- Home
- Work
- Bots, AI & automation
- An AI consultant in Telegram: what happens inside
An AI consultant in Telegram: what happens inside
A chat with a consultant for the Pine House project and its inner workings: the intent, the knowledge-base fragments found, the request to the language model, the lead form filling in and the hand-off to the CRM. The search over the base is real and runs in the browser. The bot keeps to its limits: it answers only from the base, names no price without an estimate and asks for a phone number only after consent.
- 9
- knowledge-base fragments
- BM25
- search right in the browser
- 152-FZ
- consent before the phone
The interactive demo did not start in this browser, so a static version is shown.
The task
Show how an AI sales bot works: where its answers come from, how it is kept from “hallucinating” and how a conversation becomes a lead for the manager.
What I did
- A knowledge base of 9 fragments from the project data: areas, heights, materials, the plot, services
- Hybrid search: phrasing rules plus BM25 over the base, with the fragments found and their scores
- Limits: answers only from the base, a price only after an estimate, off-topic questions politely turned back
- A lead step by step: name, personal data consent, the phone via a Telegram button, a viewing time
- A hand-off to the CRM and a message to the manager; an unanswered question becomes a task for the manager
Proof
- The bot runs on the page: ask in your own words or go through booking a viewing
What else you can hand over to me
- An AI assistant for your site that knows your prices and sends leads to Telegram
- A Telegram bot, deployed on a server
- Set up or fix the exchange between 1C and your site
- Deploy a service on a VPS and document it in a clear runbook
- A cloud pipeline for IRL streaming