AI Customer Support Assistant Prototype

I want to prove that a Retrieval-Augmented Generation (RAG) workflow can cut repetitive tickets out of our queue. The prototype must accept incoming customer emails, interpret the question in natural language, search our internal product documentation, FAQs, and support policies, and then draft a concise, helpful reply that cites the source material it relied on.

If the model’s confidence dips below an agreed threshold it should automatically pass the thread, along with its reasoning chain and conversation context, to a human agent so nothing gets lost in hand-off.

Core tech I expect to see includes:
• an embedding store or vector DB for fast retrieval,
• an LLM orchestrator that injects the retrieved snippets into the prompt, and
• a lightweight email-style chat interface where I can watch the system work in real time.

I will judge success on response accuracy, average response time, resolution and escalation rates, and a short customer-satisfaction pulse survey you’ll wire in.

Deliver to me:
• the runnable prototype (code and simple UI),
• a seeded knowledge base built from the docs I supply,
• setup instructions so I can reproduce results locally, and
• a brief evaluation report that benchmarks the metrics above.

I’m happy to clarify edge cases or provide redacted tickets once we begin.