Anyone can put a chatbot on a website. The hard part is what sits behind it. Here’s how we built a checked library for laser education, what we had to chase down along the way, and why it’s the foundation for everything else.
A typical starting point
“Let’s just give the AI our manuals”
Picture a typical laser manufacturer. CO2 cutters, fiber cutters and markers, in a dozen sizes and wattages. A manual for every model, revisions going back years, spec sheets, service bulletins, and a support team answering the same twenty questions every day: it won’t fire, it won’t home, the cut isn’t going through. Sooner or later someone says it: let’s upload all of it to an AI and put a chatbot on the website.
It’s a reasonable idea, and it demos well. Ask a question, get a fluent answer. The trouble shows up later, in the answers nobody checked:
- Specs come apart. Documents get cut into pieces before an AI searches them, and a table can be split so a value lands away from the heading it belongs to. The AI fills the gap with a guess. (VentureBeat, 2026)
- Long documents lose their middle. Researchers found language models use information at the start and end of a long input better than what’s buried in the middle. (Liu et al., 2023)
- Tables stay hard. A 2026 benchmark of questions that mix text and tables found them challenging even for the best current models and retrieval methods. (Strich et al., 2026)
- It invents what isn’t there. When one software company’s support bot made up a login policy that didn’t exist, users reportedly cancelled before the company apologized. (Fortune, 2025)
- The company owns the answer. When Air Canada’s chatbot gave a customer the wrong refund policy, a Canadian tribunal held the airline responsible for what its bot said. (Moffatt v. Air Canada, 2024 BCCRT 149)
I know this one first-hand. Years ago I fed a stack of laser owners’ manuals into an AI and tried to use it as a library. Sometimes the answer was right. Often it blended two manuals, picked the wrong model, or filled a gap with something that sounded sure and wasn’t.
That taught me the lesson this whole project is built on: raw data is too much. An AI can’t tell which part of a pile it’s drawing from, and it sounds just as confident either way. Before an AI can be trusted with information, someone has to distill it: read it, check it, keep what holds up, and throw out what doesn’t. And the problem is growing, as more of the web is written by AI and wrong information gets repeated until it looks settled.
Laser TeamWork is where I’ve been doing that distilling, in the open, so anyone can see how it works.
What we built
Laser TeamWork: a checked library with an assistant in front of it
Laser TeamWork is a free, vendor-neutral resource for teachers who run laser cutters in schools. Every answer on it comes from published sources: federal rules, the ANSI laser safety standards, OSHA, university safety offices and research. The assistant answers only from that library, shows its sources, and says so when the library doesn’t cover something.
136
answers published, each one a full page
151
distinct published sources, from 81 organizations
350
source citations across those answers
131
checked entries the assistant answers from
236
sources re-opened to check one batch of 65 answers
78
pieces of evidence checked for one reader’s question
The method
Nothing goes in until it’s been checked twice
Every answer and every library entry goes through the same four steps. The second one is the step most people skip, and it’s the one that matters.
1
Research
Find what the published sources actually say, and link every claim to the source it came from. No forums, no sales pages.
2
Check
A separate pass re-opens every source and reads it again. Anything the source doesn’t really say gets reworded or cut.
3
Library
Only what survives the check goes in. Fix something once and it’s fixed everywhere it’s used.
4
Answer
The assistant writes from the library only. Each paragraph is checked against its source before anyone sees it, and gaps are named, not filled.
What it is
- One place where every answer traces back to a published source
- A library that gets stronger as it grows, because new material goes through the same check
- An assistant that says “the library doesn’t cover that” instead of guessing
What it isn’t
- A smarter chatbot trained on the whole internet
- A pile of documents dropped into an AI and hoped for the best
- Finished: some answers still say plainly what they can’t tell you yet
Under the hood
This isn’t a standard AI chatbot
A standard chatbot is one language model doing everything: deciding what you meant, finding information, writing the answer and judging whether it’s right, all in one pass. That’s where the confident wrong answers come from.
Here, the language model does one job, and the decisions are made around it:
- Decide
no chat modelA small classifier, built to sort questions rather than write, decides the topic and whether to stop for safety. Three separate safety checks run, and any one of them stops the answer. - Fixed answers where they belong
no chat modelPrice, “talk to a person” and safety stops get answers we wrote ahead of time. No model improvises them. - Find
no chat modelA search over the checked library only, never the open web. - Write
language modelThe model writes from the entries it was handed, and names the entry under every paragraph. - Check
second passEach paragraph is checked against its entries. Anything unsupported is cut. If nothing is left, the answer says the library doesn’t cover it.
