kww5B_vXO-qZ" + }, + "source": [ + "## Installing Haystack\n", + "\n", + "Install Haystack 2.0 and other required packages with `pip`:" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": { + "colab": { + "base_uri": "https://localhost:8080/" }, - { - "cell_type": "markdown", - "metadata": { - "id": "yL8nuJdWO-qa" - }, - "source": [ - "> `InMemoryDocumentStore` is the simplest DocumentStore to get started with. It requires no external dependencies and it's a good option for smaller projects and debugging. But it doesn't scale up so well to larger Document collections, so it's not a good choice for production systems. To learn more about the different types of external databases that Haystack supports, see [DocumentStore Integrations](https://haystack.deepset.ai/integrations?type=Document+Store)." - ] + "id": "UQbU8GUfO-qZ", + "outputId": "c33579e9-5557-43bd-a3c5-63b8373770c7" + }, + "outputs": [], + "source": [ + "%%bash\n", + "\n", + "pip install haystack-ai\n", + "pip install \"datasets>=2.6.1\"\n", + "pip install \"sentence-transformers>=3.0.0\"" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "Wl_jYERtO-qa" + }, + "source": [ + "### Enabling Telemetry\n", + "\n", + "Knowing you're using this tutorial helps us decide where to invest our efforts to build a better product but you can always opt out by commenting the following line. See [Telemetry](https://docs.haystack.deepset.ai/docs/enabling-telemetry) for more details." + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": { + "id": "A76B4S49O-qa" + }, + "outputs": [], + "source": [ + "from haystack.telemetry import tutorial_running\n", + "\n", + "tutorial_running(27)" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "_lvfew16O-qa" + }, + "source": [ + "## Fetching and Indexing Documents\n", + "\n", + "You'll start creating your question answering system by downloading the data and indexing the data with its embeddings to a DocumentStore. \n", + "\n", + "In this tutorial, you will take a simple approach to writing documents and their embeddings into the DocumentStore. For a full indexing pipeline with preprocessing, cleaning and splitting, check out our tutorial on [Preprocessing Different File Types](https://haystack.deepset.ai/tutorials/30_file_type_preprocessing_index_pipeline).\n", + "\n", + "\n", + "### Initializing the DocumentStore\n", + "\n", + "Initialize a DocumentStore to index your documents. A DocumentStore stores the Documents that the question answering system uses to find answers to your questions. In this tutorial, you'll be using the `InMemoryDocumentStore`." + ] + }, + { + "cell_type": "code", + "execution_count": 3, + "metadata": { + "id": "CbVN-s5LO-qa" + }, + "outputs": [], + "source": [ + "from haystack.document_stores.in_memory import InMemoryDocumentStore\n", + "\n", + "document_store = InMemoryDocumentStore()" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "yL8nuJdWO-qa" + }, + "source": [ + "> `InMemoryDocumentStore` is the simplest DocumentStore to get started with. It requires no external dependencies and it's a good option for smaller projects and debugging. But it doesn't scale up so well to larger Document collections, so it's not a good choice for production systems. To learn more about the different types of external databases that Haystack supports, see [DocumentStore Integrations](https://haystack.deepset.ai/integrations?type=Document+Store)." + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "XvLVaFHTO-qb" + }, + "source": [ + "The DocumentStore is now ready. Now it's time to fill it with some Documents." + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "HryYZP9ZO-qb" + }, + "source": [ + "### Fetch the Data\n", + "\n", + "You'll use the Wikipedia pages of [Seven Wonders of the Ancient World](https://en.wikipedia.org/wiki/Wonders_of_the_World) as Documents. We preprocessed the data and uploaded to a Hugging Face Space: [Seven Wonders](https://huggingface.co/datasets/bilgeyucel/seven-wonders). Thus, you don't need to perform any additional cleaning or splitting.\n", + "\n", + "Fetch the data and convert it into Haystack Documents:" + ] + }, + { + "cell_type": "code", + "execution_count": 4, + "metadata": { + "colab": { + "base_uri": "https://localhost:8080/" }, - { - "cell_type": "markdown", - "metadata": { - "id": "XvLVaFHTO-qb" - }, - "source": [ - "The DocumentStore is now ready. Now it's time to fill it with some Documents." - ] + "id": "INdC3WvLO-qb", + "outputId": "1af43d0f-2999-4de4-d152-b3cca9fb49e6" + }, + "outputs": [], + "source": [ + "from datasets import load_dataset\n", + "from haystack import Document\n", + "\n", + "dataset = load_dataset(\"bilgeyucel/seven-wonders\", split=\"train\")\n", + "docs = [Document(content=doc[\"content\"], meta=doc[\"meta\"]) for doc in dataset]" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "czMjWwnxPA-3" + }, + "source": [ + "### Initalize a Document Embedder\n", + "\n", + "To store your data in the DocumentStore with embeddings, initialize a [SentenceTransformersDocumentEmbedder](https://docs.haystack.deepset.ai/docs/sentencetransformersdocumentembedder) with the model name and call `warm_up()` to download the embedding model.\n", + "\n", + "> If you'd like, you can use a different [Embedder](https://docs.haystack.deepset.ai/docs/embedders) for your documents." + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": { + "colab": { + "base_uri": "https://localhost:8080/" }, - { - "cell_type": "markdown", - "metadata": { - "id": "HryYZP9ZO-qb" - }, - "source": [ - "### Fetch the Data\n", - "\n", - "You'll use the Wikipedia pages of [Seven Wonders of the Ancient World](https://en.wikipedia.org/wiki/Wonders_of_the_World) as Documents. We preprocessed the data and uploaded to a Hugging Face Space: [Seven Wonders](https://huggingface.co/datasets/bilgeyucel/seven-wonders). Thus, you don't need to perform any additional cleaning or splitting.\n", - "\n", - "Fetch the data and convert it into Haystack Documents:" - ] + "id": "EUmAH9sEn3R7", + "outputId": "ee54b59b-4d4a-45eb-c1a9-0b7b248f1dd4" + }, + "outputs": [], + "source": [ + "from haystack.components.embedders import SentenceTransformersDocumentEmbedder\n", + "\n", + "doc_embedder = SentenceTransformersDocumentEmbedder(model=\"sentence-transformers/all-MiniLM-L6-v2\")\n", + "doc_embedder.warm_up()" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "9y4iJE_SrS4K" + }, + "source": [ + "### Write Documents to the DocumentStore\n", + "\n", + "Run the `doc_embedder` with the Documents. The embedder will create embeddings for each document and save these embeddings in Document object's `embedding` field. Then, you can write the Documents to the DocumentStore with `write_documents()` method." + ] + }, + { + "cell_type": "code", + "execution_count": 6, + "metadata": { + "colab": { + "base_uri": "https://localhost:8080/", + "height": 66, + "referenced_widgets": [ + "7d482188c12d4a7886f20a65d3402c59", + "2a3ec74419ae4a02ac0210db66133415", + "ddeff9a822404adbbc3cad97a939bc0c", + "36d341ab3a044709b5af2e8ab97559bc", + "88fc33e1ab78405e911b5eafa512c935", + "91e5d4b0ede848319ef0d3b558d57d19", + "d2428c21707d43f2b6f07bfafbace8bb", + "7fdb2c859e454e72888709a835f7591e", + "6b8334e071a3438397ba6435aac69f58", + "5f5cfa425cac4d37b2ea29e53b4ed900", + "3c59a82dac5c476b9a3e3132094e1702" + ] }, + "id": "ETpQKftLplqh", + "outputId": "b9c8658c-90c8-497c-e765-97487c0daf8e" + }, + "outputs": [ { - "cell_type": "code", - "execution_count": 4, - "metadata": { - "colab": { - "base_uri": "https://localhost:8080/" - }, - "id": "INdC3WvLO-qb", - "outputId": "1af43d0f-2999-4de4-d152-b3cca9fb49e6" - }, - "outputs": [ - { - "name": "stderr", - "output_type": "stream", - "text": [ - "/usr/local/lib/python3.10/dist-packages/huggingface_hub/utils/_token.py:88: UserWarning: \n", - "The secret `HF_TOKEN` does not exist in your Colab secrets.\n", - "To authenticate with the Hugging Face Hub, create a token in your settings tab (https://huggingface.co/settings/tokens), set it as secret in your Google Colab and restart your session.\n", - "You will be able to reuse this secret in all of your notebooks.\n", - "Please note that authentication is recommended but still optional to access public models or datasets.\n", - " warnings.warn(\n" - ] - } - ], - "source": [ - "from datasets import load_dataset\n", - "from haystack import Document\n", - "\n", - "dataset = load_dataset(\"bilgeyucel/seven-wonders\", split=\"train\")\n", - "docs = [Document(content=doc[\"content\"], meta=doc[\"meta\"]) for doc in dataset]" - ] + "name": "stderr", + "output_type": "stream", + "text": [ + "Batches: 0%| | 0/5 [00:00 If you'd like, you can use a different [Embedder](https://docs.haystack.deepset.ai/docs/embedders) for your documents." - ] + "name": "stderr", + "output_type": "stream", + "text": [ + "Batches: 100%|██████████| 5/5 [00:01<00:00, 3.38it/s]\n" + ] }, { - "cell_type": "code", - "execution_count": 5, - "metadata": { - "colab": { - "base_uri": "https://localhost:8080/" - }, - "id": "EUmAH9sEn3R7", - "outputId": "ee54b59b-4d4a-45eb-c1a9-0b7b248f1dd4" - }, - "outputs": [ - { - "name": "stderr", - "output_type": "stream", - "text": [ - "/usr/local/lib/python3.10/dist-packages/torch/_utils.py:831: UserWarning: TypedStorage is deprecated. It will be removed in the future and UntypedStorage will be the only storage class. This should only matter to you if you are using storages directly. To access UntypedStorage directly, use tensor.untyped_storage() instead of tensor.storage()\n", - " return self.fget.__get__(instance, owner)()\n" - ] - } - ], - "source": [ - "from haystack.components.embedders import SentenceTransformersDocumentEmbedder\n", - "\n", - "doc_embedder = SentenceTransformersDocumentEmbedder(model=\"sentence-transformers/all-MiniLM-L6-v2\")\n", - "doc_embedder.warm_up()" + "data": { + "text/plain": [ + "151" ] + }, + "execution_count": 6, + "metadata": {}, + "output_type": "execute_result" + } + ], + "source": [ + "docs_with_embeddings = doc_embedder.run(docs)\n", + "document_store.write_documents(docs_with_embeddings[\"documents\"])" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "IdojTxg6uubn" + }, + "source": [ + "## Building the RAG Pipeline\n", + "\n", + "The next step is to build a [Pipeline](https://docs.haystack.deepset.ai/docs/pipelines) to generate answers for the user query following the RAG approach. To create the pipeline, you first need to initialize each component, add them to your pipeline, and connect them." + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "0uyV6-u-u56P" + }, + "source": [ + "### Initialize a Text Embedder\n", + "\n", + "Initialize a text embedder to create an embedding for the user query. The created embedding will later be used by the Retriever to retrieve relevant documents from the DocumentStore.