DocuMind

AI document intelligence platform with RAG, FastAPI, async workers, and pgvector.

Architecture

• API: FastAPI + async SQLAlchemy
• Workers: Celery + Redis
• Embeddings: OpenAI + Amazon Bedrock adapters with factory selection
• Retrieval: pgvector cosine similarity via RAG pipeline
• Evaluation: golden eval harness with faithfulness, grounding, and MRR@k

Local setup

python -m venv .venv
.venv\Scripts\activate
pip install -e ".[dev]"
cp .env.example .env
python -m pytest tests/ -q

ADRs

• pgvector over managed vector DBs: simpler topology, same ACID guarantees.
• Celery/Redis over Airflow: direct FastAPI integration, lower ops overhead for ingestion frequency.
• SQLAlchemy 2.0 async ORM with explicit table inserts for SQLite compatibility.