Skip to content

HOME WORK 04 / 04

A SALES ASSISTANT GROUNDED IN YOUR OWN CATALOGUE

For a security technology firm — an assistant that answers from the company's own product documents, cites its source, and flags a sales opportunity.

FIG. 07 — ASSISTANT

01 THE PROBLEM

Most incoming questions were the same ones: what is the difference between these two devices, is there a certified model, which one fits this site. The sales team spent part of every day re-explaining the catalogue. An off-the-shelf chatbot, meanwhile, does not know the products — it assumes it does and invents.

02 THE APPROACH

We connected the assistant to the firm’s own product documents. A question first triggers a semantic search across those documents; the passages found are handed to the model, and the answer is produced from those alone.

The real problem was not answering correctly — it was being able to say “I don’t know”. Asked for a technical detail that is not in the knowledge base, the assistant does not invent one; it says so and hands the visitor to the sales team.

When a conversation carries buying intent it is flagged as a lead, which makes the assistant a channel rather than a cost line.

03 BUILT WITH

  • Python
  • FastAPI
  • Claude API
  • Vector search
  • SQLite

04 THE OUTCOME

  1. A quality suite of 100 questions across 5 scenarios, written and run

  2. Hallucination tests included; the single invented answer was caught and fixed

  3. Outside the knowledge base, the assistant hands over to the sales team

  4. Conversations carrying intent are flagged as leads

01 / 04 NEXT

FIG. 04 — FLEET

05 CONTACT

HAVE A JOB LIKE THIS?

Describe the problem; in the first call we work out the scope and a rough budget together.

Start a project