\n", + "\n", + "> ⚠️ Notice that you used `sentence-transformers/all-MiniLM-L6-v2` model to create embeddings for your documents before. This is why you need to use the same model to embed the user queries." + ] + }, + { + "cell_type": "code", + "execution_count": 7, + "metadata": { + "id": "LyJY2yW628dl" + }, + "outputs": [], + "source": [ + "from haystack.components.embedders import SentenceTransformersTextEmbedder\n", + "\n", + "text_embedder = SentenceTransformersTextEmbedder(model=\"sentence-transformers/all-MiniLM-L6-v2\")" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "0_cj-5m-O-qb" + }, + "source": [ + "### Initialize the Retriever\n", + "\n", + "Initialize a [InMemoryEmbeddingRetriever](https://docs.haystack.deepset.ai/docs/inmemoryembeddingretriever) and make it use the InMemoryDocumentStore you initialized earlier in this tutorial. This Retriever will get the relevant documents to the query." + ] + }, + { + "cell_type": "code", + "execution_count": 8, + "metadata": { + "id": "-uo-6fjiO-qb" + }, + "outputs": [], + "source": [ + "from haystack.components.retrievers.in_memory import InMemoryEmbeddingRetriever\n", + "\n", + "retriever = InMemoryEmbeddingRetriever(document_store)" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "6CEuQpB7O-qb" + }, + "source": [ + "### Define a Template Prompt\n", + "\n", + "Create a custom prompt for a generative question answering task using the RAG approach. The prompt should take in two parameters: `documents`, which are retrieved from a document store, and a `question` from the user. Use the Jinja2 looping syntax to combine the content of the retrieved documents in the prompt.\n", + "\n", + "Next, initialize a [PromptBuilder](https://docs.haystack.deepset.ai/docs/promptbuilder) instance with your prompt template. The PromptBuilder, when given the necessary values, will automatically fill in the variable values and generate a complete prompt. This approach allows for a more tailored and effective question-answering experience." + ] + }, + { + "cell_type": "code", + "execution_count": 9, + "metadata": { + "id": "ObahTh45FqOT" + }, + "outputs": [], + "source": [ + "from haystack.components.builders import ChatPromptBuilder\n", + "from haystack.dataclasses import ChatMessage\n", + "\n", + "template = [ChatMessage.from_user(\"\"\"\n", + "Given the following information, answer the question.\n", + "\n", + "Context:\n", + "{% for document in documents %}\n", + " {{ document.content }}\n", + "{% endfor %}\n", + "\n", + "Question: {{question}}\n", + "Answer:\n", + "\"\"\")]\n", + "\n", + "prompt_builder = ChatPromptBuilder(template=template)" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "HR14lbfcFtXj" + }, + "source": [ + "### Initialize a ChatGenerator\n", + "\n", + "\n", + "ChatGenerators are the components that interact with large language models (LLMs). Now, set `OPENAI_API_KEY` environment variable and initialize a [OpenAIChatGenerator](https://docs.haystack.deepset.ai/docs/OpenAIChatGenerator) that can communicate with OpenAI GPT models. As you initialize, provide a model name:" + ] + }, + { + "cell_type": "code", + "execution_count": 10, + "metadata": { + "colab": { + "base_uri": "https://localhost:8080/" }, - { - "cell_type": "markdown", - "metadata": { - "id": "9y4iJE_SrS4K" - }, - "source": [ - "### Write Documents to the DocumentStore\n", - "\n", - "Run the `doc_embedder` with the Documents. The embedder will create embeddings for each document and save these embeddings in Document object's `embedding` field. Then, you can write the Documents to the DocumentStore with `write_documents()` method." - ] + "id": "SavE_FAqfApo", + "outputId": "1afbf2e8-ae63-41ff-c37f-5123b2103356" + }, + "outputs": [], + "source": [ + "import os\n", + "from getpass import getpass\n", + "from haystack.components.generators.chat import OpenAIChatGenerator\n", + "\n", + "if \"OPENAI_API_KEY\" not in os.environ:\n", + " os.environ[\"OPENAI_API_KEY\"] = getpass(\"Enter OpenAI API key:\")\n", + "chat_generator = OpenAIChatGenerator(model=\"gpt-4o-mini\")" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "nenbo2SvycHd" + }, + "source": [ + "> You can replace `OpenAIChatGenerator` in your pipeline with another `ChatGenerator`. Check out the full list of chat generators [here](https://docs.haystack.deepset.ai/docs/generators)." + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "1bfHwOQwycHe" + }, + "source": [ + "### Build the Pipeline\n", + "\n", + "To build a pipeline, add all components to your pipeline and connect them. Create connections from `text_embedder`'s \"embedding\" output to \"query_embedding\" input of `retriever`, from `retriever` to `prompt_builder` and from `prompt_builder` to `llm`. Explicitly connect the output of `retriever` with \"documents\" input of the `prompt_builder` to make the connection obvious as `prompt_builder` has two inputs (\"documents\" and \"question\").\n", + "\n", + "For more information on pipelines and creating connections, refer to [Creating Pipelines](https://docs.haystack.deepset.ai/docs/creating-pipelines) documentation." + ] + }, + { + "cell_type": "code", + "execution_count": 11, + "metadata": { + "colab": { + "base_uri": "https://localhost:8080/", + "height": 1000 }, + "id": "f6NFmpjEO-qb", + "outputId": "89fd1b48-5189-4401-9cf8-15f55c503676" + }, + "outputs": [], + "source": [ + "from haystack import Pipeline\n", + "\n", + "basic_rag_pipeline = Pipeline()\n", + "# Add components to your pipeline\n", + "basic_rag_pipeline.add_component(\"text_embedder\", text_embedder)\n", + "basic_rag_pipeline.add_component(\"retriever\", retriever)\n", + "basic_rag_pipeline.add_component(\"prompt_builder\", prompt_builder)\n", + "basic_rag_pipeline.add_component(\"llm\", chat_generator)\n" + ] + }, + { + "cell_type": "code", + "execution_count": 12, + "metadata": {}, + "outputs": [ { - "cell_type": "code", - "execution_count": 6, - "metadata": { - "colab": { - "base_uri": "https://localhost:8080/", - "height": 66, - "referenced_widgets": [ - "7d482188c12d4a7886f20a65d3402c59", - "2a3ec74419ae4a02ac0210db66133415", - "ddeff9a822404adbbc3cad97a939bc0c", - "36d341ab3a044709b5af2e8ab97559bc", - "88fc33e1ab78405e911b5eafa512c935", - "91e5d4b0ede848319ef0d3b558d57d19", - "d2428c21707d43f2b6f07bfafbace8bb", - "7fdb2c859e454e72888709a835f7591e", - "6b8334e071a3438397ba6435aac69f58", - "5f5cfa425cac4d37b2ea29e53b4ed900", - "3c59a82dac5c476b9a3e3132094e1702" - ] - }, - "id": "ETpQKftLplqh", - "outputId": "b9c8658c-90c8-497c-e765-97487c0daf8e" - }, - "outputs": [ - { - "data": { - "application/vnd.jupyter.widget-view+json": { - "model_id": "7d482188c12d4a7886f20a65d3402c59", - "version_major": 2, - "version_minor": 0 - }, - "text/plain": [ - "Batches: 0%| | 0/5 [00:00\n", + "🚅 Components\n", + " - text_embedder: SentenceTransformersTextEmbedder\n", + " - retriever: InMemoryEmbeddingRetriever\n", + " - prompt_builder: ChatPromptBuilder\n", + " - llm: OpenAIChatGenerator\n", + "🛤️ Connections\n", + " - text_embedder.embedding -> retriever.query_embedding (List[float])\n", + " - retriever.documents -> prompt_builder.documents (List[Document])\n", + " - prompt_builder.prompt -> llm.messages (List[ChatMessage])" ] + }, + "execution_count": 12, + "metadata": {}, + "output_type": "execute_result" + } + ], + "source": [ + "# Now, connect the components to each other\n", + "basic_rag_pipeline.connect(\"text_embedder.embedding\", \"retriever.query_embedding\")\n", + "basic_rag_pipeline.connect(\"retriever\", \"prompt_builder\")\n", + "basic_rag_pipeline.connect(\"prompt_builder.prompt\", \"llm.messages\")" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "6NqyLhx7O-qc" + }, + "source": [ + "That's it! Your RAG pipeline is ready to generate answers to questions!" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "DBAyF5tVO-qc" + }, + "source": [ + "## Asking a Question\n", + "\n", + "When asking a question, use the `run()` method of the pipeline. Make sure to provide the question to both the `text_embedder` and the `prompt_builder`. This ensures that the `{{question}}` variable in the template prompt gets replaced with your specific question." + ] + }, + { + "cell_type": "code", + "execution_count": 13, + "metadata": { + "colab": { + "base_uri": "https://localhost:8080/", + "height": 86, + "referenced_widgets": [ + "4e6e97b6d54f4f80bb7e8b25aba8e616", + "1a820c06a7a049d8b6c9ff300284d06e", + "58ff4e0603a74978a134f63533859be5", + "8bdb8bfae31d4f4cb6c3b0bf43120eed", + "39a68d9a5c274e2dafaa2d1f86eea768", + "d0cfe5dacdfc431a91b4c4741123e2d0", + "e7f1e1a14bb740d18827dd78bbe7b2e3", + "3fda06f905b445a488efdd2dd08c0939", + "2bc341a780f7498ba9cd475468841bb5", + "d7218475e23b420a8c03d00ca4ab8718", + "a694abaf765f4d1b82fa0138e59c6793" + ] }, + "id": "Vnt283M5O-qc", + "outputId": "d2843a73-3ad5-4daa-8d1e-a58de7aa2bb0" + }, + "outputs": [ { - "cell_type": "markdown", - "metadata": { - "id": "IdojTxg6uubn" - }, - "source": [ - "## Building the RAG Pipeline\n", - "\n", - "The next step is to build a [Pipeline](https://docs.haystack.deepset.ai/docs/pipelines) to generate answers for the user query following the RAG approach. To create the pipeline, you first need to initialize each component, add them to your pipeline, and connect them." - ] + "name": "stderr", + "output_type": "stream", + "text": [ + "Batches: 100%|██████████| 1/1 [00:00<00:00, 1.77it/s]\n", + "huggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...\n", + "To disable this warning, you can either:\n", + "\t- Avoid using `tokenizers` before the fork if possible\n", + "\t- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)\n" + ] }, { - "cell_type": "markdown", - "metadata": { - "id": "0uyV6-u-u56P" - }, - "source": [ - "### Initialize a Text Embedder\n", - "\n", - "Initialize a text embedder to create an embedding for the user query. The created embedding will later be used by the Retriever to retrieve relevant documents from the DocumentStore.\n", - "\n", - "> ⚠️ Notice that you used `sentence-transformers/all-MiniLM-L6-v2` model to create embeddings for your documents before. This is why you need to use the same model to embed the user queries." - ] + "name": "stdout", + "output_type": "stream", + "text": [ + "The Colossus of Rhodes was a statue of the Greek sun-god Helios, standing approximately 70 cubits (about 33 meters or 108 feet) tall. Although no complete descriptions of its appearance exist, scholars believe it featured the following characteristics:\n", + "\n", + "1. **Facial Features**: The head of the statue likely had curly hair, with spikes resembling bronze or silver flames radiating outward. This style is similar to depictions found on contemporary Rhodian coins.\n", + "\n", + "2. **Posture**: While the exact pose is uncertain, it is suggested that the statue may have been constructed in a pose where Helios is depicted shielding his eyes with one hand, a common representation of someone looking toward the sun.\n", + "\n", + "3. **Construction Materials**: The structure was built using iron tie bars and brass plates, which formed the skin of the statue. The interior was filled with stone blocks.\n", + "\n", + "4. **Height and Scale**: The Colossus was positioned on a 15-metre-high (49-foot) pedestal, making it one of the tallest statues of the ancient world, towering over the harbor entrance.\n", + "\n", + "5. **Symbolic Representation**: The statue was meant to symbolize the victory and freedom of the Rhodians after successfully defending their city against an invader.\n", + "\n", + "Overall, the Colossus of Rhodes was an impressive and monumental statue designed to celebrate and symbolize the strength and resilience of the city of Rhodes.