Added option to output the raw json additionally to the output.

Author: Maximilian-Winter

examples/07_Memory/VirtualGameMaster/main.py

@@ -7,7 +7,8 @@
 from llama_cpp_agent.llm_agent import SystemPromptModule, SystemPromptModulePosition
 from llama_cpp_agent.providers import LlamaCppServerProvider
 from memory import output_settings, agent_core_memory, agent_retrieval_memory, agent_event_memory, update_memory_section
-from prompts import game_master_prompt, examples, memory_prompt, wrap_function_response_in_xml_tags_json_mode, generate_fake_write_message, generate_write_message_with_examples, wrap_player_message_in_xml_tags_json_mode
+from prompts import game_master_prompt, examples, memory_prompt, wrap_function_response_in_xml_tags_json_mode, \
+    generate_fake_write_message, generate_write_message_with_examples, wrap_player_message_in_xml_tags_json_mode
 
 provider = LlamaCppServerProvider("http://localhost:8080")
 
@@ -29,46 +30,53 @@
 date_time_section = SystemPromptModule("current_date_time", "The following section shows the current date and time:")
 
 example_section = SystemPromptModule("examples",
-                                     "The following examples show you which kind of responses you should write to the user based on the current scenario:", suffix="Always remember to never write actions or dialogue for the user! Always let the user decide on actions or dialogue!")
+                                     "The following examples show you which kind of responses you should write to the user based on the current scenario:",
+                                     suffix="Always remember to never write actions or dialogue for the user! Always let the user decide on actions or dialogue!")
 example_section.set_content(examples)
 memory_intro_section = SystemPromptModule("memory_intro",
                                           "To support you in your task as a game master and to help you remembering things, you have access to 3 different types of memory.",
                                           position=SystemPromptModulePosition.after_system_instructions)
 memory_intro_section.set_content(memory_prompt)
+output_settings.output_structured_output_and_raw_json_string = True
 while True:
     user_input = input(">")
     if user_input == "exit":
         break
     update_memory_section(memory_section)
     date_time_section.set_content(datetime.datetime.now().strftime("%d.%m.%Y") + "\nFormat: dd.mm.yyyy")
 
-    agent_event_memory.add_event_to_queue(Roles.user, wrap_player_message_in_xml_tags_json_mode(user_input))
-    agent_output = agent.get_chat_response(
+    agent_event_memory.add_event(Roles.user, wrap_player_message_in_xml_tags_json_mode(user_input))
+    agent_output, json_output = agent.get_chat_response(
         chat_history=agent_event_memory.get_event_memory_manager().build_chat_history(),
         llm_sampling_settings=settings,
         system_prompt_modules=[memory_intro_section, memory_section, date_time_section],
         structured_output_settings=output_settings)
 
+    agent_event_memory.add_event(Roles.assistant, json_output)
     while True:
         update_memory_section(memory_section)
         date_time_section.set_content(datetime.datetime.now().strftime("%d.%m.%Y") + "\nFormat: dd.mm.yyyy")
 
         if agent_output[0]["function"] == "write_message_to_player":
-            agent_event_memory.add_event_to_queue(Roles.tool, generate_write_message_with_examples(examples=example_section.get_formatted_content()))
             output = agent.get_chat_response(
+                generate_write_message_with_examples(examples=example_section.get_formatted_content()),
+                role=Roles.tool,
                 chat_history=agent_event_memory.get_event_memory_manager().build_chat_history(),
                 add_message_to_chat_history=False, add_response_to_chat_history=False,
                 system_prompt_modules=[memory_intro_section, memory_section, date_time_section],
                 llm_sampling_settings=settings)
-            agent_event_memory.add_event_to_queue(Roles.tool, generate_fake_write_message())
-            agent_event_memory.add_event_to_queue(Roles.assistant, output)
+            agent_event_memory.add_event(Roles.tool, generate_fake_write_message())
+            agent_event_memory.add_event(Roles.assistant, output)
 
             print(output)
             break
 
-        agent_event_memory.add_event_to_queue(Roles.tool, wrap_function_response_in_xml_tags_json_mode(agent_output[0]["return_value"]))
-        agent_output = agent.get_chat_response(chat_history=agent_event_memory.get_event_memory_manager().build_chat_history(),
-                                               llm_sampling_settings=settings,
-                                               system_prompt_modules=[memory_intro_section, memory_section,
-                                                                      date_time_section],
-                                               structured_output_settings=output_settings)
+        agent_event_memory.add_event(Roles.tool, wrap_function_response_in_xml_tags_json_mode(
+            agent_output[0]["return_value"]))
+        agent_output, json_output = agent.get_chat_response(
+            chat_history=agent_event_memory.get_event_memory_manager().build_chat_history(),
+            llm_sampling_settings=settings,
+            system_prompt_modules=[memory_intro_section, memory_section,
+                                   date_time_section],
+            structured_output_settings=output_settings)
+        agent_event_memory.add_event(Roles.assistant, json_output)

src/llama_cpp_agent/agent_memory/event_memory_manager.py

@@ -1,5 +1,5 @@
 from sqlalchemy.orm import Session
-from .event_memory import Event, EventType
+from .event_memory import Event
 import datetime
 import json
 
@@ -24,7 +24,7 @@ def build_chat_history(self):
         messages = self.build_event_memory_context()
         for message in messages:
             history.add_message(message)
-        return messages
+        return history
 
     def add_event_to_queue(self, event_type: Roles, content: str, metadata: dict):
         new_event = Event(

src/llama_cpp_agent/agent_memory/memory_tools.py

@@ -6,8 +6,9 @@
 from sqlalchemy import create_engine
 from sqlalchemy.orm import sessionmaker, scoped_session
 
