generic-llm-memorizer 0.1.3

A library for managing LLM conversations and context

Generic LLM Memorizer

A Python library for managing LLM conversations and context with automatic fact extraction and structured data handling.

Installation

pip install generic-llm-memorizer

Usage

Basic Conversation Management

from generic_llm_memorizer import Conversation

# Create a new conversation
conv = Conversation(system_prompt="You are a helpful assistant.")

# Add messages
conv.add_user_message("Hello, how are you?")
conv.add_assistant_message("I'm doing well, thank you! How can I help you today?")

# Get conversation history
messages = conv.get_messages()

# Format for LLM API (expects list of dicts with role and content)
from generic_llm_memorizer import format_conversation_for_prompt
formatted = format_conversation_for_prompt(conv)
# Output: [{"role": "system", "content": "You are a helpful assistant."}, ...]

from generic_llm_memorizer import Conversation

# Create a new conversation
conv = Conversation(system_prompt="You are a helpful assistant.")

# Add messages
conv.add_user_message("Hello, how are you?")
conv.add_assistant_message("I'm doing well, thank you! How can I help you today?")

# Get conversation history
messages = conv.get_messages()

# Format for LLM API (expects list of dicts with role and content)
from generic_llm_memorizer import format_conversation_for_prompt
formatted = format_conversation_for_prompt(conv)
# Output: [{"role": "system", "content": "You are a helpful assistant."}, ...]

Persisting Conversations

from generic_llm_memorizer import Conversation

# Save conversation to JSON
conv_json = conv.to_json()
with open("conversation.json", "w") as f:
    f.write(conv_json)

# Load conversation from JSON
with open("conversation.json", "r") as f:
    conv_json = f.read()
loaded_conv = Conversation.from_json(conv_json)

from generic_llm_memorizer import Conversation

# Save conversation to JSON
conv_json = conv.to_json()
with open("conversation.json", "w") as f:
    f.write(conv_json)

# Load conversation from JSON
with open("conversation.json", "r") as f:
    conv_json = f.read()
loaded_conv = Conversation.from_json(conv_json)

Utility Functions

Extending with Custom Models

You can create your own Pydantic models for domain-specific extractions:

from pydantic import BaseModel
from typing import List
from enum import Enum

class Priority(str, Enum):
    LOW = "low"
    MEDIUM = "medium"
    HIGH = "high"

class Task(BaseModel):
    title: str
    description: str
    priority: Priority
    assignee: str

class TaskList(BaseModel):
    tasks: List[Task]

# Use with the generic llm_call function
tasks = llm_call("Extract tasks from this conversation", TaskList)

Advanced Usage with Structured Data Extraction

The library now supports a generic llm_call function that works with Pydantic models for structured data extraction:

from typing import Type, TypeVar
from pydantic import BaseModel
from generic_llm_memorizer.conversation import Conversation
from generic_llm_memorizer.memory import Summary, Memory, Fact

# Define a generic type variable bounded by BaseModel
T = TypeVar("T", bound=BaseModel)

def llm_call(prompt: str, model_class: Type[T]) -> T:
    """
    Generic LLM call function that returns structured data.
    
    Args:
        prompt: The prompt to send to the LLM
        model_class: The Pydantic model class to parse the response into
        
    Returns:
        An instance of the specified model_class
    """
    # In a real implementation, this would call an actual LLM API
    # and parse the response into the specified Pydantic model
    
    # Mock implementation for demonstration
    if model_class == Summary:
        return Summary(content="Mock summary of the conversation")
    elif model_class == Memory:
        return Memory(facts=[
            Fact(
                content="Example fact extracted from conversation",
                kind="example",
                timestamp="2023-01-01T10:00:00Z",
                validity="forever"
            )
        ])
    
    # Return a default instance of the model
    return model_class()

# Usage with LLMMemorizer
from generic_llm_memorizer.llm_memorizer import LLMMemorizer

conversation = Conversation(system_prompt="You are a helpful assistant")
conversation.add_user_message("Hello, my name is John Doe and I'm a software engineer")

memorizer = LLMMemorizer(
    user_id="user_123",
    session_id="session_456",
    llm_call=llm_call,
    conversation=conversation
)

# Extract structured facts
memory = memorizer.extract_facts()
print(f"Extracted {len(memory.facts)} facts")

# Generate structured summary
summary = memorizer.summarize()
print(f"Summary: {summary.content}")

## API Reference

```python
from generic_llm_memorizer import extract_last_user_message, extract_last_assistant_message

# Extract last user message
last_user = extract_last_user_message(conv)

# Extract last assistant message
last_assistant = extract_last_assistant_message(conv)

API Reference

Conversation Class

Manages a conversation with an LLM.

Methods

• add_message(role, content, metadata=None): Add a message to the conversation
• add_user_message(content, metadata=None): Add a user message
• add_assistant_message(content, metadata=None): Add an assistant message
• get_messages(): Get all messages
• get_recent_messages(count): Get the most recent messages
• clear(): Clear all messages
• to_dict(): Convert to dictionary format
• to_json(indent=2): Convert to JSON string
• from_dict(data): Create from dictionary data
• from_json(json_str): Create from JSON string

Message Dataclass

Represents a single message in a conversation.

Attributes

• role: Role of the message sender ("user", "assistant", "system", etc.)
• content: Message content
• timestamp: ISO format timestamp (automatically set)
• metadata: Additional metadata dictionary

Utility Functions

• format_conversation_for_prompt(conversation): Format conversation for LLM API
• extract_last_user_message(conversation): Get last user message content
• extract_last_assistant_message(conversation): Get last assistant message content

License

MIT