// the find
QuivrHQ/quivr
Opiniated RAG for integrating GenAI in your apps 🧠 Focus on your product rather than the RAG. Easy integration in existing products with customisation! Any LLM: GPT4, Groq, Llama. Any Vectorstore: PGVector, Faiss. Any Files. Anyway you want.
Quivr is a Python library that wraps LangChain-based RAG into a higher-level `Brain` abstraction — ingest files, ask questions, done. It targets developers who want document Q&A in their app without assembling the retrieval pipeline themselves. The hosted quivr.com product has been stripped out; what remains is the `quivr-core` pip package.
YAML-driven workflow configuration is genuinely useful — you can swap retrieval strategies, reranker models, and history window sizes without touching code. Processor registry pattern makes adding custom file parsers clean and composable. LangGraph integration for the RAG graph means you get proper conditional branching and streaming out of the box rather than a hand-rolled chain. Test coverage is solid for the processor layer with real fixture files across PDF, EPUB, DOCX, ODT.
Heavy LangChain dependency means you inherit all of its abstraction debt — debugging a retrieval failure requires tracing through multiple layers you didn't write. The pivot from full-stack app to library left visible scars: the repo still carries frontend scaffolding, Docker files, and docs that reference features that no longer exist in this codebase. Cohere is hardcoded as the example reranker in the default YAML config, which will silently fail for anyone without a Cohere key who didn't read carefully. No built-in chunking strategy control beyond what LangChain exposes — if the default splitter doesn't work for your content type, you're digging through the abstraction.