IntelliDocs AI — RAG Document Q&A Platform

Upload documents and ask questions — answers come back with their sources

The challenge

Teams needed to upload PDFs, Word docs, images, and text, then ask questions with source citations — while keeping each customer's data isolated and blocking prompt-injection attacks.

Approach

LangGraph ingestion pipeline (PyMuPDF, Unstructured, python-docx, Tesseract OCR) splits files into searchable chunks. Each customer's documents live in a separate Chroma collection. Presidio masks PII before ingestion; LiteLLM routes between Gemini, Cohere, and Groq with fallback. LangGraph + Celery + RabbitMQ run background ingestion jobs. Keycloak handles auth.

Results

  • Blocks prompt-injection attacks in 5 languages (English, German, Spanish, French, Nepali).
  • Every answer is checked for leaked secrets before returning.
  • Quality gates run Ragas, DeepEval, and attack tests in CI on every change.
  • Answers stream back with source citations.

Metrics

AI providers

Gemini, Cohere, Groq (LiteLLM router)

Customer isolation

One Chroma collection per customer

Injection languages

5

Ingestion stack

LangGraph + Celery + RabbitMQ

Quality checks

Ragas, DeepEval, attack tests

Technologies

  • Python
  • FastAPI
  • LangGraph
  • LiteLLM
  • ChromaDB
  • LangChain
  • PostgreSQL
  • Celery
  • RabbitMQ
  • Keycloak
  • Presidio
  • Docker
Live site

Want similar work? Get in touch.