Enterprise RAG Platform

A design for question answering over an organisation's own documents: ingestion, hybrid retrieval, re-ranking and generation constrained to cite its sources, with an offline evaluation loop. It extends the FAISS retrieval used in the AI project management system to a document-grounded setting.

Organisations want answers from their own documents, but a language model alone invents details and cannot cite sources. The system has to retrieve the right passages, keep answers inside that evidence, respect document permissions, and show the user where each claim came from.

Ingestion workers parse documents, split them into structure-aware chunks and write embeddings to a vector index next to a keyword index. A query is expanded, answered by hybrid retrieval (vector and keyword), re-ranked by a cross-encoder, filtered by the caller's permissions, and passed to the generator with a strict citation contract. Every step emits traces for an evaluation store.

Embeddings for semantic recall, lexical scoring for exact terms and identifiers, a re-ranker for precision, and a generator constrained to cite retrieved chunks. An offline evaluation set measures retrieval recall and answer faithfulness before any change reaches users.

Permissions are enforced at retrieval time, not in the prompt; answers without supporting passages are refused rather than guessed; and the index is rebuilt from source documents, never edited in place.

Design objective: grounded answers with citations for every claim, measurable retrieval recall and faithfulness on a fixed evaluation set, and no document visible to a user who could not open it directly.