// the take
RAGLight is a Python RAG framework that wraps together embeddings, vector stores, and LLM providers (Ollama, OpenAI, Mistral, Gemini, Bedrock, LMStudio, vLLM) behind a Builder/Config API, plus a CLI wizard and a FastAPI server mode. It's aimed at people who want a RAG pipeline without wiring LangChain/LlamaIndex primitives by hand, and increasingly at agentic RAG with MCP tool access.
Hybrid search is a real feature, not a checkbox — BM25 + dense vectors merged with RRF, available on both Chroma and Qdrant backends. The `raglight serve` command gives you a working FastAPI + optional Streamlit UI configured entirely from env vars, which is more useful out of the box than most RAG libraries that only ship a Python API. Query reformulation for multi-turn follow-ups and conversation history with a `max_history` cap are handled for you across every provider, which is normally a hand-rolled piece of glue code.
The provider matrix (7 LLM backends × 5 embedding backends × 2 vector stores) is a lot of surface area for one maintainer to keep correct — expect uneven testing depth on the less common combinations like Bedrock or Gemini embeddings. Vector store choice is limited to Chroma and Qdrant; no pgvector, Pinecone, or Weaviate despite the 'plug in anything' pitch. The README's breadth (VLM PDF processing, Langfuse tracing, MCP, agentic pipelines, hybrid search, streaming) reads like feature accretion rather than a hardened core, and there's no mention of async I/O anywhere, which matters if you're actually running `serve` under real concurrency.
We read the README, the directory tree and a key source file in September 2026. Nobody pays to be listed. How we judge →