A Modular RAG Toolkit That Won't Lock You Into One LLM

What it does

RAGLight is a Python library that assembles Retrieval-Augmented Generation pipelines from interchangeable parts. It handles document ingestion, embedding, retrieval, and text generation behind a unified interface, and ships with both a CLI wizard and a FastAPI server for teams that want a working interface without writing boilerplate. You can feed it PDFs, code files, or GitHub repos and start querying through Python, the terminal, or a browser.

The interesting bit

The framework treats LLMs and vector stores as replaceable cartridges rather than fixed dependencies: Ollama, OpenAI, Mistral, Gemini, LMStudio, vLLM, and AWS Bedrock all implement the same internal contract. Recent additions include an agentic pipeline, MCP tool integration for external capabilities like code execution, and hybrid search that fuses BM25 keyword scoring with dense vector retrieval.

Key highlights

• Swap LLMs, embedding models, or vector stores (ChromaDB, Qdrant) without rewriting pipeline logic
• Hybrid search with BM25 + semantic + reciprocal rank fusion for retrieval that catches keyword and meaning matches
• Query reformulation and full conversation history across all supported providers
• CLI wizard and FastAPI/Streamlit server for zero-code chat and ingestion
• Optional Langfuse tracing and MCP server integration for tool-augmented answers

Verdict

RAGLight is a solid choice for developers who want a vendor-neutral RAG sandbox or a quick on-premise document chat. If you are building a heavily customized retrieval system from scratch, this is well-organized glue code around existing APIs rather than a research breakthrough.