CrustData Customer Support Agent

TL;DR

• Next.js app and FastAPI backend for support workflows.
• RAG pipeline with citations to reduce hallucinations.
• Admin tools to update sources and review responses.

Artifacts

Context

Support teams needed accurate answers grounded in internal docs, not generic chatbot replies.

Problem

Existing bots could not handle nuanced technical questions and made it hard to verify sources.

Approach

• Build an ingestion pipeline with chunking and metadata tags.
• Use vector search with filters to retrieve relevant sources.
• Generate responses with citations and guardrails.
• Create an admin UI for source management and feedback.

Tradeoffs

• Balanced recall and latency by limiting chunk fan-out.
• Kept retrieval logic simple for easier maintenance.
• Optimized for clarity over creative generation.

Testing and Reliability

• Golden question set for regression checks.
• Manual review workflow for edge cases.

Deployment and Ops

• Dockerized services with Supabase auth and storage.
• Environment-based config for model and retrieval tuning.

Outcome

• Faster first-response times for support teams.
• More consistent answers with traceable sources.
• Simplified updates as documentation changes.

If I had two more weeks

• Add evaluation dashboard with answer quality metrics.
• Support multi-tenant routing and RBAC.