Documentation · Chapter 02

AI agent knowledge base

The AI agent knowledge base in Kolio is the collection of facts about your company — services, prices, opening hours, locations, terms — from which the agent answers customers on the phone, in website chat and by email. This chapter covers the sources (website, PDF, manual text), the AI build, versions with restore, limits and credit usage.

By the Kolio teamupdated 11 min read15 min

What is the knowledge base?

AI agent knowledge base is, in Kolio, one structured document of facts about your company that the AI builds from your sources or that you write by hand, and that the agent receives in every conversation.

You manage the knowledge base in AI agents → Knowledge base (permission baza-znanja). The page has three parts: adding sources (Website, PDF document), the Built knowledge base card with the Build or Refresh button, and the editor with Version history. The newest version is always the active one.

Kolio AI agent knowledge base — website sources with fetched pages, the built knowledge base and version history with the active badge
AI agents → Knowledge base: sources, built base and version history.

How the knowledge base reaches the agent

In every conversation the active version of the knowledge base is inserted into the agent instructions at the {{BAZA_ZNANJA}} placeholder (see AI agent instructions). In addition the agent has the search_knowledge_base tool, which finds up to 3 most relevant sections of the base (at most 1,500 characters) for the customer's question by keyword scoring — no vector embeddings, because it is one document per account.

Which sources are available?

A knowledge base in Kolio is made from three kinds of sources: a captured website, a PDF document and text written or corrected by hand. Website and PDF are "sources" in the Sources list from which the AI builds the base; manual text goes directly into the editor.

Knowledge base sources in Kolio
SourceHow to add itStatus
WebsiteWebsite → enter the URL (your own or a third-party domain, e.g. a manufacturer's manual) → choose Only this page or Discovered subpages too → Check → tick pages → Download selectedworks; you can add several sources
XML sitemapXML sitemap → enter the sitemap address (e.g. https://company.com/sitemap.xml; sitemap indexes and .xml.gz are supported) → Check → tick pages → Download selectedKolio lists every URL from the sitemap without following links; it downloads only the ticked ones and the source list shows “X of Y found selected”
PDF documentPDF document → Upload PDF (max 20 MB, PDF with a text layer only)Kolio checks the file, extracts its text and adds the source to the list; a scanned PDF without text is rejected
Manual textEdit knowledge base → write or correct Content → Save new versionworks; without any sources you can write the whole base by hand

The total size of all sources is limited to 20 MB; the line "Used … of 20 MB" below the source list shows the usage.

How does Kolio capture a website?

Kolio does not crawl a website blindly: it first reads the sitemap, then ranks the candidates by usefulness for customer conversations and captures at most 15 pages with at most 12,000 characters of text per page. Pages with facts take priority over the catalogue, while foreign-language versions and shopping flows are skipped.

  1. The sitemap comes first. Kolio looks for the sitemap in robots.txt (the Sitemap: line), at /sitemap.xml and /sitemap_index.xml; it reads at most 5 sitemap files and scores at most 5,000 URLs. When there is no sitemap, candidates come from links discovered during the visit (BFS).
  2. Ranking by usefulness. Every URL gets a score: matches of fact patterns (about, contact, shipping, payment, returns and refunds, warranty, terms, FAQ, opening hours, services, pricing, locations, inquiry) earn a large bonus, top-level pages (/contact) a flat bonus, deep subpages and query parameters a penalty. The entry URL is always first.
  3. Skipping useless pages. Cart, checkout, login, search, categories and tags, individual products (/product/…), yearly blog archives, authors and system paths (/wp-json, /feed) are not captured; the same goes for paths with a language prefix (/en/, /de/, /hr/ …), because for the agent they are duplicate content in another language.
  4. The 15-page limit. When the site is larger, the source label says "first 15 pages" and a Fetched pages list appears under the source. The capture fails with an error if the page has fewer than 50 characters of usable text.

How do you build the knowledge base?

Building the knowledge base in Kolio turns all sources with text into one structured document: the AI extracts the facts useful for customer conversations (offering, procedures, terms, opening hours, locations, contact, FAQ) and drops navigation, buttons, cookie notices and marketing copy. The result is saved as a new version with origin AI-generated.