\n" + ] + } + ], + "source": [ + "question = \"What does Rhodes Statue look like?\"\n", + "\n", + "response = basic_rag_pipeline.run({\"text_embedder\": {\"text\": question}, \"prompt_builder\": {\"question\": question}})\n", + "\n", + "print(response[\"llm\"][\"replies\"][0].text)" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "IWQN-aoGO-qc" + }, + "source": [ + "Here are some other example questions to test:" + ] + }, + { + "cell_type": "code", + "execution_count": 14, + "metadata": { + "id": "_OHUQ5xxO-qc" + }, + "outputs": [], + "source": [ + "examples = [\n", + " \"Where is Gardens of Babylon?\",\n", + " \"Why did people build Great Pyramid of Giza?\",\n", + " \"What does Rhodes Statue look like?\",\n", + " \"Why did people visit the Temple of Artemis?\",\n", + " \"What is the importance of Colossus of Rhodes?\",\n", + " \"What happened to the Tomb of Mausolus?\",\n", + " \"How did Colossus of Rhodes collapse?\",\n", + "]" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "XueCK3y4O-qc" + }, + "source": [ + "## What's next\n", + "\n", + "🎉 Congratulations! You've learned how to create a generative QA system for your documents with the RAG approach.\n", + "\n", + "If you liked this tutorial, you may also enjoy:\n", + "- [Filtering Documents with Metadata](https://haystack.deepset.ai/tutorials/31_metadata_filtering)\n", + "- [Preprocessing Different File Types](https://haystack.deepset.ai/tutorials/30_file_type_preprocessing_index_pipeline)\n", + "- [Creating a Hybrid Retrieval Pipeline](https://haystack.deepset.ai/tutorials/33_hybrid_retrieval)\n", + "\n", + "To stay up to date on the latest Haystack developments, you can [subscribe to our newsletter](https://landing.deepset.ai/haystack-community-updates) and [join Haystack discord community](https://discord.gg/haystack).\n", + "\n", + "Thanks for reading!" + ] + } + ], + "metadata": { + "accelerator": "GPU", + "colab": { + "gpuType": "T4", + "provenance": [] + }, + "kernelspec": { + "display_name": "Python 3", + "name": "python3" + }, + "language_info": { + "codemirror_mode": { + "name": "ipython", + "version": 3 + }, + "file_extension": ".py", + "mimetype": "text/x-python", + "name": "python", + "nbconvert_exporter": "python", + "pygments_lexer": "ipython3", + "version": "3.9.6" + }, + "orig_nbformat": 4, + "vscode": { + "interpreter": { + "hash": "31f2aee4e71d21fbe5cf8b01ff0e069b9275f58929596ceb00d14d90e3e16cd6" + } + }, + "widgets": { + "application/vnd.jupyter.widget-state+json": { + "1a820c06a7a049d8b6c9ff300284d06e": { + "model_module": "@jupyter-widgets/controls", + "model_module_version": "1.5.0", + "model_name": "HTMLModel", + "state": { + "_dom_classes": [], + "_model_module": "@jupyter-widgets/controls", + "_model_module_version": "1.5.0", + "_model_name": "HTMLModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/controls", + "_view_module_version": "1.5.0", + "_view_name": "HTMLView", + "description": "", + "description_tooltip": null, + "layout": "IPY_MODEL_d0cfe5dacdfc431a91b4c4741123e2d0", + "placeholder": "​", + "style": "IPY_MODEL_e7f1e1a14bb740d18827dd78bbe7b2e3", + "value": "Batches: 100%" + } }, - { - "cell_type": "code", - "execution_count": 7, - "metadata": { - "id": "LyJY2yW628dl" - }, - "outputs": [], - "source": [ - "from haystack.components.embedders import SentenceTransformersTextEmbedder\n", - "\n", - "text_embedder = SentenceTransformersTextEmbedder(model=\"sentence-transformers/all-MiniLM-L6-v2\")" - ] + "2a3ec74419ae4a02ac0210db66133415": { + "model_module": "@jupyter-widgets/controls", + "model_module_version": "1.5.0", + "model_name": "HTMLModel", + "state": { + "_dom_classes": [], + "_model_module": "@jupyter-widgets/controls", + "_model_module_version": "1.5.0", + "_model_name": "HTMLModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/controls", + "_view_module_version": "1.5.0", + "_view_name": "HTMLView", + "description": "", + "description_tooltip": null, + "layout": "IPY_MODEL_91e5d4b0ede848319ef0d3b558d57d19", + "placeholder": "​", + "style": "IPY_MODEL_d2428c21707d43f2b6f07bfafbace8bb", + "value": "Batches: 100%" + } }, - { - "cell_type": "markdown", - "metadata": { - "id": "0_cj-5m-O-qb" - }, - "source": [ - "### Initialize the Retriever\n", - "\n", - "Initialize a [InMemoryEmbeddingRetriever](https://docs.haystack.deepset.ai/docs/inmemoryembeddingretriever) and make it use the InMemoryDocumentStore you initialized earlier in this tutorial. This Retriever will get the relevant documents to the query." - ] + "2bc341a780f7498ba9cd475468841bb5": { + "model_module": "@jupyter-widgets/controls", + "model_module_version": "1.5.0", + "model_name": "ProgressStyleModel", + "state": { + "_model_module": "@jupyter-widgets/controls", + "_model_module_version": "1.5.0", + "_model_name": "ProgressStyleModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/base", + "_view_module_version": "1.2.0", + "_view_name": "StyleView", + "bar_color": null, + "description_width": "" + } }, - { - "cell_type": "code", - "execution_count": 8, - "metadata": { - "id": "-uo-6fjiO-qb" - }, - "outputs": [], - "source": [ - "from haystack.components.retrievers.in_memory import InMemoryEmbeddingRetriever\n", - "\n", - "retriever = InMemoryEmbeddingRetriever(document_store)" - ] + "36d341ab3a044709b5af2e8ab97559bc": { + "model_module": "@jupyter-widgets/controls", + "model_module_version": "1.5.0", + "model_name": "HTMLModel", + "state": { + "_dom_classes": [], + "_model_module": "@jupyter-widgets/controls", + "_model_module_version": "1.5.0", + "_model_name": "HTMLModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/controls", + "_view_module_version": "1.5.0", + "_view_name": "HTMLView", + "description": "", + "description_tooltip": null, + "layout": "IPY_MODEL_5f5cfa425cac4d37b2ea29e53b4ed900", + "placeholder": "​", + "style": "IPY_MODEL_3c59a82dac5c476b9a3e3132094e1702", + "value": " 5/5 [00:01<00:00,  3.35it/s]" + } }, - { - "cell_type": "markdown", - "metadata": { - "id": "6CEuQpB7O-qb" - }, - "source": [ - "### Define a Template Prompt\n", - "\n", - "Create a custom prompt for a generative question answering task using the RAG approach. The prompt should take in two parameters: `documents`, which are retrieved from a document store, and a `question` from the user. Use the Jinja2 looping syntax to combine the content of the retrieved documents in the prompt.\n", - "\n", - "Next, initialize a [PromptBuilder](https://docs.haystack.deepset.ai/docs/promptbuilder) instance with your prompt template. The PromptBuilder, when given the necessary values, will automatically fill in the variable values and generate a complete prompt. This approach allows for a more tailored and effective question-answering experience." - ] + "39a68d9a5c274e2dafaa2d1f86eea768": { + "model_module": "@jupyter-widgets/base", + "model_module_version": "1.2.0", + "model_name": "LayoutModel", + "state": { + "_model_module": "@jupyter-widgets/base", + "_model_module_version": "1.2.0", + "_model_name": "LayoutModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/base", + "_view_module_version": "1.2.0", + "_view_name": "LayoutView", + "align_content": null, + "align_items": null, + "align_self": null, + "border": null, + "bottom": null, + "display": null, + "flex": null, + "flex_flow": null, + "grid_area": null, + "grid_auto_columns": null, + "grid_auto_flow": null, + "grid_auto_rows": null, + "grid_column": null, + "grid_gap": null, + "grid_row": null, + "grid_template_areas": null, + "grid_template_columns": null, + "grid_template_rows": null, + "height": null, + "justify_content": null, + "justify_items": null, + "left": null, + "margin": null, + "max_height": null, + "max_width": null, + "min_height": null, + "min_width": null, + "object_fit": null, + "object_position": null, + "order": null, + "overflow": null, + "overflow_x": null, + "overflow_y": null, + "padding": null, + "right": null, + "top": null, + "visibility": null, + "width": null + } }, - { - "cell_type": "code", - "execution_count": 9, - "metadata": { - "id": "ObahTh45FqOT" - }, - "outputs": [], - "source": [ - "from haystack.components.builders import PromptBuilder\n", - "\n", - "template = \"\"\"\n", - "Given the following information, answer the question.\n", - "\n", - "Context:\n", - "{% for document in documents %}\n", - " {{ document.content }}\n", - "{% endfor %}\n", - "\n", - "Question: {{question}}\n", - "Answer:\n", - "\"\"\"\n", - "\n", - "prompt_builder = PromptBuilder(template=template)" - ] + "3c59a82dac5c476b9a3e3132094e1702": { + "model_module": "@jupyter-widgets/controls", + "model_module_version": "1.5.0", + "model_name": "DescriptionStyleModel", + "state": { + "_model_module": "@jupyter-widgets/controls", + "_model_module_version": "1.5.0", + "_model_name": "DescriptionStyleModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/base", + "_view_module_version": "1.2.0", + "_view_name": "StyleView", + "description_width": "" + } }, - { - "cell_type": "markdown", - "metadata": { - "id": "HR14lbfcFtXj" - }, - "source": [ - "### Initialize a Generator\n", - "\n", - "\n", - "Generators are the components that interact with large language models (LLMs). Now, set `OPENAI_API_KEY` environment variable and initialize a [OpenAIGenerator](https://docs.haystack.deepset.ai/docs/OpenAIGenerator) that can communicate with OpenAI GPT models. As you initialize, provide a model name:" - ] + "3fda06f905b445a488efdd2dd08c0939": { + "model_module": "@jupyter-widgets/base", + "model_module_version": "1.2.0", + "model_name": "LayoutModel", + "state": { + "_model_module": "@jupyter-widgets/base", + "_model_module_version": "1.2.0", + "_model_name": "LayoutModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/base", + "_view_module_version": "1.2.0", + "_view_name": "LayoutView", + "align_content": null, + "align_items": null, + "align_self": null, + "border": null, + "bottom": null, + "display": null, + "flex": null, + "flex_flow": null, + "grid_area": null, + "grid_auto_columns": null, + "grid_auto_flow": null, + "grid_auto_rows": null, + "grid_column": null, + "grid_gap": null, + "grid_row": null, + "grid_template_areas": null, + "grid_template_columns": null, + "grid_template_rows": null, + "height": null, + "justify_content": null, + "justify_items": null, + "left": null, + "margin": null, + "max_height": null, + "max_width": null, + "min_height": null, + "min_width": null, + "object_fit": null, + "object_position": null, + "order": null, + "overflow": null, + "overflow_x": null, + "overflow_y": null, + "padding": null, + "right": null, + "top": null, + "visibility": null, + "width": null + } }, - { - "cell_type": "code", - "execution_count": 10, - "metadata": { - "colab": { - "base_uri": "https://localhost:8080/" - }, - "id": "SavE_FAqfApo", - "outputId": "1afbf2e8-ae63-41ff-c37f-5123b2103356" - }, - "outputs": [ - { - "name": "stdout", - "output_type": "stream", - "text": [ - "Enter OpenAI API key: ··········\n" - ] - } + "4e6e97b6d54f4f80bb7e8b25aba8e616": { + "model_module": "@jupyter-widgets/controls", + "model_module_version": "1.5.0", + "model_name": "HBoxModel", + "state": { + "_dom_classes": [], + "_model_module": "@jupyter-widgets/controls", + "_model_module_version": "1.5.0", + "_model_name": "HBoxModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/controls", + "_view_module_version": "1.5.0", + "_view_name": "HBoxView", + "box_style": "", + "children": [ + "IPY_MODEL_1a820c06a7a049d8b6c9ff300284d06e", + "IPY_MODEL_58ff4e0603a74978a134f63533859be5", + "IPY_MODEL_8bdb8bfae31d4f4cb6c3b0bf43120eed" ], - "source": [ - "import os\n", - "from getpass import getpass\n", - "from haystack.components.generators import OpenAIGenerator\n", - "\n", - "if \"OPENAI_API_KEY\" not in os.environ:\n", - " os.environ[\"OPENAI_API_KEY\"] = getpass(\"Enter OpenAI API key:\")\n", - "generator = OpenAIGenerator(model=\"gpt-4o-mini\")" - ] + "layout": "IPY_MODEL_39a68d9a5c274e2dafaa2d1f86eea768" + } }, - { - "cell_type": "markdown", - "metadata": { - "id": "nenbo2SvycHd" - }, - "source": [ - "> You can replace `OpenAIGenerator` in your pipeline with another `Generator`. Check out the full list of generators [here](https://docs.haystack.deepset.ai/docs/generators)." - ] + "58ff4e0603a74978a134f63533859be5": { + "model_module": "@jupyter-widgets/controls", + "model_module_version": "1.5.0", + "model_name": "FloatProgressModel", + "state": { + "_dom_classes": [], + "_model_module": "@jupyter-widgets/controls", + "_model_module_version": "1.5.0", + "_model_name": "FloatProgressModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/controls", + "_view_module_version": "1.5.0", + "_view_name": "ProgressView", + "bar_style": "success", + "description": "", + "description_tooltip": null, + "layout": "IPY_MODEL_3fda06f905b445a488efdd2dd08c0939", + "max": 1, + "min": 0, + "orientation": "horizontal", + "style": "IPY_MODEL_2bc341a780f7498ba9cd475468841bb5", + "value": 1 + } }, - { - "cell_type": "markdown", - "metadata": { - "id": "1bfHwOQwycHe" - }, - "source": [ - "### Build the Pipeline\n", - "\n", - "To build a pipeline, add all components to your pipeline and connect them. Create connections from `text_embedder`'s \"embedding\" output to \"query_embedding\" input of `retriever`, from `retriever` to `prompt_builder` and from `prompt_builder` to `llm`. Explicitly connect the output of `retriever` with \"documents\" input of the `prompt_builder` to make the connection obvious as `prompt_builder` has two inputs (\"documents\" and \"question\").\n", - "\n", - "For more information on pipelines and creating connections, refer to [Creating Pipelines](https://docs.haystack.deepset.ai/docs/creating-pipelines) documentation." - ] + "5f5cfa425cac4d37b2ea29e53b4ed900": { + "model_module": "@jupyter-widgets/base", + "model_module_version": "1.2.0", + "model_name": "LayoutModel", + "state": { + "_model_module": "@jupyter-widgets/base", + "_model_module_version": "1.2.0", + "_model_name": "LayoutModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/base", + "_view_module_version": "1.2.0", + "_view_name": "LayoutView", + "align_content": null, + "align_items": null, + "align_self": null, + "border": null, + "bottom": null, + "display": null, + "flex": null, + "flex_flow": null, + "grid_area": null, + "grid_auto_columns": null, + "grid_auto_flow": null, + "grid_auto_rows": null, + "grid_column": null, + "grid_gap": null, + "grid_row": null, + "grid_template_areas": null, + "grid_template_columns": null, + "grid_template_rows": null, + "height": null, + "justify_content": null, + "justify_items": null, + "left": null, + "margin": null, + "max_height": null, + "max_width": null, + "min_height": null, + "min_width": null, + "object_fit": null, + "object_position": null, + "order": null, + "overflow": null, + "overflow_x": null, + "overflow_y": null, + "padding": null, + "right": null, + "top": null, + "visibility": null, + "width": null + } }, - { - "cell_type": "code", - "execution_count": 11, - "metadata": { - "colab": { - "base_uri": "https://localhost:8080/", - "height": 1000 - }, - "id": "f6NFmpjEO-qb", - "outputId": "89fd1b48-5189-4401-9cf8-15f55c503676" - }, - "outputs": [ - { - "data": { - "image/jpeg": "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", - "text/plain": [ - "" - ] - }, - "metadata": {}, - "output_type": "display_data" - }, - { - "data": { - "text/plain": [] - }, - "execution_count": 11, - "metadata": {}, - "output_type": "execute_result" - } + "6b8334e071a3438397ba6435aac69f58": { + "model_module": "@jupyter-widgets/controls", + "model_module_version": "1.5.0", + "model_name": "ProgressStyleModel", + "state": { + "_model_module": "@jupyter-widgets/controls", + "_model_module_version": "1.5.0", + "_model_name": "ProgressStyleModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/base", + "_view_module_version": "1.2.0", + "_view_name": "StyleView", + "bar_color": null, + "description_width": "" + } + }, + "7d482188c12d4a7886f20a65d3402c59": { + "model_module": "@jupyter-widgets/controls", + "model_module_version": "1.5.0", + "model_name": "HBoxModel", + "state": { + "_dom_classes": [], + "_model_module": "@jupyter-widgets/controls", + "_model_module_version": "1.5.0", + "_model_name": "HBoxModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/controls", + "_view_module_version": "1.5.0", + "_view_name": "HBoxView", + "box_style": "", + "children": [ + "IPY_MODEL_2a3ec74419ae4a02ac0210db66133415", + "IPY_MODEL_ddeff9a822404adbbc3cad97a939bc0c", + "IPY_MODEL_36d341ab3a044709b5af2e8ab97559bc" ], - "source": [ - "from haystack import Pipeline\n", - "\n", - "basic_rag_pipeline = Pipeline()\n", - "# Add components to your pipeline\n", - "basic_rag_pipeline.add_component(\"text_embedder\", text_embedder)\n", - "basic_rag_pipeline.add_component(\"retriever\", retriever)\n", - "basic_rag_pipeline.add_component(\"prompt_builder\", prompt_builder)\n", - "basic_rag_pipeline.add_component(\"llm\", generator)\n", - "\n", - "# Now, connect the components to each other\n", - "basic_rag_pipeline.connect(\"text_embedder.embedding\", \"retriever.query_embedding\")\n", - "basic_rag_pipeline.connect(\"retriever\", \"prompt_builder.documents\")\n", - "basic_rag_pipeline.connect(\"prompt_builder\", \"llm\")" - ] + "layout": "IPY_MODEL_88fc33e1ab78405e911b5eafa512c935" + } }, - { - "cell_type": "markdown", - "metadata": { - "id": "6NqyLhx7O-qc" - }, - "source": [ - "That's it! Your RAG pipeline is ready to generate answers to questions!" - ] + "7fdb2c859e454e72888709a835f7591e": { + "model_module": "@jupyter-widgets/base", + "model_module_version": "1.2.0", + "model_name": "LayoutModel", + "state": { + "_model_module": "@jupyter-widgets/base", + "_model_module_version": "1.2.0", + "_model_name": "LayoutModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/base", + "_view_module_version": "1.2.0", + "_view_name": "LayoutView", + "align_content": null, + "align_items": null, + "align_self": null, + "border": null, + "bottom": null, + "display": null, + "flex": null, + "flex_flow": null, + "grid_area": null, + "grid_auto_columns": null, + "grid_auto_flow": null, + "grid_auto_rows": null, + "grid_column": null, + "grid_gap": null, + "grid_row": null, + "grid_template_areas": null, + "grid_template_columns": null, + "grid_template_rows": null, + "height": null, + "justify_content": null, + "justify_items": null, + "left": null, + "margin": null, + "max_height": null, + "max_width": null, + "min_height": null, + "min_width": null, + "object_fit": null, + "object_position": null, + "order": null, + "overflow": null, + "overflow_x": null, + "overflow_y": null, + "padding": null, + "right": null, + "top": null, + "visibility": null, + "width": null + } }, - { - "cell_type": "markdown", - "metadata": { - "id": "DBAyF5tVO-qc" - }, - "source": [ - "## Asking a Question\n", - "\n", - "When asking a question, use the `run()` method of the pipeline. Make sure to provide the question to both the `text_embedder` and the `prompt_builder`. This ensures that the `{{question}}` variable in the template prompt gets replaced with your specific question." - ] + "88fc33e1ab78405e911b5eafa512c935": { + "model_module": "@jupyter-widgets/base", + "model_module_version": "1.2.0", + "model_name": "LayoutModel", + "state": { + "_model_module": "@jupyter-widgets/base", + "_model_module_version": "1.2.0", + "_model_name": "LayoutModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/base", + "_view_module_version": "1.2.0", + "_view_name": "LayoutView", + "align_content": null, + "align_items": null, + "align_self": null, + "border": null, + "bottom": null, + "display": null, + "flex": null, + "flex_flow": null, + "grid_area": null, + "grid_auto_columns": null, + "grid_auto_flow": null, + "grid_auto_rows": null, + "grid_column": null, + "grid_gap": null, + "grid_row": null, + "grid_template_areas": null, + "grid_template_columns": null, + "grid_template_rows": null, + "height": null, + "justify_content": null, + "justify_items": null, + "left": null, + "margin": null, + "max_height": null, + "max_width": null, + "min_height": null, + "min_width": null, + "object_fit": null, + "object_position": null, + "order": null, + "overflow": null, + "overflow_x": null, + "overflow_y": null, + "padding": null, + "right": null, + "top": null, + "visibility": null, + "width": null + } }, - { - "cell_type": "code", - "execution_count": 13, - "metadata": { - "colab": { - "base_uri": "https://localhost:8080/", - "height": 86, - "referenced_widgets": [ - "4e6e97b6d54f4f80bb7e8b25aba8e616", - "1a820c06a7a049d8b6c9ff300284d06e", - "58ff4e0603a74978a134f63533859be5", - "8bdb8bfae31d4f4cb6c3b0bf43120eed", - "39a68d9a5c274e2dafaa2d1f86eea768", - "d0cfe5dacdfc431a91b4c4741123e2d0", - "e7f1e1a14bb740d18827dd78bbe7b2e3", - "3fda06f905b445a488efdd2dd08c0939", - "2bc341a780f7498ba9cd475468841bb5", - "d7218475e23b420a8c03d00ca4ab8718", - "a694abaf765f4d1b82fa0138e59c6793" - ] - }, - "id": "Vnt283M5O-qc", - "outputId": "d2843a73-3ad5-4daa-8d1e-a58de7aa2bb0" - }, - "outputs": [ - { - "data": { - "application/vnd.jupyter.widget-view+json": { - "model_id": "4e6e97b6d54f4f80bb7e8b25aba8e616", - "version_major": 2, - "version_minor": 0 - }, - "text/plain": [ - "Batches: 0%| | 0/1 [00:00 This tutorial uses Haystack 2.0. To learn more, read the [Haystack 2.0 announcement](https://haystack.deepset.ai/blog/haystack-2-release) or visit the [Haystack 2.0 Documentation](https://docs.haystack.deepset.ai/docs/intro)..\n", - "\n", - "## Overview\n", - "This tutorial demonstrates how to use Haystack 2.0's advanced [looping pipelines](https://docs.haystack.deepset.ai/docs/pipelines#loops) with LLMs for more dynamic and flexible data processing. You'll learn how to extract structured data from unstructured data using an LLM, and to validate the generated output against a predefined schema.