RAGatouille

RAGatouille packages ColBERT, a late-interaction retrieval model from research, into a Python library that slots directly into RAG pipelines. It covers the full workflow: training and fine-tuning ColBERT models with built-in data processing and hard negative mining, indexing document collections into compressed indices persisted to disk, and querying with single or batch searches. Where single-vector dense embeddings collapse a document into one vector, ColBERT matches queries against documents token by token, an approach that tends to generalize better to new domains, stay data efficient, and handle multilingual content well. Indices and models interoperate with the official ColBERT implementation, Vespa, Intel's fastRAG, and LlamaIndex. The library targets Python 3.9 to 3.11 and has limited Windows support, with WSL2 recommended. Maintained under the AnswerDotAI organization as open source, it is used by developers who want contemporary retrieval quality without wrestling with research code.