töökratt
Job boards dump hundreds of listings on you. I wanted something I could ask in plain language (backend in Denmark, remote ok) and get answers grounded in real posts, with a way to say I don't know instead of inventing roles.
I built a RAG pipeline: ingest jobs from The Hub API, embed them in Qdrant, and chat over the index (FastAPI, Gemini 2.5 Flash) with citations. The first production request OOMed because the embedding model loaded in-process on a 512MB box. I moved embeddings to Qdrant Cloud Inference, benchmarked two free models on the real 1,015-document corpus, rejected one because its scores couldn't support a hard cutoff, and reindexed without taking search down.
The copilot is live. I also run evals (unit, retrieval, generation) and Grafana/Loki, including an alert on prompt-injection attempts.