What grounding actually is
When a customer asks a question, the bot searches your documents, finds the passages that match, and answers from them. It is not remembering your return policy, it is reading it, at the moment of the question.
That distinction matters for two reasons. Your policy can change on Tuesday and be correct in answers on Tuesday. And when an answer is wrong, the cause is almost always that the bot found the wrong passage, or missed the right one, rather than a flaw in the AI itself.
Searching by meaning alone is not enough
Most bots search by meaning, which handles paraphrase well. Someone asking “can I send this back” finds a passage about returns without sharing a single word with it.
It handles exact codes badly. A part number like RX-4410-B, an order number, a specific error code: these are exactly the questions where searching by meaning is least useful, because the code that matters has no meaning to match.
The fix is to do hybrid search: search both ways at once, by meaning and by the exact words. ProxyAI does this by default, so a question containing a product code works as well as one written as a sentence.
If you are comparing platforms, ask this directly. “We use AI search” says nothing about whether a customer quoting a part number will get an answer.
How your documents are written matters more than any setting
When you import a document, ProxyAI’s document importer splits it into short passages by itself, so there is nothing for you to set up. What you control is how clearly each document is laid out. A document with clear headings, lists and tables is split in sensible places. A wall of plain text is not, and a passage can end up without the context that made it meaningful, such as a price with no sign of which product it belongs to.
ProxyAI’s AI formatter does the tidying for you. After you upload a document, press Clean convert markdown (AI powered) on the review screen. It restores headings, lists and tables, and removes page numbers and repeated headers, without rewriting or dropping anything. It uses a little of your credit.
Rewriting your top ten support documents with real headings, or running them through the AI formatter, is usually a bigger accuracy win than any tuning.
The discipline that matters most: what you do not upload
The instinct is to upload everything. It is the wrong instinct.
The bot finds what best matches the question, not what is most current. Leave last year’s shipping policy next to this year’s, under a different filename, and a customer will eventually be told the wrong thing, confidently, with the outdated passage as the reason. The bot did not hallucinate, it read a document you gave it.
Three rules that prevent most of this:
One current version of each thing. Delete superseded documents rather than keeping them “for reference”. The bot has no concept of reference.
Keep the same filename when you update. In ProxyAI, uploading a document with the same filename replaces the old version, so the policy updates instead of piling up as two contradictory copies. A new name, such as shipping-2026.pdf next to shipping-2025.pdf, keeps both.
Review before you publish. Reading text out of PDFs is imperfect. A table that turned into a wall of numbers is worse than no table, and you can only see that by looking.
Scanned PDFs need a decision, not a default
A scanned or photographed PDF has no readable text, so the bot sees nothing. Text recognition (OCR) fixes this, but it adds cost and is never perfect.
So decide per document: turn it on for documents that really are scans, and leave it off for the rest. ProxyAI does this on your own computer, so scanned documents are not sent to a third-party service, worth knowing if the scans are contracts.
Each bot should only see its own documents
If you run more than one bot, each one needs its own separate set of documents.
Keeping them separate protects you in two ways. The first is privacy: one client’s bot can never see another client’s documents. The second is getting the right answer: if two brands with different return windows shared one set of documents, the bot could answer with the wrong company’s policy. With separate sets, each bot only knows its own.
How to tell whether it is working
Test in both directions, and the second one is the one people skip.
Ask what is covered. Questions your documents answer directly. Check the answer matches the source rather than a plausible paraphrase of it.
Ask what is not covered. Questions near your documents but not in them. The correct behaviour is admitting it does not know. A bot that confidently improvises here will do the same thing to a customer, and you will not find out until they act on it.
Then look at the questions the bot could not answer. Those are your content gaps, stated as questions real people asked, a better list of what to write next than anything you would produce in a planning meeting.