-from .event_memory import EventType, Base
+from .event_memory import Base
 from .event_memory_manager import EventMemoryManager
+from ..chat_history.messages import Roles
 from ..function_calling import LlamaCppFunctionTool
 from .core_memory_manager import CoreMemoryManager
 from .retrieval_memory_manager import RetrievalMemoryManager, RetrievalMemory
@@ -39,10 +40,10 @@ def run(self, event_memory_manager: EventMemoryManager):
             self.page = 0
         return event_memory_manager.query_events(
             event_types=[
-                EventType.UserMessage,
-                EventType.AgentMessage,
-                EventType.SystemMessage,
-                EventType.FunctionMessage,
+                Roles.user,
+                Roles.assistant,
+                Roles.system,
+                Roles.tool,
             ],
             content_keywords=self.keywords,
             start_date=parsed_start_datetime,
@@ -83,10 +84,10 @@ def run(self, event_memory_manager: EventMemoryManager):
 
         return event_memory_manager.query_events(
             event_types=[
-                EventType.UserMessage,
-                EventType.AgentMessage,
-                EventType.SystemMessage,
-                EventType.FunctionMessage,
+                Roles.user,
+                Roles.assistant,
+                Roles.system,
+                Roles.tool,
             ],
             start_date=parsed_start_datetime,
             end_date=parsed_end_datetime,
@@ -279,6 +280,8 @@ def __init__(self, event_queue_file=None, db_path="sqlite:///events.db"):
             conversation_search_date, event_memory_manager=self.event_memory_manager
         )
 
+    def add_event(self, role: Roles, content:str):
+        self.event_memory_manager.add_event_to_queue(role, content, {})
     def get_event_memory_manager(self):
         return self.event_memory_manager
 

src/llama_cpp_agent/llm_agent.py

@@ -427,10 +427,17 @@ def get_response_role_and_completion(
                     "content": message,
                 },
             )
+
         if system_prompt:
-            messages[0]["content"] = system_prompt
+            if messages[0]["role"] != Roles.system:
+                messages.insert(0, {"role": Roles.system, "content": system_prompt})
+            else:
+                messages[0]["content"] = system_prompt
         else:
-            messages[0]["content"] = self.system_prompt
+            if messages[0]["role"] != Roles.system:
+                messages.insert(0, {"role": Roles.system, "content": self.system_prompt})
+            else:
+                messages[0]["content"] = self.system_prompt
 
         additional_suffix = ""
         if self.add_tools_and_structures_documentation_to_system_prompt:

src/llama_cpp_agent/llm_output_settings/settings.py

@@ -111,7 +111,10 @@ class LlmStructuredOutputSettings(BaseModel):
         False,
         description="If the output should be just the generated JSON string by the LLM",
     )
-
+    output_structured_output_and_raw_json_string: Optional[bool] = Field(
+        False,
+        description="If the output should be a tuple of the output and the generated JSON string by the LLM",
+    )
     class Config:
         arbitrary_types_allowed = True
 
@@ -144,7 +147,8 @@ def from_llama_cpp_function_tools(
 
     @staticmethod
     def from_pydantic_models(
-            models: List[type[BaseModel]], output_type: LlmStructuredOutputType, add_thoughts_and_reasoning_field: bool = False
+            models: List[type[BaseModel]], output_type: LlmStructuredOutputType,
+            add_thoughts_and_reasoning_field: bool = False
     ):
         """
         Create settings from a list of Pydantic models with a specific output type.
@@ -612,22 +616,21 @@ def add_all_current_functions_to_heartbeat_list(self, excluded: list[str] = None
             [tool.model.__name__ for tool in self.function_tools if tool.model.__name__ not in excluded]
         )
 
-
     def handle_structured_output(self, llm_output: str, prompt_suffix: str = None):
         if self.output_raw_json_string:
             return llm_output
 
         if prompt_suffix:
             llm_output = llm_output.replace(prompt_suffix, "", 1)
 
-
         if (
                 self.output_type is LlmStructuredOutputType.function_calling
                 or self.output_type is LlmStructuredOutputType.parallel_function_calling
         ):
             output = parse_json_response(llm_output)
             output = self.clean_keys(output)
-
+            if self.output_structured_output_and_raw_json_string:
+                return self.handle_function_call(output), llm_output
             return self.handle_function_call(output)
         elif self.output_type == LlmStructuredOutputType.object_instance:
             output = parse_json_response(llm_output)
@@ -636,6 +639,8 @@ def handle_structured_output(self, llm_output: str, prompt_suffix: str = None):
             model_attributes = output[self.output_model_attributes_field_name]
             for model in self.pydantic_models:
                 if model_name == model.__name__:
+                    if self.output_structured_output_and_raw_json_string:
+                        return model(**model_attributes), llm_output
                     return model(**model_attributes)
 
         elif self.output_type == LlmStructuredOutputType.list_of_objects:
@@ -648,6 +653,8 @@ def handle_structured_output(self, llm_output: str, prompt_suffix: str = None):
                     model_attributes = out[self.output_model_attributes_field_name]
                     if model_name == model.__name__:
                         models.append(model(**model_attributes))
+            if self.output_structured_output_and_raw_json_string:
+                return models, llm_output
             return models
         return llm_output