Home / Docs / The Knowledge Base

The AnroAgents Knowledge Base

Uploading documents so the agent answers from your real material.

The Knowledge Base of an AnroAgents agent is where you upload the documents the agent should be able to search: price lists, catalogues, brochures, terms and conditions, technical specifications, internal FAQ. When a visitor asks something, the platform finds the passages most relevant to that question and gives them to the agent as the ground for its answer.

Find the Knowledge Base in the AnroAgents dashboard inside the agent form, as its own step, and also as a panel on the agent edit screen. Uploading a document requires the agent to be saved first.

What you can upload

The AnroAgents Knowledge Base accepts PDF, DOCX, TXT, CSV, MD, HTML, JSON and XML files. Text is extracted from each one, split into passages and indexed for search.

How many sources you can add, how large each file may be, how many website pages are read, and how much text in total the agent can hold depend on your plan. Every plan, Free included, has a Knowledge Base; see the AnroAgents pricing document for the exact limits per plan.

Adding your whole website

The fastest way to fill the AnroAgents Knowledge Base is your own website. Paste its address in the Knowledge Base panel and press "Whole website". AnroAgents reads the site's sitemap and follows the links between its pages, reads each page's text, and indexes all of it as one source. Every passage the agent retrieves carries the title and address of the page it came from, so the agent can point a visitor to the right page.

A website counts as one source however many pages it has; what bounds it is the plan's page limit and character limit. When the site is larger than the plan reads, the source is marked "plan limit reached" and the pages nearest the home page and those listed first in the sitemap are the ones kept.

When the website is a shop, AnroAgents reads each product page as a product rather than as a page: it takes the shop's catalogue from its platform where the platform offers one (Shopify and WooCommerce publish their product list, so the whole catalogue is read in a few requests), and otherwise from the product data the shop publishes for search engines (schema.org product markup, as Magento and most shop systems output it) and makes one card per product with its name, category, price, availability, brand, colour, material, size, variants, the captions of its photos and its link. Product pages count against the plan's product limit, not its page limit, so a shop with thousands of products is read in full — up to 50 products on Free, 500 on Starter, 5,000 on Pro and 20,000 on Business. When a shop has more products than the plan reads, the owner gets an email and a dashboard notification once, saying how many products were read and which plan reads the whole catalogue; products past the limit are not in the knowledge base until then. Each product is indexed as its own passage, so a question naming several features at once finds the product that has all of them. Category pages that only list products, and the second and later pages of a listing, are read for their links but not kept: their products are already there as cards, and a category page would otherwise come up in place of the product a visitor asked for. Pages about delivery, returns, the company and the like are kept as before.

Reading runs in the background and takes from under a minute for a small site to several minutes for a large one; a shop with thousands of products is read gently, a few pages at a time, and can take up to an hour. The panel shows "Reading the website…" with the number of products read so far, and then the number of pages and products read. Press the refresh icon on a website source to read it again after you change the site — the new reading replaces the old one, so there are never two versions of the same page. On the Pro plan a website source is also read again automatically every week, and on Business every day, so new prices, stock and products reach the agent without anyone pressing a button; if nothing on the site changed, the reading is not indexed again. The panel shows the date of the last reading. When an automatic reading fails — most often because the site's host started showing bots a security check — the agent keeps answering from the last good reading and the owner gets an email and a dashboard notification saying what went wrong.

The website is read the way a search engine reads it. Pages that robots.txt closes to bots are skipped, pages behind a login cannot be read, a site whose host shows bots a security check (a Vercel or Cloudflare challenge) cannot be read until the host allows our crawler, whose user agent is AnroAgentsBot — the source then says so — and a site that builds its content with JavaScript only shows an empty page to any reader of this kind; for such a site upload the content as a file instead. Headers, menus and footers that repeat on every page are left out except on the first page, where the footer usually holds the address and phone number.

When you create an agent by importing it from a website on the Basics step, AnroAgents offers to read that whole website into the Knowledge Base as well. The option is on by default and runs as soon as the agent is saved.

To index a single page instead — a price page or a policy page — paste its address and press "This page only". A single page takes one source of its own.

What makes a good knowledge document

A good AnroAgents knowledge document is written for someone who knows nothing about your business. Passages are retrieved individually, so a paragraph that says "it costs €40 including installation" is useless if nothing in that paragraph says what "it" is. Name the product in the sentence that gives its price.

Split a large AnroAgents knowledge document by topic rather than uploading one enormous file. Ten focused documents retrieve better than one combined one, because each passage then sits among related material instead of between two unrelated subjects.

Remove navigation, headers, footers and legal boilerplate before uploading a file to the AnroAgents Knowledge Base (a website added as a whole has this done for it). Every character counts against your plan's limit, and repeated boilerplate makes retrieval noisier.

Keep prices in the AnroAgents Knowledge Base only if you are prepared to keep them current. An agent quoting last season's price is worse than an agent saying it will check.

Keeping documents up to date

Uploading a new version of a document to the AnroAgents Knowledge Base does not replace the old one — delete the outdated file first, then upload the replacement. Two versions of the same price list will both be searched, and the agent has no way to know which one you meant.

Each file in the AnroAgents Knowledge Base shows its processing status. Processing runs in the background and takes from a few seconds to a couple of minutes depending on the size of the document; the agent starts using the material as soon as the status shows the file is ready.

If a file fails to process in the AnroAgents Knowledge Base, the usual causes are a scanned PDF with no text layer, a password-protected document, or a file that would push the agent past its character limit. A scanned PDF has to be run through OCR before uploading.

A source can also come out ready with a note that some parts were skipped. That means a few passages could not be indexed — the rest of the source works, and the agent answers from it as usual. A very large source that does not finish indexing in the time allowed shows an error saying how much of it was indexed; read it again, and if it keeps failing, split the content into smaller sources.

What the Knowledge Base is not

The AnroAgents Knowledge Base is a search over text, not a database with filters. It answers "what does the warranty cover on the X200" well, and "list every product under €50 in stock in Berlin" badly. Structured filtering of that kind belongs in the Shopify integration or in a service call, not in uploaded documents.

The AnroAgents Knowledge Base is private to your agent. Documents are stored for your account, indexed for your agent, and never shared with other customers or used to train any model.