A Sales Chatbot for Heavy-Equipment Automation

The client sells and installs guidance systems for excavators, dozers and graders, priced in the tens of thousands of dollars per machine. The idea, as I put it to the owner, was that buyers see the content, get curious and ask their real questions on the site before talking to anyone: will it fit my machine, what's the difference between 2D and 3D, what does it cost. Their new site already had a step-by-step system builder and payback calculators. The chatbot had to answer those questions with the same facts, on any page, and save every question so the team can see what buyers ask.

The first version showed what goes wrong. A question about a Cat 320 excavator came back with Volvo passages, and on the first run of the 14-question test set, 9 questions got answers, 3 timed out and 2 errored. Price needed its own rule: a figure in chat has to match what the site's builder charges for the same system.

I gave the chatbot tools instead of letting it improvise. It searches a prebuilt index of passages from manufacturer manuals and brochures, plus a glossary of trade terms. When a buyer names a machine make, results are filtered and ranked to that make, and for a fit or install question that names a make and a model number, installation and compatibility documents rank above troubleshooting pages. An answer cites at most 8 sources, each with its document title, plus the page number and a page or product image when the record carries them. Prices work differently. Dollar figures are stripped from the passages before the model reads them, its instructions forbid quoting a price from a passage, and its pricing tool calls the same function the site's builder uses. Items without a sourced price show as price on request. Two routes skip the model entirely: agriculture questions, a market the client doesn't serve, get a polite decline, and a request for a person goes straight to a handoff. When a buyer is ready, the chatbot prepares a quote, demo or callback request with the machine, job and location from the conversation filled in, and the buyer confirms it before anything is sent. Model calls run on tight clocks: 6 seconds for the first chunk of a response and 20 seconds per attempt, and a labeled quick answer built from the retrieved passages when the model path falls back. Conversations are saved on the buyer's device, and the site sends buyers' questions to a log the team can read. The client's internal platform has its own chat over its indexed document library, which searches by meaning and keyword together, cites the page when the passage carries one, and refuses when it finds no evidence.

• Prebuilt passage index with make-aware ranking
• Shared pricing function with the online builder
• Tool calls for fit, sources, builds, prices and requests
• Per-attempt time limits with a sourced quick-answer fallback
• Hybrid meaning-and-keyword search with page citations
• Question log in Convex for the content team

On the same 14-question test set, answered questions went from 9 to 14, with no timeouts and no errors. In a 20-request load run, all 20 finished and none errored. Questions about Cat, Deere and Komatsu machines now return documents for those makes. Answers show their sources, and the pricing tool returns the same figure as the builder for the same build.

A buyer gets an answer with its sources without waiting for a call, and a quote request the buyer confirms starts with the machine, job and location already filled in. The question log is there for the team's weekly content plan: what buyers ask is what the next posts and pages should answer.

Keep prices out of what the model reads. Here dollar figures are stripped from the passages before the model sees them, and its pricing tool runs the builder's own code, so the chatbot's route to a price is the same function the builder uses.