RAGLLMVECTOR-DB 2025
AI Support Agent
agentic retrieval over a corpus of support tickets
A production agent that answers customer-support questions by retrieving relevant past tickets, drafting an answer, and routing edge cases to a human. Built on a hybrid lexical-plus-vector index with rubric-based offline evals.
The hard part wasn’t the model — it was the boring infrastructure: chunking strategy, embedding refresh, eval datasets, and the graceful-degradation logic that decides when the agent should give up. I wrote a longer post on this under writing.