A document assistant that cites its sources

Answers only from your files, cites the file and section, says when the answer isn't there instead of inventing one, and flags instructions hidden in documents.

Python
FastAPI
PostgreSQL
pgvector
llama.cpp
Docker

Measured result

20/20

evaluation cases

The problem

General-purpose assistants invent convincing answers, and staff can't tell where they came from. A single document with hidden instructions can change the assistant's behaviour.

What I built

Documents are chunked and embedded into PostgreSQL with pgvector, and the closest passages are retrieved for each question.
Structured answers: status, answer, and citations with a verbatim quote from the passage.
Guards against instructions injected inside documents, and an explicit refusal when the information isn't there.
Works with OpenAI, Claude, or a local model on the company's server.

Measured result

A 20-case evaluation including unanswerable questions and injection attempts: 20/20 on the second run (19/20 on the first), using a local 35B model on my own server.
38 unit tests pass without needing a model or a database.

On my modest hardware an answer took about a minute; a stronger server or a cloud model is much faster.

The assistant's evaluation results and an Arabic answer with its source

A real answer: a document with a hidden instructionjson

{
  "status": "answered",
  "answer": "Approved vendor invoices are paid within 45 days of the invoice receipt date [C1].",
  "citations": [{
    "id": "C1",
    "doc": "vendor_payments_faq.txt",
    "section": "Payment terms / شروط الدفع",
    "quote": "Approved vendor invoices are paid within 45 days of the invoice receipt date."
  }],
  "guard_flags": ["injection_in_document:vendor_payments_faq-004-60d9a0f5aa"]
}

Need something similar?

All case studies

13/13tests passing

WhatsApp Cloud API audit and a secure webhook

A toolkit that shows whether a WhatsApp integration is really official and who owns it, plus a webhook that verifies Meta's signature and never processes a message twice.

30/30audio tracks, 0% packet loss

A self-hosted LiveKit voice server for a Flutter app

Audio rooms on your own server instead of per-minute billing, with a one-command install and audio-only tokens.

34tests passing

An offline-first Odoo server that syncs to the cloud

The shop always works on a local Odoo, and a background daemon pushes invoices and customers to the cloud when the internet returns, without duplicates.