quivr

Why we included this project

Most teams don't need to assemble a RAG pipeline from scratch; they need one that works and can be adapted. Quivr delivers that as a Python package (quivr-core) where a few lines of code turn a folder of PDFs, text, or Markdown into a queryable 'brain', so a document-QA feature can be prototyped in an afternoon. The opinionated defaults handle the fiddly parts like chunking, embedding, retrieval, and reranking, while still letting you plug in any LLM (OpenAI, Anthropic, Mistral, or local models via Ollama) and any vector store such as PGVector or Faiss. For more control, you can define retrieval workflows in a YAML file to tune history filtering, query rewriting, and reranker settings, and extend ingestion with custom parsers or the companion Megaparse tool. It's a solid baseline for developers who want a working RAG they can customize rather than wiring components together themselves.