I built a "chat with your documents" app that refuses to make things up.
Upload PDFs, DOCX or TXT → ask questions → every answer cites the exact document and page. No source, no answer.
Next.js + FastAPI + Postgres/pgvector + OpenAI. Sharing the whole build, one day at a time 👇 #buildinpublic
🚀 I’m Open to New Opportunities | Full-Time & Freelance
Due to funding challenges at my previous startup, I’m currently exploring new opportunities.
I have 9+ years of software engineering experience, specializing in:
🔹 React.js, Next.js, TypeScript, Node.js, Python
🔹 AI-powered video generation & automation
🔹 HeyGen, Seedance, Veo, Kling, OmniHuman
🔹 AI Agents & workflow automation
🔹 LLM integrations and backend systems
I’ve built AI Video Agents, automated ad-generation pipelines, and scalable full-stack applications.
📍 Location: Bangalore, India
🌍 Open to: Remote / Hybrid / On-site
🤝 Available for: Full-time roles & Freelance projects
If your team is hiring, you need help building AI products, or you know someone looking for these skills, I’d love to connect.
Reposts and referrals would mean a lot!
Let’s build something amazing together. 🚀
#OpenToWork #FreelanceDeveloper #AIEngineer #FullStackDeveloper #ReactJS #GenerativeAI #AIVideo
@srbh_here Agreed for the quotes themselves: in KnowledgePilot the server checks every quoted span against the retrieved chunk, and drops the citation if it doesn’t match.
The open problem is the step after: a real quote that doesn’t actually support the claim.
1/ I built KnowledgePilot AI: ask questions about your documents, get answers you can check. Every answer cites the document + page. 🎥
2/ Citations aren't just labels. The server checks each quote against the passages the model was given, and drops any it can't verify. No source → "I couldn't find that."
3/ The scariest bug was one I almost shipped: the newer model rejects temperature with a 400. A "one-line model change" would have broken every answer. Now each model gets only the parameters it accepts.
4/ Security catch: Supabase exposes public tables through the anon key, which is public. A migration revokes access and enables RLS. Checked on prod: 42501 permission denied.
5/ Load test: 72 concurrent searches with 0 errors; kill -9 mid-ingestion → 24/24 docs indexed, none twice. 1,099 + 393 tests passing.
6/ Live on Vercel + Render + Supabase free tiers, with caps of $0.50/day and $3/month. Try it 👇
@zeal_m_ Teams with documents where a wrong answer is costly: policies, handbooks, contracts, research. The first users are people who’d rather hear “not in your docs” than a confident guess. The demo is free to try: 2 searches and 2 questions.
I built a RAG app that refuses to guess 🎥
Every answer cites its document + page, and citations are verified server-side.
Next.js · FastAPI · Postgres/pgvector · OpenAI Deployed for ₹0/month hosting, with hard caps on AI spend.
Live demo below 👇 #buildinpublic
@BlakeXu_ Agreed, and it’s the cheapest part to add. Showing a citation is easy; checking that the quote is really in the retrieved text is a few lines of server code. It’s what makes the citations worth clicking.
@VladSoftware The server checks each cited quote against the passages the model was actually given. A citation that doesn’t match gets dropped, not shown. If nothing in the answer can be verified, it returns “I couldn’t find that” instead of an uncited answer.
Agreed, and it’s half there. If no citation maps to a passage the model was actually given, the answer is replaced with “I couldn’t find that.” Pages come from the stored chunk, never the model.
The gap: a real source with a quote that doesn’t match is still shown. Tightening that next.
@itsvlady Page isn’t string-matched; the model can’t set it. It cites source IDs it was given (S1, S2); the server builds the citation from the stored chunk’s page.
Quotes are string-matched after normalising (case, punctuation, PDF hyphens), with a 6+ word run as fallback.
@TenetConcept Right now each claim is cited to its own doc + page, so disagreements are visible, not hidden. Explicitly flagging conflicts (and preferring the newer doc) is on my roadmap.