\n", - "\n", - "This tutorial uses `gpt-4o-mini` to change unstructured passages into JSON outputs that follow the [Pydantic](https://github.com/pydantic/pydantic) schema. It uses a custom OutputValidator component to validate the JSON and loop back to make corrections, if necessary." - ] - }, - { - "cell_type": "markdown", - "metadata": { - "id": "jmiAHh1oGsKI" - }, - "source": [ - "## Preparing the Colab Environment\n", - "\n", - "Enable the debug mode of logging:" - ] - }, - { - "cell_type": "code", - "execution_count": null, - "metadata": { - "id": "Vor9IHuNRvEh" - }, - "outputs": [], - "source": [ - "import logging\n", - "\n", - "logging.basicConfig()\n", - "logging.getLogger(\"canals.pipeline.pipeline\").setLevel(logging.DEBUG)" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "id": "ljbWiyJkKiPw" - }, - "source": [ - "## Installing Dependencies\n", - "Install Haystack and [colorama](https://pypi.org/project/colorama/) with pip:" - ] - }, - { - "cell_type": "code", - "execution_count": null, - "metadata": { - "colab": { - "base_uri": "https://localhost:8080/" - }, - "id": "kcc1AlLQd_jI", - "outputId": "efc4bbab-a9fe-46ee-d8af-9d86edacaf04" - }, - "outputs": [], - "source": [ - "%%bash\n", - "\n", - "pip install haystack-ai\n", - "pip install colorama" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "id": "nTA5fdvCLMKD" - }, - "source": [ - "### Enabling Telemetry\n", - "\n", - "Enable telemetry to let us know you're using this tutorial. (You can always opt out by commenting out this line). For details, see [Telemetry](https://docs.haystack.deepset.ai/docs/telemetry)." - ] - }, - { - "cell_type": "code", - "execution_count": null, - "metadata": { - "id": "Apay3QSQLKdM" - }, - "outputs": [], - "source": [ - "from haystack.telemetry import tutorial_running\n", - "\n", - "tutorial_running(28)" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "id": "Cmjfa8CiCeFl" - }, - "source": [ - "## Defining a Schema to Parse the JSON Object\n", - "\n", - "Define a simple JSON schema for the data you want to extract from a text passsage using the LLM. As the first step, define two [Pydantic models](https://docs.pydantic.dev/1.10/usage/models/), `City` and `CitiesData`, with suitable fields and types." - ] - }, - { - "cell_type": "code", - "execution_count": null, - "metadata": { - "id": "xwKrDOOGdaAz" - }, - "outputs": [], - "source": [ - "from typing import List\n", - "from pydantic import BaseModel\n", - "\n", - "\n", - "class City(BaseModel):\n", - " name: str\n", - " country: str\n", - " population: int\n", - "\n", - "\n", - "class CitiesData(BaseModel):\n", - " cities: List[City]" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "id": "zv-6-l_PCeFl" - }, - "source": [ - "> You can change these models according to the format you wish to extract from the text." - ] - }, - { - "cell_type": "markdown", - "metadata": { - "id": "ouk1mAOUCeFl" - }, - "source": [ - "Then, generate a JSON schema from Pydantic models using `schema_json()`. You will later on use this schema in the prompt to instruct the LLM.\n", - "\n", - "To learn more about the JSON schemas, visit [Pydantic Schema](https://docs.pydantic.dev/1.10/usage/schema/). " - ] - }, - { - "cell_type": "code", - "execution_count": null, - "metadata": { - "id": "8Lg9_72jCeFl" - }, - "outputs": [], - "source": [ - "json_schema = CitiesData.schema_json(indent=2)" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "id": "KvNhg0bP7kfg" - }, - "source": [ - "## Creating a Custom Component: OutputValidator\n", - "\n", - "`OutputValidator` is a custom component that validates if the JSON object the LLM generates complies with the provided [Pydantic model](https://docs.pydantic.dev/1.10/usage/models/). If it doesn't, OutputValidator returns an error message along with the incorrect JSON object to get it fixed in the next loop.\n", - "\n", - "For more details about custom components, see [Creating Custom Components](https://docs.haystack.deepset.ai/docs/custom-components)." - ] - }, - { - "cell_type": "code", - "execution_count": null, - "metadata": { - "id": "yr6D8RN2d7Vy" - }, - "outputs": [], - "source": [ - "import json\n", - "import random\n", - "import pydantic\n", - "from pydantic import ValidationError\n", - "from typing import Optional, List\n", - "from colorama import Fore\n", - "from haystack import component\n", - "\n", - "# Define the component input parameters\n", - "@component\n", - "class OutputValidator:\n", - " def __init__(self, pydantic_model: pydantic.BaseModel):\n", - " self.pydantic_model = pydantic_model\n", - " self.iteration_counter = 0\n", - "\n", - " # Define the component output\n", - " @component.output_types(valid_replies=List[str], invalid_replies=Optional[List[str]], error_message=Optional[str])\n", - " def run(self, replies: List[str]):\n", - "\n", - " self.iteration_counter += 1\n", - "\n", - " ## Try to parse the LLM's reply ##\n", - " # If the LLM's reply is a valid object, return `\"valid_replies\"`\n", - " try:\n", - " output_dict = json.loads(replies[0])\n", - " self.pydantic_model.parse_obj(output_dict)\n", - " print(\n", - " Fore.GREEN\n", - " + f\"OutputValidator at Iteration {self.iteration_counter}: Valid JSON from LLM - No need for looping: {replies[0]}\"\n", - " )\n", - " return {\"valid_replies\": replies}\n", - "\n", - " # If the LLM's reply is corrupted or not valid, return \"invalid_replies\" and the \"error_message\" for LLM to try again\n", - " except (ValueError, ValidationError) as e:\n", - " print(\n", - " Fore.RED\n", - " + f\"OutputValidator at Iteration {self.iteration_counter}: Invalid JSON from LLM - Let's try again.\\n\"\n", - " f\"Output from LLM:\\n {replies[0]} \\n\"\n", - " f\"Error from OutputValidator: {e}\"\n", - " )\n", - " return {\"invalid_replies\": replies, \"error_message\": str(e)}" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "id": "vQ_TfSBkCeFm" - }, - "source": [ - "Then, create an OutputValidator instance with `CitiesData` that you have created before." - ] - }, - { - "cell_type": "code", - "execution_count": null, - "metadata": { - "id": "bhPCLCBCCeFm" - }, - "outputs": [], - "source": [ - "output_validator = OutputValidator(pydantic_model=CitiesData)" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "id": "xcIWKjW4k42r" - }, - "source": [ - "## Creating the Prompt\n", - "\n", - "Write instructions for the LLM for converting a passage into a JSON format. Ensure the instructions explain how to identify and correct errors if the JSON doesn't match the required schema. Once you create the prompt, initialize PromptBuilder to use it. \n", - "\n", - "For information about Jinja2 template and PromptBuilder, see [PromptBuilder](https://docs.haystack.deepset.ai/docs/promptbuilder)." - ] - }, - { - "cell_type": "code", - "execution_count": null, - "metadata": { - "id": "ohPpNALjdVKt" - }, - "outputs": [], - "source": [ - "from haystack.components.builders import PromptBuilder\n", - "\n", - "prompt_template = \"\"\"\n", - "Create a JSON object from the information present in this passage: {{passage}}.\n", - "Only use information that is present in the passage. Follow this JSON schema, but only return the actual instances without any additional schema definition:\n", - "{{schema}}\n", - "Make sure your response is a dict and not a list.\n", - "{% if invalid_replies and error_message %}\n", - " You already created the following output in a previous attempt: {{invalid_replies}}\n", - " However, this doesn't comply with the format requirements from above and triggered this Python exception: {{error_message}}\n", - " Correct the output and try again. Just return the corrected output without any extra explanations.\n", - "{% endif %}\n", - "\"\"\"\n", - "prompt_builder = PromptBuilder(template=prompt_template)" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "id": "KM9-Zq2FL7Nn" - }, - "source": [ - "## Initalizing the Generator\n", - "\n", - "[OpenAIGenerator](https://docs.haystack.deepset.ai/docs/openaigenerator) generates\n", - "text using OpenAI's `gpt-4o-mini` model by default. Set the `OPENAI_API_KEY` variable and provide a model name to the Generator." - ] - }, - { - "cell_type": "code", - "execution_count": null, - "metadata": { - "id": "Z4cQteIgunUR" - }, - "outputs": [], - "source": [ - "import os\n", - "from getpass import getpass\n", - "\n", - "from haystack.components.generators import OpenAIGenerator\n", - "\n", - "if \"OPENAI_API_KEY\" not in os.environ:\n", - " os.environ[\"OPENAI_API_KEY\"] = getpass(\"Enter OpenAI API key:\")\n", - "generator = OpenAIGenerator()" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "id": "zbotIOgXHkC5" - }, - "source": [ - "## Building the Pipeline\n", - "\n", - "Add all components to your pipeline and connect them. Add connections from `output_validator` back to the `prompt_builder` for cases where the produced JSON doesn't comply with the JSON schema. Set `max_runs_per_component` to avoid infinite looping." - ] - }, - { - "cell_type": "code", - "execution_count": null, - "metadata": { - "id": "eFglN9YEv-1W" - }, - "outputs": [], - "source": [ - "from haystack import Pipeline\n", - "\n", - "pipeline = Pipeline(max_runs_per_component=5)\n", - "\n", - "# Add components to your pipeline\n", - "pipeline.add_component(instance=prompt_builder, name=\"prompt_builder\")\n", - "pipeline.add_component(instance=generator, name=\"llm\")\n", - "pipeline.add_component(instance=output_validator, name=\"output_validator\")\n", - "\n", - "# Now, connect the components to each other\n", - "pipeline.connect(\"prompt_builder\", \"llm\")\n", - "pipeline.connect(\"llm\", \"output_validator\")\n", - "# If a component has more than one output or input, explicitly specify the connections:\n", - "pipeline.connect(\"output_validator.invalid_replies\", \"prompt_builder.invalid_replies\")\n", - "pipeline.connect(\"output_validator.error_message\", \"prompt_builder.error_message\")" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "id": "-UKW5wtIIT7w" - }, - "source": [ - "### Visualize the Pipeline\n", - "\n", - "Draw the pipeline with the [`draw()`](https://docs.haystack.deepset.ai/docs/drawing-pipeline-graphs) method to confirm the connections are correct. You can find the diagram in the Files section of this Colab." - ] - }, - { - "cell_type": "code", - "execution_count": null, - "metadata": { - "id": "RZJg6YHId300" - }, - "outputs": [], - "source": [ - "pipeline.draw(\"auto-correct-pipeline.png\")" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "id": "kV_kexTjImpo" - }, - "source": [ - "## Testing the Pipeline\n", - "\n", - "Run the pipeline with an example passage that you want to convert into a JSON format and the `json_schema` you have created for `CitiesData`. For the given example passage, the generated JSON object should be like:\n", - "```json\n", - "{\n", - " \"cities\": [\n", - " {\n", - " \"name\": \"Berlin\",\n", - " \"country\": \"Germany\",\n", - " \"population\": 3850809\n", - " },\n", - " {\n", - " \"name\": \"Paris\",\n", - " \"country\": \"France\",\n", - " \"population\": 2161000\n", - " },\n", - " {\n", - " \"name\": \"Lisbon\",\n", - " \"country\": \"Portugal\",\n", - " \"population\": 504718\n", - " }\n", - " ]\n", - "}\n", - "```\n", - "The output of the LLM should be compliant with the `json_schema`. If the LLM doesn't generate the correct JSON object, it will loop back and try again." - ] - }, - { - "cell_type": "code", - "execution_count": null, - "metadata": { - "colab": { - "base_uri": "https://localhost:8080/" - }, - "id": "yIoMedb6eKia", - "outputId": "4a9ef924-cf26-4908-d83f-b0bc0dc03b54" - }, - "outputs": [], - "source": [ - "passage = \"Berlin is the capital of Germany. It has a population of 3,850,809. Paris, France's capital, has 2.161 million residents. Lisbon is the capital and the largest city of Portugal with the population of 504,718.