03Requirements
Answer questions over supplied documents with traceable citations. Unsupported questions should produce a clear decline instead of an invented answer.
OPENING THE SYSTEM
Loading the next page…
Loading the next page…
Document question answering with hybrid retrieval, reranking, source citations and an explicit path to decline unsupported answers.
An end-to-end RAG pipeline ingests documents, persists local embeddings in ChromaDB, combines BM25 and semantic retrieval, reranks the candidates and generates cited answers with Groq. Weak evidence triggers abstention.
Document question answering needs to find relevant evidence and recognise when the available context cannot support an answer.
Load documents, split them into chunks and preserve source metadata for later citations.
Answer questions over supplied documents with traceable citations. Unsupported questions should produce a clear decline instead of an invented answer.
Load documents, split them into chunks and preserve source metadata for later citations.
Document ingestion, chunking, local embeddings and persistent ChromaDB storage form the knowledge layer. Source metadata travels with retrieved chunks.
BM25 complements semantic retrieval for exact terminology. Reciprocal-rank fusion merges the ranked lists; a cross-encoder then reranks the candidate evidence.
Groq generates answers from retrieved context. Citation validation and confidence-based abstention make evidence and uncertainty part of the output contract.
The repository extends the pipeline with a FastAPI query interface, health checks and Docker packaging. It also documents authentication, document access controls and per-user rate limits.
The project includes regression and test workflows for retrieval and citation behaviour. Repository-reported regression results are available with the source; they are not presented here as a general production-accuracy guarantee.
Keep ingestion, retrieval, reranking and generation separate so each stage can be inspected and evaluated. Treat refusal to answer as a valid outcome when the evidence is weak.
Expand representative evaluation cases, measure retrieval and answer quality together, and document latency, cost and failure behaviour under realistic workloads.