The problem: most "chat with your documents" tools answer confidently even when the documents don't support it. That's unusable for policies, contracts or research.
What I built: a multi-tenant RAG platform built on one rule, no source, no answer. Users upload PDFs, Word or text files to a workspace and ask questions. Answers stream in with numbered citations to the exact document and page, and the server verifies every citation against the retrieved text before showing it.
Engineering highlights:
Citations verified server-side; unsupported answers return "I couldn't find that"
Semantic search with PostgreSQL + pgvector; streaming over Server-Sent Events
Tenant isolation in every query; Supabase's public data API locked down (verified on production)
A crash-safe job queue in PostgreSQL: a kill -9 under load lost nothing
AI costs reserved before every call, with per-user, per-workspace and deployment-wide caps
A data-driven model migration: 5 configurations evaluated, about 74% cheaper per answer
Results: 1,099 backend and 393 frontend tests passing; load-tested (0 errors at 72 concurrent searches); encrypted, restore-verified backups; deployed live at ₹0/month hosting.
Live demo: https://knowledgepilot-ai-olive.vercel.app
Code: https://github.com/vivek0108gupta/knowledgepilot-ai
0 likes · 0 views · Nov 18, 2026
Solo full-stack build of a multi-tenant RAG app: answers cite the exact document and page, and every citation is verified server-side. Load-tested and deployed