Summary

Lyra-stellAI/BYO-LLM-WIKI offers a local-first framework for building a personal, LLM-native knowledge library. It transforms unstructured data into a coherent, searchable knowledge base featuring contextual RAG, a layered knowledge graph, and reinforced memory. The system enables automatic generation of agent skills, gated by human review, providing a powerful tool for knowledge management and agentic task automation.

How It Works

This project utilizes a local-first architecture, storing state in plain JSON/SQLite files accessible via Flask UI, JSON API, and CLI. Knowledge ingestion populates a multi-layered knowledge graph and HNSW vector index. Contextual RAG retrieves information using a two-layer index, enhanced by LLM re-ranker/MMR for grounded, cited answers. A deepagents agent manages the KG and memory, while an agent-skill loop uses LangGraph for durable, human-gated skill generation and refinement. Integration with external MCP servers is supported.

Quick Start & Requirements

• Install: pip install -r requirements.txt (add ragas for RAGAS metrics).
• Prerequisites: Python 3, LLM provider API keys (e.g., OPENAI_API_KEY for embeddings). Configure via .env.example to .env.
• Run: python app.py (Flask UI) or python runner.py (CLI).
• Links: Live Demo: https://byo-wiki-demo.onrender.com

Highlighted Details

• Layered Knowledge Graph: Organizes information hierarchically (source -> section -> chunk -> entity -> topic -> synthesis -> memory -> skill).
• Gated Agent Skills: Automated pipeline generates skills, evaluated and requiring human review before integration.
• Contextual RAG with MMR: Two-layer HNSW index with MMR improves retrieval recall and synthesis over graph traversal.
• MCP Integration: Connects to external MCP servers (Supabase, GitHub) and can act as one, with gated writes.
• Robust Evaluation: Includes retrieval metrics, RAGAS, and a multi-provider LLM-judge panel.

Maintenance & Community

The project is inspired by LangChain and uses common libraries. Specific details on active maintenance, core contributors, or community channels are not provided in the README.

Licensing & Compatibility

• License: Not specified in the README, posing a significant adoption risk.
• Compatibility: Local-first design ensures broad compatibility with standard Python environments. MCP integration allows flexible external service connections.

Limitations & Caveats

Primarily designed for local-first use; cloud features are opt-in. The agent skill pipeline requires human review, potentially creating a bottleneck. Reliance on external LLM APIs incurs costs and vendor dependencies. The absence of a declared license is a critical adoption blocker.