\"\n", - "result = pipeline.run({\"prompt_builder\": {\"passage\": passage, \"schema\": json_schema}})" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "id": "WWxmPgADS_Fa" - }, - "source": [ - "> If you encounter `PipelineMaxLoops: Maximum loops count (5) exceeded for component 'prompt_builder'.` error, consider increasing the maximum loop count or simply rerun the pipeline." - ] - }, - { - "cell_type": "markdown", - "metadata": { - "id": "eWPawSjgSJAM" - }, - "source": [ - "### Print the Correct JSON\n", - "If you didn't get any error, you can now print the corrected JSON." - ] + "cells": [ + { + "cell_type": "markdown", + "metadata": { + "id": "AVBtOVlNJ51C" + }, + "source": [ + "# Tutorial: Generating Structured Output with Loop-Based Auto-Correction\n", + "\n", + "- **Level**: Intermediate\n", + "- **Time to complete**: 15 minutes\n", + "- **Prerequisites**: You must have an API key from an active OpenAI account as this tutorial is using the gpt-4o-mini model by OpenAI.\n", + "- **Components Used**: `PromptBuilder`, `OpenAIChatGenerator`, `OutputValidator` (Custom component)\n", + "- **Goal**: After completing this tutorial, you will have built a system that extracts unstructured data, puts it in a JSON schema, and automatically corrects errors in the JSON output from a large language model (LLM) to make sure it follows the specified structure.\n", + "\n", + "> This tutorial uses Haystack 2.0. To learn more, read the [Haystack 2.0 announcement](https://haystack.deepset.ai/blog/haystack-2-release) or visit the [Haystack 2.0 Documentation](https://docs.haystack.deepset.ai/docs/intro)..\n", + "\n", + "## Overview\n", + "This tutorial demonstrates how to use Haystack 2.0's advanced [looping pipelines](https://docs.haystack.deepset.ai/docs/pipelines#loops) with LLMs for more dynamic and flexible data processing. You'll learn how to extract structured data from unstructured data using an LLM, and to validate the generated output against a predefined schema.\n", + "\n", + "This tutorial uses `gpt-4o-mini` to change unstructured passages into JSON outputs that follow the [Pydantic](https://github.com/pydantic/pydantic) schema. It uses a custom OutputValidator component to validate the JSON and loop back to make corrections, if necessary." + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "jmiAHh1oGsKI" + }, + "source": [ + "## Preparing the Colab Environment\n", + "\n", + "Enable the debug mode of logging:" + ] + }, + { + "cell_type": "code", + "execution_count": 1, + "metadata": { + "id": "Vor9IHuNRvEh" + }, + "outputs": [], + "source": [ + "import logging\n", + "\n", + "logging.basicConfig()\n", + "logging.getLogger(\"canals.pipeline.pipeline\").setLevel(logging.DEBUG)" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "ljbWiyJkKiPw" + }, + "source": [ + "## Installing Dependencies\n", + "Install Haystack and [colorama](https://pypi.org/project/colorama/) with pip:" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": { + "colab": { + "base_uri": "https://localhost:8080/" }, + "id": "kcc1AlLQd_jI", + "outputId": "efc4bbab-a9fe-46ee-d8af-9d86edacaf04" + }, + "outputs": [], + "source": [ + "%%bash\n", + "\n", + "pip install haystack-ai\n", + "pip install colorama" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "nTA5fdvCLMKD" + }, + "source": [ + "### Enabling Telemetry\n", + "\n", + "Enable telemetry to let us know you're using this tutorial. (You can always opt out by commenting out this line). For details, see [Telemetry](https://docs.haystack.deepset.ai/docs/telemetry)." + ] + }, + { + "cell_type": "code", + "execution_count": 3, + "metadata": { + "id": "Apay3QSQLKdM" + }, + "outputs": [ { - "cell_type": "code", - "execution_count": null, - "metadata": { - "colab": { - "base_uri": "https://localhost:8080/" - }, - "id": "BVO47gXQQnDC", - "outputId": "460a10d4-a69a-49cd-bbb2-fc4980907299" - }, - "outputs": [], - "source": [ - "valid_reply = result[\"output_validator\"][\"valid_replies\"][0]\n", - "valid_json = json.loads(valid_reply)\n", - "print(valid_json)" - ] - }, + "name": "stderr", + "output_type": "stream", + "text": [ + "/Users/amna.mubashar/Library/Python/3.9/lib/python/site-packages/tqdm/auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html\n", + " from .autonotebook import tqdm as notebook_tqdm\n" + ] + } + ], + "source": [ + "from haystack.telemetry import tutorial_running\n", + "\n", + "tutorial_running(28)" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "Cmjfa8CiCeFl" + }, + "source": [ + "## Defining a Schema to Parse the JSON Object\n", + "\n", + "Define a simple JSON schema for the data you want to extract from a text passsage using the LLM. As the first step, define two [Pydantic models](https://docs.pydantic.dev/1.10/usage/models/), `City` and `CitiesData`, with suitable fields and types." + ] + }, + { + "cell_type": "code", + "execution_count": 4, + "metadata": { + "id": "xwKrDOOGdaAz" + }, + "outputs": [], + "source": [ + "from typing import List\n", + "from pydantic import BaseModel\n", + "\n", + "\n", + "class City(BaseModel):\n", + " name: str\n", + " country: str\n", + " population: int\n", + "\n", + "\n", + "class CitiesData(BaseModel):\n", + " cities: List[City]" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "zv-6-l_PCeFl" + }, + "source": [ + "> You can change these models according to the format you wish to extract from the text." + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "ouk1mAOUCeFl" + }, + "source": [ + "Then, generate a JSON schema from Pydantic models using `schema_json()`. You will later on use this schema in the prompt to instruct the LLM.\n", + "\n", + "To learn more about the JSON schemas, visit [Pydantic Schema](https://docs.pydantic.dev/1.10/usage/schema/). " + ] + }, + { + "cell_type": "code", + "execution_count": 5, + "metadata": { + "id": "8Lg9_72jCeFl" + }, + "outputs": [], + "source": [ + "json_schema = CitiesData.schema_json(indent=2)" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "KvNhg0bP7kfg" + }, + "source": [ + "## Creating a Custom Component: OutputValidator\n", + "\n", + "`OutputValidator` is a custom component that validates if the JSON object the LLM generates complies with the provided [Pydantic model](https://docs.pydantic.dev/1.10/usage/models/). If it doesn't, OutputValidator returns an error message along with the incorrect JSON object to get it fixed in the next loop.\n", + "\n", + "For more details about custom components, see [Creating Custom Components](https://docs.haystack.deepset.ai/docs/custom-components)." + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": { + "id": "yr6D8RN2d7Vy" + }, + "outputs": [], + "source": [ + "import json\n", + "import random\n", + "import pydantic\n", + "from pydantic import ValidationError\n", + "from typing import Optional, List\n", + "from colorama import Fore\n", + "from haystack import component\n", + "from haystack.dataclasses import ChatMessage\n", + "\n", + "\n", + "# Define the component input parameters\n", + "@component\n", + "class OutputValidator:\n", + " def __init__(self, pydantic_model: pydantic.BaseModel):\n", + " self.pydantic_model = pydantic_model\n", + " self.iteration_counter = 0\n", + "\n", + " # Define the component output\n", + " @component.output_types(valid_replies=List[str], invalid_replies=Optional[List[str]], error_message=Optional[str])\n", + " def run(self, replies: List[ChatMessage]):\n", + "\n", + " self.iteration_counter += 1\n", + "\n", + " ## Try to parse the LLM's reply ##\n", + " # If the LLM's reply is a valid object, return `\"valid_replies\"`\n", + " try:\n", + " output_dict = json.loads(replies[0].text)\n", + " self.pydantic_model.parse_obj(output_dict)\n", + " print(\n", + " Fore.GREEN\n", + " + f\"OutputValidator at Iteration {self.iteration_counter}: Valid JSON from LLM - No need for looping: {replies[0]}\"\n", + " )\n", + " return {\"valid_replies\": replies}\n", + "\n", + " # If the LLM's reply is corrupted or not valid, return \"invalid_replies\" and the \"error_message\" for LLM to try again\n", + " except (ValueError, ValidationError) as e:\n", + " print(\n", + " Fore.RED\n", + " + f\"OutputValidator at Iteration {self.iteration_counter}: Invalid JSON from LLM - Let's try again.\\n\"\n", + " f\"Output from LLM:\\n {replies[0]} \\n\"\n", + " f\"Error from OutputValidator: {e}\"\n", + " )\n", + " return {\"invalid_replies\": replies, \"error_message\": str(e)}" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "vQ_TfSBkCeFm" + }, + "source": [ + "Then, create an OutputValidator instance with `CitiesData` that you have created before." + ] + }, + { + "cell_type": "code", + "execution_count": 7, + "metadata": { + "id": "bhPCLCBCCeFm" + }, + "outputs": [], + "source": [ + "output_validator = OutputValidator(pydantic_model=CitiesData)" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "xcIWKjW4k42r" + }, + "source": [ + "## Creating the Prompt\n", + "\n", + "Write instructions for the LLM for converting a passage into a JSON format. Ensure the instructions explain how to identify and correct errors if the JSON doesn't match the required schema. Once you create the prompt, initialize PromptBuilder to use it. \n", + "\n", + "For information about Jinja2 template and ChatPromptBuilder, see [ChatPromptBuilder](https://docs.haystack.deepset.ai/docs/chatpromptbuilder)." + ] + }, + { + "cell_type": "code", + "execution_count": 8, + "metadata": { + "id": "ohPpNALjdVKt" + }, + "outputs": [], + "source": [ + "from haystack.components.builders import ChatPromptBuilder\n", + "\n", + "\n", + "prompt_template = [ChatMessage.from_user(\"\"\"\n", + "Create a JSON object from the information present in this passage: {{passage}}.\n", + "Only use information that is present in the passage. Follow this JSON schema, but only return the actual instances without any additional schema definition:\n", + "{{schema}}\n", + "Make sure your response is a dict and not a list.\n", + "{% if invalid_replies and error_message %}\n", + " You already created the following output in a previous attempt: {{invalid_replies}}\n", + " However, this doesn't comply with the format requirements from above and triggered this Python exception: {{error_message}}\n", + " Correct the output and try again. Just return the corrected output without any extra explanations.\n", + "{% endif %}\n", + "\"\"\")]\n", + "prompt_builder = ChatPromptBuilder(template=prompt_template)" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "KM9-Zq2FL7Nn" + }, + "source": [ + "## Initalizing the ChatGenerator\n", + "\n", + "[OpenAIChatGenerator](https://docs.haystack.deepset.ai/docs/openaichatgenerator) generates\n", + "text using OpenAI's `gpt-4o-mini` model by default. Set the `OPENAI_API_KEY` variable and provide a model name to the ChatGenerator." + ] + }, + { + "cell_type": "code", + "execution_count": 9, + "metadata": { + "id": "Z4cQteIgunUR" + }, + "outputs": [], + "source": [ + "import os\n", + "from getpass import getpass\n", + "\n", + "from haystack.components.generators.chat import OpenAIChatGenerator\n", + "\n", + "if \"OPENAI_API_KEY\" not in os.environ:\n", + " os.environ[\"OPENAI_API_KEY\"] = getpass(\"Enter OpenAI API key:\")\n", + "chat_generator = OpenAIChatGenerator()" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "zbotIOgXHkC5" + }, + "source": [ + "## Building the Pipeline\n", + "\n", + "Add all components to your pipeline and connect them. Add connections from `output_validator` back to the `prompt_builder` for cases where the produced JSON doesn't comply with the JSON schema. Set `max_runs_per_component` to avoid infinite looping." + ] + }, + { + "cell_type": "code", + "execution_count": 10, + "metadata": { + "id": "eFglN9YEv-1W" + }, + "outputs": [ { - "cell_type": "markdown", - "metadata": { - "id": "Egz_4h2vI_QL" - }, - "source": [ - "## What's next\n", - "\n", - "🎉 Congratulations! You've built a system that generates structured JSON out of unstructured text passages, and auto-corrects it by using the looping functionality of Haystack pipelines.