// blog
An AI That Answers Once: How Our Repair Assistant Needs Less AI Every Month
Our board viewer looks in its own answer library before it calls the AI. If the same question has been asked about the same board data, the saved answer comes back with no AI call. If not, the AI answers, each value is checked against the board’s data, and a clean answer is saved for the next person. So the cost of the AI is limited by how many different questions exist, not by how many people ask them.
Most AI features have an awkward shape: the more people use them, the more they cost, and it never stops. A question asked a thousand times is paid for a thousand times.
Our free board viewer has an AI assistant for electronics repair, and we didn’t want that shape. Repair questions repeat. The number of boards is finite, and so are the questions worth asking about each one. So the assistant answers a question once, and our own database answers it after that.
The library comes first
When you press Ask, the viewer sends your question and the data for the part or net you’ve selected. The server then does this:
- It looks for the same question with the same data in the library. If it’s there, you get the saved answer at once, with a note saying how many times it’s been served. No AI is called, and it doesn’t count against your daily limit.
- If it isn’t there, the AI answers.
- The server checks the answer. If it’s clean, it goes into the library for the next person.
Case, spacing and punctuation don’t make a question new. Neither does which part you have selected, if the question is about the board as a whole, or which export of the board you opened. Rewording does, for now, so the library fills fastest with the viewer’s example questions. Matching reworded questions is the obvious next step.
Why that makes the bill predictable
The spending has a limit you can work out in advance. Say there are 5,000 boards and 20 common questions for each. That’s 100,000 answers at most, once each, and at an AI cost well under a cent per answer the whole library comes to a fixed sum, even if a million people ask.
Real use won’t be that tidy, because the same question about a different selected net is a different question. So there’s also a hard ceiling: each visitor gets 10 new AI answers a day, the whole site a fixed number a day, and the server adds up what each call cost and stops calling the AI when the month’s budget is spent. Whatever happens, we know the most a month can cost. And answers already in the library keep working after that.
The safeguards
A shared library raises the stakes. One made-up reading, saved and served to everyone, is worse than one bad answer in a private chat. So an answer has to earn its place:
- Only the data you can see. The AI gets the selected part or net, its readings and the repair notes, marked as data and not as instructions. Your board file stays in your browser.
- Every value is checked. A number in an answer has to trace to that data, to your question or to a rail name such as PP3V3_S5. If it doesn’t, you see a warning and the answer isn’t saved.
- Every name is checked. A part or net that isn’t on your board is marked. The ones that are become links.
- Only repair questions. Anything else gets one line back, and that line is never saved.
- Nothing personal is shared. Every question and answer is saved, but anything leaning on your own readings or an earlier question is never shown to anyone else. Email addresses, phone numbers and links are removed before saving. No IP address is saved with an answer.
- A forged request can’t poison it. A saved answer is tied to the exact data it came from. Someone sending fake data can only ever match their own fake data, so they can’t plant an answer that reaches anyone else.
- You can overrule it. Every saved answer has a Helpful button and an Ask the AI again button. The second one gets you a new answer on the spot. Three of those, when they outnumber Helpful, take the old answer out of the library.
When a board’s readings change, the data changes, so the old answer stops matching and a new one is written.
What the checks don’t do
They catch values and names that came from nowhere. They don’t prove an answer right. An explanation can sound sensible and still be wrong, which is why the page tells you to measure before you replace anything.
The same idea works elsewhere
Any business answering the same questions from the same data can do this: support, product questions, policy lookups. Save what’s been checked, answer from your own database first, and call the AI only for what’s truly new. If you’d like AI in your business with a cost you can predict, see how we do it.
