Train a WhatsApp AI agent on your website, FAQ and files

What an agent should read

In Ohanvi the agent answers from a knowledge base, and only from it. A source can be any of six things.

Your website. Give it the address and it follows links on the same site, page by page. It respects robots.txt, waits half a second between pages, and skips any page over 5 MB.

A sitemap. It reads the pages the sitemap lists, which is the cleaner choice when your site has pages you would rather it did not read.

An FAQ. Questions customers ask, with the answer you want given. Each question and its answer is kept as one passage, so an answer is never split in half.

A file. PDF, DOCX, XLSX, TXT, Markdown, CSV or HTML, up to 25 MB. PDFs are read page by page, up to 1,000 pages; a spreadsheet sheet by sheet, row by row, up to 10,000 rows. That is how a price list or a delivery-pincode sheet becomes something the agent can quote.

Pasted text. Policies, scripts or a price list pasted straight in, for the things that live in nobody’s file.

A YouTube video with captions. If your product demo explains it better than any page, the captions become knowledge too.

Test before a customer sees it

Each source is read in the background, and its status updates by itself. Once it is ready, ask it what a customer would. The answer comes back with the passages it was built from, best match first, so you can see whether it found the right part of your price list or the wrong one.

You can test a single source or the whole knowledge base. If an answer is wrong, the fix is almost always in the knowledge, not the model: the delivery charge is missing from the page it read, or two pages disagree.

The test chat on the AI screen works the same way, as a full conversation. Nothing is sent to any customer, and each reply there uses one AI message.

Decide what happens when it does not know

This is the setting that matters most. Every reply is scored on how closely the knowledge it used matches the question. You choose the minimum. Below it, the agent does not answer: the customer gets your handoff line and your team sees what the agent would have said, as a suggestion.

You can lean either way. Set it high and the agent answers only when it is sure, and more chats come to your team. Set it lower and it answers more, with more risk of a weak answer. Most businesses start cautious and loosen it once they have read a week of replies.

One rule is not a setting. A customer who asks for a person, with words like “agent”, “human” or “talk to someone”, always gets one.

Teach it from its mistakes

When a reply is wrong, rate it low and type what it should have said, or say it, in any language. The agent uses your answer the next time a customer asks something similar. You can rate replies in the test chat and straight from the Team Inbox.

A training run goes further. The agent re-reads its websites, looks at what customers asked that it could not answer well, and drafts answers for those gaps. Nothing is added until you review it. Training runs are free.

The agent also keeps a short note on each customer as it talks to them: what they bought, their sizes, the language they use, problems they had. It reads that note the next time they write, so a returning customer is not a stranger.

Keeping it current

Websites change. Re-crawl a source, or every website and sitemap source at once, and only the content that changed is read again. Remove a source and it stops being used in answers immediately.

Pages that are gone or blocked by robots.txt are dropped at the next read; a page that fails briefly keeps its last version, so a short outage on your site does not empty the agent’s memory.

What it costs

Each AI reply to a customer is ₹2 on our key, or 50 paise on your own model key, with no monthly AI fee on any plan. Reading a source costs one AI message per read, and a re-read bills only what changed. Training runs are free.