\n", - "\n", - "To stay up to date on the latest Haystack developments, you can [subscribe to our newsletter](https://landing.deepset.ai/haystack-community-updates) and [join Haystack discord community](https://discord.gg/haystack).\n", - "\n", - "Thanks for reading!" + "data": { + "text/plain": [ + "\n", + "🚅 Components\n", + " - prompt_builder: ChatPromptBuilder\n", + " - llm: OpenAIChatGenerator\n", + " - output_validator: OutputValidator\n", + "🛤️ Connections\n", + " - prompt_builder.prompt -> llm.messages (List[ChatMessage])\n", + " - llm.replies -> output_validator.replies (List[ChatMessage])\n", + " - output_validator.invalid_replies -> prompt_builder.invalid_replies (Optional[List[str]])\n", + " - output_validator.error_message -> prompt_builder.error_message (Optional[str])" ] + }, + "execution_count": 10, + "metadata": {}, + "output_type": "execute_result" } - ], - "metadata": { - "accelerator": "GPU", + ], + "source": [ + "from haystack import Pipeline\n", + "\n", + "pipeline = Pipeline(max_runs_per_component=5)\n", + "\n", + "# Add components to your pipeline\n", + "pipeline.add_component(instance=prompt_builder, name=\"prompt_builder\")\n", + "pipeline.add_component(instance=chat_generator, name=\"llm\")\n", + "pipeline.add_component(instance=output_validator, name=\"output_validator\")\n", + "\n", + "# Now, connect the components to each other\n", + "pipeline.connect(\"prompt_builder.prompt\", \"llm.messages\")\n", + "pipeline.connect(\"llm.replies\", \"output_validator\")\n", + "# If a component has more than one output or input, explicitly specify the connections:\n", + "pipeline.connect(\"output_validator.invalid_replies\", \"prompt_builder.invalid_replies\")\n", + "pipeline.connect(\"output_validator.error_message\", \"prompt_builder.error_message\")" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "-UKW5wtIIT7w" + }, + "source": [ + "### Visualize the Pipeline\n", + "\n", + "Draw the pipeline with the [`draw()`](https://docs.haystack.deepset.ai/docs/drawing-pipeline-graphs) method to confirm the connections are correct. You can find the diagram in the Files section of this Colab." + ] + }, + { + "cell_type": "code", + "execution_count": 11, + "metadata": { + "id": "RZJg6YHId300" + }, + "outputs": [], + "source": [ + "pipeline.draw(\"auto-correct-pipeline.png\")" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "kV_kexTjImpo" + }, + "source": [ + "## Testing the Pipeline\n", + "\n", + "Run the pipeline with an example passage that you want to convert into a JSON format and the `json_schema` you have created for `CitiesData`. For the given example passage, the generated JSON object should be like:\n", + "```json\n", + "{\n", + " \"cities\": [\n", + " {\n", + " \"name\": \"Berlin\",\n", + " \"country\": \"Germany\",\n", + " \"population\": 3850809\n", + " },\n", + " {\n", + " \"name\": \"Paris\",\n", + " \"country\": \"France\",\n", + " \"population\": 2161000\n", + " },\n", + " {\n", + " \"name\": \"Lisbon\",\n", + " \"country\": \"Portugal\",\n", + " \"population\": 504718\n", + " }\n", + " ]\n", + "}\n", + "```\n", + "The output of the LLM should be compliant with the `json_schema`. If the LLM doesn't generate the correct JSON object, it will loop back and try again." + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": { "colab": { - "gpuType": "T4", - "provenance": [] + "base_uri": "https://localhost:8080/" }, - "kernelspec": { - "display_name": "Python 3", - "name": "python3" + "id": "yIoMedb6eKia", + "outputId": "4a9ef924-cf26-4908-d83f-b0bc0dc03b54" + }, + "outputs": [], + "source": [ + "passage = \"Berlin is the capital of Germany. It has a population of 3,850,809. Paris, France's capital, has 2.161 million residents. Lisbon is the capital and the largest city of Portugal with the population of 504,718.\"\n", + "result = pipeline.run({\"prompt_builder\": {\"passage\": passage, \"schema\": json_schema}})" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "WWxmPgADS_Fa" + }, + "source": [ + "> If you encounter `PipelineMaxLoops: Maximum loops count (5) exceeded for component 'prompt_builder'.` error, consider increasing the maximum loop count or simply rerun the pipeline." + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "eWPawSjgSJAM" + }, + "source": [ + "### Print the Correct JSON\n", + "If you didn't get any error, you can now print the corrected JSON." + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": { + "colab": { + "base_uri": "https://localhost:8080/" }, - "language_info": { - "name": "python" + "id": "BVO47gXQQnDC", + "outputId": "460a10d4-a69a-49cd-bbb2-fc4980907299" + }, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "{'cities': [{'name': 'Berlin', 'country': 'Germany', 'population': 3850809}, {'name': 'Paris', 'country': 'France', 'population': 2161000}, {'name': 'Lisbon', 'country': 'Portugal', 'population': 504718}]}\n" + ] } + ], + "source": [ + "valid_reply = result[\"output_validator\"][\"valid_replies\"][0].text\n", + "valid_json = json.loads(valid_reply)\n", + "print(valid_json)" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "Egz_4h2vI_QL" + }, + "source": [ + "## What's next\n", + "\n", + "🎉 Congratulations! You've built a system that generates structured JSON out of unstructured text passages, and auto-corrects it by using the looping functionality of Haystack pipelines.\n", + "\n", + "To stay up to date on the latest Haystack developments, you can [subscribe to our newsletter](https://landing.deepset.ai/haystack-community-updates) and [join Haystack discord community](https://discord.gg/haystack).\n", + "\n", + "Thanks for reading!" + ] + } + ], + "metadata": { + "accelerator": "GPU", + "colab": { + "gpuType": "T4", + "provenance": [] + }, + "kernelspec": { + "display_name": "Python 3", + "name": "python3" }, - "nbformat": 4, - "nbformat_minor": 0 + "language_info": { + "codemirror_mode": { + "name": "ipython", + "version": 3 + }, + "file_extension": ".py", + "mimetype": "text/x-python", + "name": "python", + "nbconvert_exporter": "python", + "pygments_lexer": "ipython3", + "version": "3.9.6" + } + }, + "nbformat": 4, + "nbformat_minor": 0 } diff --git a/tutorials/29_Serializing_Pipelines.ipynb b/tutorials/29_Serializing_Pipelines.ipynb index 9e6d3493..b98eaf2c 100644 --- a/tutorials/29_Serializing_Pipelines.ipynb +++ b/tutorials/29_Serializing_Pipelines.ipynb @@ -10,7 +10,7 @@ "\n", "- **Level**: Beginner\n", "- **Time to complete**: 10 minutes\n", - "- **Components Used**: [`HuggingFaceLocalGenerator`](https://docs.haystack.deepset.ai/docs/huggingfacelocalgenerator), [`PromptBuilder`](https://docs.haystack.deepset.ai/docs/promptbuilder)\n", + "- **Components Used**: [`HuggingFaceLocalChatGenerator`](https://docs.haystack.deepset.ai/docs/huggingfacelocalchatgenerator), [`ChatPromptBuilder`](https://docs.haystack.deepset.ai/docs/chatpromptbuilder)\n", "- **Prerequisites**: None\n", "- **Goal**: After completing this tutorial, you'll understand how to serialize and deserialize between YAML and Python code.\n", "\n", @@ -65,48 +65,7 @@ "id": "CagzMFdkeBBp", "outputId": "e304450a-24e3-4ef8-e642-1fbb573e7d55" }, - "outputs": [ - { - "name": "stdout", - "output_type": "stream", - "text": [ - "Requirement already satisfied: haystack-ai in /usr/local/lib/python3.10/dist-packages (2.0.0b5)\n", - "Requirement already satisfied: boilerpy3 in /usr/local/lib/python3.10/dist-packages (from haystack-ai) (1.0.7)\n", - "Requirement already satisfied: haystack-bm25 in /usr/local/lib/python3.10/dist-packages (from haystack-ai) (1.0.2)\n", - "Requirement already satisfied: jinja2 in /usr/local/lib/python3.10/dist-packages (from haystack-ai) (3.1.3)\n", - "Requirement already satisfied: lazy-imports in /usr/local/lib/python3.10/dist-packages (from haystack-ai) (0.3.1)\n", - "Requirement already satisfied: more-itertools in /usr/local/lib/python3.10/dist-packages (from haystack-ai) (10.1.0)\n", - "Requirement already satisfied: networkx in /usr/local/lib/python3.10/dist-packages (from haystack-ai) (3.2.1)\n", - "Requirement already satisfied: openai>=1.1.0 in /usr/local/lib/python3.10/dist-packages (from haystack-ai) (1.10.0)\n", - "Requirement already satisfied: pandas in /usr/local/lib/python3.10/dist-packages (from haystack-ai) (1.5.3)\n", - "Requirement already satisfied: posthog in /usr/local/lib/python3.10/dist-packages (from haystack-ai) (3.3.3)\n", - "Requirement already satisfied: pyyaml in /usr/local/lib/python3.10/dist-packages (from haystack-ai) (6.0.1)\n", - "Requirement already satisfied: tenacity in /usr/local/lib/python3.10/dist-packages (from haystack-ai) (8.2.3)\n", - "Requirement already satisfied: tqdm in /usr/local/lib/python3.10/dist-packages (from haystack-ai) (4.66.1)\n", - "Requirement already satisfied: typing-extensions in /usr/local/lib/python3.10/dist-packages (from haystack-ai) (4.9.0)\n", - "Requirement already satisfied: anyio<5,>=3.5.0 in /usr/local/lib/python3.10/dist-packages (from openai>=1.1.0->haystack-ai) (3.7.1)\n", - "Requirement already satisfied: distro<2,>=1.7.0 in /usr/lib/python3/dist-packages (from openai>=1.1.0->haystack-ai) (1.7.0)\n", - "Requirement already satisfied: httpx<1,>=0.23.0 in /usr/local/lib/python3.10/dist-packages (from openai>=1.1.0->haystack-ai) (0.26.0)\n", - "Requirement already satisfied: pydantic<3,>=1.9.0 in /usr/local/lib/python3.10/dist-packages (from openai>=1.1.0->haystack-ai) (1.10.14)\n", - "Requirement already satisfied: sniffio in /usr/local/lib/python3.10/dist-packages (from openai>=1.1.0->haystack-ai) (1.3.0)\n", - "Requirement already satisfied: numpy in /usr/local/lib/python3.10/dist-packages (from haystack-bm25->haystack-ai) (1.23.5)\n", - "Requirement already satisfied: MarkupSafe>=2.0 in /usr/local/lib/python3.10/dist-packages (from jinja2->haystack-ai) (2.1.4)\n", - "Requirement already satisfied: python-dateutil>=2.8.1 in /usr/local/lib/python3.10/dist-packages (from pandas->haystack-ai) (2.8.2)\n", - "Requirement already satisfied: pytz>=2020.1 in /usr/local/lib/python3.10/dist-packages (from pandas->haystack-ai) (2023.3.post1)\n", - "Requirement already satisfied: requests<3.0,>=2.7 in /usr/local/lib/python3.10/dist-packages (from posthog->haystack-ai) (2.31.0)\n", - "Requirement already satisfied: six>=1.5 in /usr/local/lib/python3.10/dist-packages (from posthog->haystack-ai) (1.16.0)\n", - "Requirement already satisfied: monotonic>=1.5 in /usr/local/lib/python3.10/dist-packages (from posthog->haystack-ai) (1.6)\n", - "Requirement already satisfied: backoff>=1.10.0 in /usr/local/lib/python3.10/dist-packages (from posthog->haystack-ai) (2.2.1)\n", - "Requirement already satisfied: idna>=2.8 in /usr/local/lib/python3.10/dist-packages (from anyio<5,>=3.5.0->openai>=1.1.0->haystack-ai) (3.6)\n", - "Requirement already satisfied: exceptiongroup in /usr/local/lib/python3.10/dist-packages (from anyio<5,>=3.5.0->openai>=1.1.0->haystack-ai) (1.2.0)\n", - "Requirement already satisfied: certifi in /usr/local/lib/python3.10/dist-packages (from httpx<1,>=0.23.0->openai>=1.1.0->haystack-ai) (2023.11.17)\n", - "Requirement already satisfied: httpcore==1.