No More Hallucinated Citations: A Domain-Specific RAG System with Ollama, ChromaDB and AI Agents

A local, domain-specific RAG system was built on 2,514 multilingual academic PDFs about urban art.

The pipeline uses paragraph-level ingestion, MariaDB, ChromaDB, Ollama, FastAPI, and six Claude Code agents.

It supports end-to-end research with zero hallucinated citations by grounding answers in queryable source documents.

The approach shows how self-hosted retrieval can improve citation accuracy and research quality without sending data off-device.