Sanad — Arabic Source-Grounded RAG Assistant

Overview

Sanad is an Arabic question-answering assistant grounded in a curated library of ~80K passages across 10 knowledge categories. Instead of answering from model memory, it retrieves relevant passages, generates an answer in Arabic, and cites the exact book, author, and page — and it declines to answer when nothing relevant is retrieved.

Challenge

General-purpose LLMs hallucinate and rarely cite sources, which is unacceptable for Arabic religious and scholarly content. The challenge was to make every answer traceable to a real source, keep retrieval precise across a large Arabic corpus, and refuse gracefully when the corpus has no answer.

Solution

A retrieval layer on PostgreSQL + pgvector with tunable top-k and similarity thresholds, category-filtered search, and Arabic prompt templates; a pluggable LLM provider factory (OpenAI, Gemini, OpenRouter, local LLMs); and Celery + Redis workers for asynchronous chunking, embedding, and indexing.

Architecture

Async FastAPI · SQLAlchemy 2 + Alembic · PostgreSQL + pgvector · Celery + Redis workers · pluggable LLM providers · Docker Compose

Results

• Corpus: ~80K passages
• Categories: 10
• Citations: Book · author · page
• Source-cited answers with refusal when no relevant source is found
• Tunable top-k and similarity thresholds to balance precision and recall
• Pluggable LLM providers (OpenAI, Gemini, OpenRouter, local)
• JWT auth with admin approval and audit log, rate limiting, multi-turn conversations
• pytest unit + integration tests, Docker Compose deployment

Technologies

• Python
• FastAPI
• PostgreSQL
• pgvector
• Celery
• Redis
• SQLAlchemy
• Docker
• pytest
• RAG