RAG Support Agent
Retrieval-grounded ticket resolution over the brochure corpus, with structured output and an explicit escalation path to a human for refunds, complaints and account issues.
Designed the retrieval pipeline, output contract and escalation logic
The brochure corpus is chunked at 800 tokens with 80 overlap into a Qdrant vector store using Gemini embeddings, retrieved top-8. The model is constrained to answer only from retrieved passages and emits a structured object carrying both the reply and an escalation flag. The constraint is the product — a confidently wrong reply about a refund costs more than no reply at all.
