Problem: LLM agents struggle to consistently learn from user feedback without requiring costly model retraining or complex infrastructure.
Solution: This library provides a simple system for capturing, storing, and reusing feedback for LLM tasks. It works by:
Collecting feedback on LLM outputs
Storing this feedback with embeddings of the original task
Retrieving relevant feedback for similar future tasks (feedback selection layer: only openai right now)
Enhancing prompts with the feedback to improve results
All of this happens without any model retraining - just by enhancing prompts with contextual feedback.
✨ Features
Simple API: Just a few methods to enhance prompts and save feedback
Multiple Embedding Models: Support for OpenAI and HuggingFace models (MiniLM, BGE-small)
Local-First: Uses JSON files for storage with no external DB requirements
Smart Feedback Selection: Uses OpenAI to choose the most relevant feedback for a task
Async Support: Both synchronous and asynchronous APIs for better performance
Customizable: Configurable thresholds, formatters, and memory handling
Zero Infrastructure: Works out of the box with minimal setup
Framework Agnostic: Works with any LLM provider (OpenAI, Anthropic, etc.)
Integration Examples: Ready-to-use examples with LangChain, Agno, and more