Air-Gapped Document Intelligence: AI Answers That Never Leave Your Network

The client needed AI answers over a sensitive document corpus, but that data legally could not be sent to OpenAI or Anthropic. The usual 'chat with your docs' build was off the table — not on cost, but on jurisdiction. A second problem was subtler: a retrieval system that invents a citation is worse than one that stays silent, and they had no way to measure which they would get.

We built a self-hosted, provider-agnostic RAG platform — an engine for document-grounded assistants rather than a single chatbot. Retrieval is hybrid: vector search over pgvector/HNSW fused with Postgres full-text via reciprocal rank fusion, so exact-keyword and semantic matches both surface. Answers are citation-grounded with fabricated-citation rejection, so the system refuses when the answer is not in the corpus instead of improvising. A RAGAS-style evaluation harness measures faithfulness, answer relevance, context precision and context recall via CLI, UI and CSV export. Every model provider sits behind a lint-enforced adapter boundary, so switching between a hosted model and a local one is a config change.