A customer-support chatbot that answers from your own help docs, not from thin air. Retrieval-augmented generation grounds every reply in real sources, cites them inline, and hands off to a human the moment it isn't sure.
The support queue was 70% the same dozen questions — refunds, setup, billing — and the team answered each one by hand. Off-the-shelf chatbots made it worse: they made up confident, wrong answers that created more tickets than they closed.
The brief: a chatbot that only speaks from the company's own documentation, shows its sources, and admits when it doesn't know — escalating to a human instead of guessing.
Help center, product docs, policies and macros are chunked, embedded, and stored in pgvector. A hybrid of keyword (BM25) and semantic search finds the right passages — even when the customer's words don't match the docs'.
For each question the top passages are retrieved, reranked, and handed to the model with a strict instruction: answer only from these sources, cite them, and refuse if they don't cover it. Every reply links the exact docs it used.
When confidence is low or no source fits, the chat hands off to a human with the full transcript. An analytics view tracks deflection, grounding rate, and the questions the docs don't yet answer.
It answers the repetitive questions in our own words, with the link to prove it — and when it doesn't know, it passes the ticket to us instead of inventing an answer.
The repetitive majority — refunds, setup, billing — answered instantly and correctly, freeing the team for the cases that actually need them.
Retrieval, reranking, and a streamed response land a cited answer in about a second — faster than finding the article yourself.
Every reply is backed by retrieved passages and cites them; with no supporting source, it refuses and escalates rather than guessing.
RAG chat, knowledge base, and analytics — the full support copilot loop in one product.
Every answer drawn from real docs with inline citations.
Help docs, policies, macros — chunked, embedded, and searchable.
Deflection rate, grounding rate, and unanswered question gaps.
Flagged conversations routed to a human when confidence drops.
Tone, fallback behaviour, and escalation rules — all configurable.