* in /usr/local/lib/python3.10/dist-packages (from httpx<1,>=0.23.0->openai>=1.1.0->haystack-ai) (1.0.2)\n", - "Requirement already satisfied: h11<0.15,>=0.13 in /usr/local/lib/python3.10/dist-packages (from httpcore==1.*->httpx<1,>=0.23.0->openai>=1.1.0->haystack-ai) (0.14.0)\n", - "Requirement already satisfied: charset-normalizer<4,>=2 in /usr/local/lib/python3.10/dist-packages (from requests<3.0,>=2.7->posthog->haystack-ai) (3.3.2)\n", - "Requirement already satisfied: urllib3<3,>=1.21.1 in /usr/local/lib/python3.10/dist-packages (from requests<3.0,>=2.7->posthog->haystack-ai) (2.0.7)\n" - ] - } - ], + "outputs": [], "source": [ "%%bash\n", "\n", @@ -126,11 +85,22 @@ }, { "cell_type": "code", - "execution_count": null, + "execution_count": 2, "metadata": { "id": "ikIM1o9cHNcS" }, - "outputs": [], + "outputs": [ + { + "name": "stderr", + "output_type": "stream", + "text": [ + "/Users/amna.mubashar/Library/Python/3.9/lib/python/site-packages/urllib3/__init__.py:35: NotOpenSSLWarning: urllib3 v2 only supports OpenSSL 1.1.1+, currently the 'ssl' module is compiled with 'LibreSSL 2.8.3'. See: https://github.com/urllib3/urllib3/issues/3020\n", + " warnings.warn(\n", + "/Users/amna.mubashar/Library/Python/3.9/lib/python/site-packages/tqdm/auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html\n", + " from .autonotebook import tqdm as notebook_tqdm\n" + ] + } + ], "source": [ "from haystack.telemetry import tutorial_running\n", "\n", @@ -150,35 +120,55 @@ }, { "cell_type": "code", - "execution_count": null, + "execution_count": 3, "metadata": { "id": "odZJjD7KgO1g" }, - "outputs": [], + "outputs": [ + { + "data": { + "text/plain": [ + "\n", + "🚅 Components\n", + " - builder: ChatPromptBuilder\n", + " - llm: HuggingFaceLocalChatGenerator\n", + "🛤️ Connections\n", + " - builder.prompt -> llm.messages (List[ChatMessage])" + ] + }, + "execution_count": 3, + "metadata": {}, + "output_type": "execute_result" + } + ], "source": [ "from haystack import Pipeline\n", - "from haystack.components.builders import PromptBuilder\n", - "from haystack.components.generators import HuggingFaceLocalGenerator\n", + "from haystack.components.builders import ChatPromptBuilder\n", + "from haystack.dataclasses import ChatMessage\n", + "from haystack.components.generators.chat import HuggingFaceLocalChatGenerator\n", "\n", - "template = \"\"\"\n", + "template = [ChatMessage.from_user(\"\"\"\n", "Please create a summary about the following topic:\n", "{{ topic }}\n", - "\"\"\"\n", - "builder = PromptBuilder(template=template)\n", - "llm = HuggingFaceLocalGenerator(\n", - " model=\"google/flan-t5-large\", task=\"text2text-generation\", generation_kwargs={\"max_new_tokens\": 150}\n", + "\"\"\")]\n", + "chat_template = \"{% for message in messages %}{% if message['role'] == 'user' %}{{ ' ' }}{% endif %}{{ message['content'] }}{% if not loop.last %}{{ ' ' }}{% endif %}{% endfor %}{{ eos_token }}\"\n", + "\n", + "builder = ChatPromptBuilder(template=template)\n", + "llm = HuggingFaceLocalChatGenerator(\n", + " model=\"google/flan-t5-large\", task=\"text2text-generation\", generation_kwargs={\"max_new_tokens\": 150},\n", + " chat_template=chat_template\n", ")\n", "\n", "pipeline = Pipeline()\n", "pipeline.add_component(name=\"builder\", instance=builder)\n", "pipeline.add_component(name=\"llm\", instance=llm)\n", "\n", - "pipeline.connect(\"builder\", \"llm\")" + "pipeline.connect(\"builder.prompt\", \"llm.messages\")" ] }, { "cell_type": "code", - "execution_count": null, + "execution_count": 4, "metadata": { "colab": { "base_uri": "https://localhost:8080/" @@ -198,7 +188,7 @@ "source": [ "topic = \"Climate change\"\n", "result = pipeline.run(data={\"builder\": {\"topic\": topic}})\n", - "print(result[\"llm\"][\"replies\"][0])" + "print(result[\"llm\"][\"replies\"][0].text)" ] }, { @@ -214,7 +204,7 @@ }, { "cell_type": "code", - "execution_count": null, + "execution_count": 16, "metadata": { "colab": { "base_uri": "https://localhost:8080/" @@ -230,21 +220,39 @@ "components:\n", " builder:\n", " init_parameters:\n", - " template: \"\\nPlease create a summary about the following topic: \\n{{ topic }}\\n\"\n", - " type: haystack.components.builders.prompt_builder.PromptBuilder\n", + " required_variables: null\n", + " template:\n", + " - content: '\n", + "\n", + " Please create a summary about the following topic:\n", + "\n", + " {{ topic }}\n", + "\n", + " '\n", + " meta: {}\n", + " name: null\n", + " role: user\n", + " variables: null\n", + " type: haystack.components.builders.chat_prompt_builder.ChatPromptBuilder\n", " llm:\n", " init_parameters:\n", " generation_kwargs:\n", " max_new_tokens: 150\n", + " stop_sequences: []\n", " huggingface_pipeline_kwargs:\n", - " device: cpu\n", + " device: mps\n", " model: google/flan-t5-large\n", " task: text2text-generation\n", - " token: null\n", - " stop_words: null\n", - " type: haystack.components.generators.hugging_face_local.HuggingFaceLocalGenerator\n", + " streaming_callback: null\n", + " token:\n", + " env_vars:\n", + " - HF_API_TOKEN\n", + " - HF_TOKEN\n", + " strict: false\n", + " type: env_var\n", + " type: haystack.components.generators.chat.hugging_face_local.HuggingFaceLocalChatGenerator\n", "connections:\n", - "- receiver: llm.prompt\n", + "- receiver: llm.messages\n", " sender: builder.prompt\n", "max_runs_per_component: 100\n", "metadata: {}\n", @@ -270,21 +278,40 @@ "components:\n", " builder:\n", " init_parameters:\n", - " template: \"\\nPlease create a summary about the following topic: \\n{{ topic }}\\n\"\n", - " type: haystack.components.builders.prompt_builder.PromptBuilder\n", + " required_variables: null\n", + " template:\n", + " - content: '\n", + "\n", + " Please create a summary about the following topic:\n", + "\n", + " {{ topic }}\n", + "\n", + " '\n", + " meta: {}\n", + " name: null\n", + " role: user\n", + " variables: null\n", + " type: haystack.components.builders.chat_prompt_builder.ChatPromptBuilder\n", " llm:\n", " init_parameters:\n", + " init_parameters:\n", " generation_kwargs:\n", " max_new_tokens: 150\n", + " stop_sequences: []\n", " huggingface_pipeline_kwargs:\n", " device: cpu\n", " model: google/flan-t5-large\n", " task: text2text-generation\n", - " token: null\n", - " stop_words: null\n", - " type: haystack.components.generators.hugging_face_local.HuggingFaceLocalGenerator\n", + " streaming_callback: null\n", + " token:\n", + " env_vars:\n", + " - HF_API_TOKEN\n", + " - HF_TOKEN\n", + " strict: false\n", + " type: env_var\n", + " type: haystack.components.generators.chat.hugging_face_local.HuggingFaceLocalChatGenerator\n", "connections:\n", - "- receiver: llm.prompt\n", + "- receiver: llm.messages\n", " sender: builder.prompt\n", "max_runs_per_component: 100\n", "metadata: {}\n", @@ -300,12 +327,12 @@ "source": [ "## Editing a Pipeline in YAML\n", "\n", - "Let's see how we can make changes to serialized pipelines. For example, below, let's modify the promptbuilder's template to translate provided `sentence` to French:" + "Let's see how we can make changes to serialized pipelines. For example, below, let's modify the `ChatPromptBuilder`'s template to translate provided `sentence` to French:" ] }, { "cell_type": "code", - "execution_count": null, + "execution_count": 5, "metadata": { "id": "U332-VjovFfn" }, @@ -315,21 +342,33 @@ "components:\n", " builder:\n", " init_parameters:\n", - " template: \"\\nPlease translate the following to French: \\n{{ sentence }}\\n\"\n", - " type: haystack.components.builders.prompt_builder.PromptBuilder\n", + " template:\n", + " - content: 'Please translate the following to French: \\n{{ sentence }}\\n'\n", + " meta: {}\n", + " name: null\n", + " role: user\n", + " variables: null\n", + " type: haystack.components.builders.chat_prompt_builder.ChatPromptBuilder\n", " llm:\n", " init_parameters:\n", " generation_kwargs:\n", " max_new_tokens: 150\n", + " stop_sequences: []\n", " huggingface_pipeline_kwargs:\n", " device: cpu\n", " model: google/flan-t5-large\n", " task: text2text-generation\n", - " token: null\n", - " stop_words: null\n", - " type: haystack.components.generators.hugging_face_local.HuggingFaceLocalGenerator\n", + " streaming_callback: null\n", + " chat_template : \"{% for message in messages %}{% if message['role'] == 'user' %}{{ ' ' }}{% endif %}{{ message['content'] }}{% if not loop.last %}{{ ' ' }}{% endif %}{% endfor %}{{ eos_token }}\"\n", + " token:\n", + " env_vars:\n", + " - HF_API_TOKEN\n", + " - HF_TOKEN\n", + " strict: false\n", + " type: env_var\n", + " type: haystack.components.generators.chat.hugging_face_local.HuggingFaceLocalChatGenerator\n", "connections:\n", - "- receiver: llm.prompt\n", + "- receiver: llm.messages\n", " sender: builder.prompt\n", "max_runs_per_component: 100\n", "metadata: {}\n", @@ -349,15 +388,15 @@ }, { "cell_type": "code", - "execution_count": null, + "execution_count": 6, "metadata": { "id": "OdlLnw-9wVN-" }, "outputs": [], "source": [ "from haystack import Pipeline\n", - "from haystack.components.builders import PromptBuilder\n", - "from haystack.components.generators import HuggingFaceLocalGenerator\n", + "from haystack.components.builders import ChatPromptBuilder\n", + "from haystack.components.generators.chat import HuggingFaceLocalChatGenerator\n", "\n", "new_pipeline = Pipeline.loads(yaml_pipeline)" ] @@ -368,12 +407,12 @@ "id": "eVPh2cV6wcu9" }, "source": [ - "Now we can run the new pipeline we defined in YAML. We had changed it so that the `PromptBuilder` expects a `sentence` and translates the sentence to French:" + "Now we can run the new pipeline we defined in YAML. We had changed it so that the `ChatPromptBuilder` expects a `sentence` and translates the sentence to French:" ] }, { "cell_type": "code", - "execution_count": null, + "execution_count": 7, "metadata": { "colab": { "base_uri": "https://localhost:8080/" @@ -385,10 +424,10 @@ { "data": { "text/plain": [ - "{'llm': {'replies': ['Je me félicite des capybaras !']}}" + "{'llm': {'replies': [ChatMessage(content='Je me félicite des capybaras', role=, name=None, meta={'finish_reason': 'stop', 'index': 0, 'model': 'google/flan-t5-large', 'usage': {'completion_tokens': 13, 'prompt_tokens': 16, 'total_tokens': 29}})]}}" ] }, - "execution_count": 15, + "execution_count": 7, "metadata": {}, "output_type": "execute_result" } @@ -421,7 +460,16 @@ "name": "python3" }, "language_info": { - "name": "python" + "codemirror_mode": { + "name": "ipython", + "version": 3 + }, + "file_extension": ".py", + "mimetype": "text/x-python", + "name": "python", + "nbconvert_exporter": "python", + "pygments_lexer": "ipython3", + "version": "3.9.6" } }, "nbformat": 4,