Guide · step 2 of 4
How to train an AI agent on your own data
The four stages of a working corpus: the initial read, curation, staying fresh, and retrieval at answer time. Why the agent never arbitrates your contradictions — and why there is no embedding index.
Training is a misleading word: no model is retrained. You give it documents, and the quality of the answers is the quality of those documents — no more, no less. That is good news, because it means everything is fixed by editing text.
1. The initial read
The agent reads your public pages and turns them into entries. That is a starting point, not a result: what is online is rarely what you would tell a customer on the phone.
2. Curation — and you are the one who arbitrates
If two prices coexist because an old page is still up, the agent does not pick one. It must not: a model silently "resolving" a contradiction in your pricing is exactly the failure we avoid. You delete the stale entry, or fix the page and resync.
3. Staying fresh
You declare the URLs that move; they resync on their own, daily. An agent quoting last year’s price is not a model error — it is reading an entry nobody replaced.
4. Retrieval at answer time
On every turn the agent sees a summary of everything it knows, and fetches an entry’s full text when it becomes relevant. So it always knows where to look, without depending on lexical luck.
Frequently asked questions
How many pages does it take to work?
Fewer than you think. Ten accurate entries beat two hundred scraped pages: the agent sees a summary of each on every turn, so a short, clean corpus gives it a complete view of what it knows. Volume becomes useful once curated, not before.
Why no vector index or embedding RAG?
Because it is deterministic: the same question finds the same entries, every time. When an answer is wrong you open the entry, fix the text, done — there is no opaque embedding space to debug and no silent index drift after a resync. The trade-off is real: exotic paraphrases match less magically. Seeing every summary on every turn compensates.
My visitors write in a different language from my site. Does it work?
Search works in whatever language your corpus is written in. If your entries are in English and a visitor writes in French, the query may find nothing even though the answer exists. The fix is not a setting: it is having the words your customers actually use inside the corpus.
What about what the agent cannot read on my site?
That is what is worth the most, and only you have it: your real lead times, what you refuse to do, what is already taken. It cannot be scanned — it gets written once, and that is the whole job.