Harran University AI Assistant

Harran University needed a general-purpose assistant that could answer questions about the university in Turkish without sending institutional content to a third-party API.

I built a local RAG pipeline: an async crawler collects pages and office documents, page hashes detect new, changed and deleted content, and only the changed pages are re-embedded. Text is chunked at 1,000 characters with a 50-character overlap, embedded with BAAI/bge-m3 and stored in ChromaDB.

Answers stream from a local Ollama model and include a source URL. If the retrieved context is not enough, the assistant says so explicitly instead of inventing an answer. I also added latency metrics and a tool that compares Ollama models on VRAM, time to first token and tokens per second.