For hands-on troubleshooting the same idea goes further, into a decision engine. A flowchart built from a technician’s real cases decides every step, safest and most likely first. The model’s only job is to understand the customer’s reply (“the fan is spinning but nothing fires”) and match it to the right branch. It never invents a check.
The receipts
Things we had to chase down
This is what “checked” means in practice. Each of these read fine at first. Each one failed when someone went back to the source.
“Enclosed laser cutters are often Class 1.”
It’s repeated everywhere. Then we looked at four manufacturers’ own certifications: Class 2, Class 2, Class 2 and Class 4.
The claim was deleted
“Operator authorization commonly lasts one academic year.”
A tidy, believable rule for a school. We looked for a source behind it and found none anywhere.
Removed: it was invented
“How often do laser cutters start fires?”
A fair question with no honest number. The old national fire-reporting system had no code for a laser cutter, and its replacement’s finest category is “shop tools.” The missing statistic isn’t a gap in the research. It doesn’t exist.
The page says so instead of guessing
“Class 4 additionally has skin hazards.”
Sounds right. But the FDA’s own table already lists a skin hazard from the direct beam at Class 3B. What Class 4 adds is the hazard from reflected and scattered light.
Reworded to match the federal table
A university safety sheet, dated July 2024.
That’s how it had been cited. When the PDF was re-opened, it read “Reviewed July 2026” on both pages, with no 2024 date anywhere.
Corrected: dates get checked too
“Texas requires registration before a school uses a Class 3B or 4 laser at all.”
Close, but stronger than the rule we could point to. The checked version says what the Texas code actually supports: schools must register those lasers.
Narrowed to what the source says
These aren’t rare. When we checked a batch of 65 answers before publishing them, 41 needed a fix, and each one got a sourced edit before it went live. That’s the work an AI can’t do for itself: it repeats what it read, whether or not the source really said it.
One question, start to finish
“What’s the difference between Class 3B and Class 4?”
A reader asked the assistant this during a test. Here’s what happened.
- The first answerThe library didn’t cover the details yet, so the assistant explained what it could and said plainly what it couldn’t. It didn’t make up the difference.
- ResearchA research pass went to the federal laser standard, the ANSI education standard, OSHA, and state rules, and wrote up what they say, with a link for every claim.
- The checkA separate pass re-opened all 78 pieces of evidence. 69 held up as written. The rest were reworded, downgraded or rejected.
- Into the libraryFour new entries went in. From then on, every answer that touches laser classes uses them.
- TestedThe reader’s exact question was run three times. It answered correctly every time, with the federal and ANSI sources behind it.
That was most of a day’s work for one question. It’s also why the answer can be trusted.
Why it matters beyond one site
AI tools are looking for exactly this
AI search tools tend to quote sources that are clear, consistent and cited. A single site where every answer has been checked is what they’re looking for. So the goal isn’t to compete with ChatGPT. If ChatGPT starts answering laser-safety questions from Laser TeamWork, that’s a win: the right answer travels further.
And the library is a foundation, not a feature. Anything new, whether it’s a reader’s question, a teacher’s experience, or a correction, goes on top of what’s already been checked, through the same check. So it gets stronger as it grows instead of looser.
For laser manufacturers
Aligning an AI with your products and the way you work
You’ve probably heard “alignment” used about AI. In practice it means getting an AI to answer the way you would, from the information you trust, and to stop when it doesn’t know. For a laser manufacturer, that’s your manuals, your spec sheets, your service history and what your best people know, distilled the same way Laser TeamWork is. Here’s where it pays off:
- Sales: a buyer says what they want to cut or mark. The material usually decides the technology, CO2, fiber or UV, and then it’s a matter of the power and the bed size that fit the work. The answer narrows to the right machine from your own spec data, with the reason shown. Live details like price and stock are read fresh, never stored.
- Support: a customer whose laser won’t fire is walked through it the way your best technician would do it, in the order your real cases say works, with the safety stops first and anything inside the cabinet handed to a technician.
- The handoff: when it reaches the edge of what’s been checked, it passes the case to a person with a summary of what was asked and what was already tried, so nobody starts over.
- It keeps learning: every answer your team gives that the AI couldn’t goes through the same check and into the library. Your people answer questions; the process does the writing up.
A shared base and your own layer: general, published knowledge, the kind Laser TeamWork is built from, sits underneath, and your product knowledge sits on top and stays yours. It starts the same way Laser TeamWork did: with the distilling, not the chatbot.
Watch it
The two-minute version
One question, start to finish: the gap, the research, the check, and the answer it gives now. The narration is my AI voice clone and the background pictures are AI-generated; the screens and the numbers are real.