  1. Add a source

    In the Website section enter the address of the site (for example https://www.your-company.com) and click Fetch content. The source first appears as pending and, after the capture, with a page-count badge; on failure the cause is shown under the source.

    AI agents → Knowledge base → Website
  2. Check the fetched pages

    Expand Fetched pages under the source. If an important subpage is missing, add it as a separate source; if several sources refer to the same site, the Overlapping sources warning appears.

  3. Build the base

    In the Built knowledge base card click Build (Refresh when a base already exists). The build takes from a few seconds to about a minute depending on the number of sources (every 4 sources are one model call). On failure Build failed is shown with the cause.

    AI agents → Knowledge base → Built knowledge base → Build
  4. Review and correct

    Read the built content in the editor. Correct missing or wrong data and save with Save new version (origin manually edited). The meta line shows the number of characters, the number of sources and the build date.

  5. Test in the simulator

    In AI agents → Call simulator ask about a price, opening hours and a location. The answers must come from the base; for a question without an answer the agent must offer handover to your team.

    AI agents → Call simulator

What happens to duplicate pages during the build

Before sending anything to the AI, Kolio removes pages that repeat across sources — by URL, or for older records by a content signature (the first 2,000 characters). It keeps the first occurrence; a source left without pages drops out. The number of sources after this step determines the price of the build, and the number of unique pages is recorded with the version.

The build runs in batches: 4 sources per model call, at most 16,000 characters of text per source; if several partial summaries result, a second call merges them into one document. All calls of one build are one group in the Usage overview.

How do you edit the base and restore an older version?

Every change to the knowledge base in Kolio is a new version: a build creates a version with origin AI-generated, saving in the editor a manually edited version, and a restore writes the old content as a new version. The history is never deleted and never overwritten.

Actions on knowledge base versions
ActionWhereWhat results
Build / refreshBuilt knowledge base → Build / Refreshnew AI-generated version; supersedes manual edits (the old version stays in history)
Manual correctionEdit knowledge base → Content + Version note → Save new versionnew manually edited version; at least 20 and at most 200,000 characters
Discard unsaved changesDiscard changesthe editor returns to the active version, nothing is saved
RestoreVersion history → Restore on the chosen versionnew manually edited version with the note "Obnovljena verzija z dne …" (Slovenian for "restored version from …"); the history is kept

The history shows the last 10 versions with date, origin, note, character count and the email of the person who saved them.

How does a build count against credits?

Building the knowledge base in Kolio costs 50 credits per every started 10 sources (after duplicate pages are removed) and is charged once per build, regardless of the number of model calls in it. Capturing a website, manual editing and restoring a version consume no credits.

Price of a knowledge base build by number of sources
Sources in the buildCredits per buildExample: included plan credits
1–1050Trial: 250 credits → up to 5 builds if there is no other usage
11–20100Starter: 2,500 credits per month
21–30150Growth: 10,000 credits per month

Included plan credits are from Kolio's pricing; the full usage overview is in Usage and billing.

When credits are used up, a build is not possible and the button returns an error with the code quota_credits; in the meantime the agent runs in limited mode (it only takes messages). Buy additional credits or wait for the next billing period — see Subscription.

Limits and what happens when you exceed them

The knowledge base limits in Kolio are defined in code and identical in all plans: 20 MB total source size, 15 pages per website source, 12,000 characters per page during capture, 16,000 characters per source during the build and 200,000 characters of manually edited content. The number of sources itself is not limited.

Knowledge base limits
LimitValueWhat happens when exceeded
Total source size20 MB (UTF-8 text)a new source is rejected with an error about the 20 MB total limit; delete redundant sources
Single PDF20 MBan oversized file is rejected in the browser; the PDF size counts towards the total source limit and the plan's storage
Pages per website source15 most usefulremaining pages are not captured; the source label says "first 15 pages"
Text per captured page12,000 characterslonger text is cut off
Text per source during the build16,000 characters per source, 4 sources per model callthe tail of the source does not reach the AI; split long sources into several sources (subpages)
Manually edited content20 to 200,000 characterssaving is rejected with a notice that the content is too short or too long
search_knowledge_base tool response3 sections, 1,500 charactersthe agent gets only the most relevant sections; the whole base is in the prompt anyway

Common problems and how to solve them

Most knowledge base problems come from the sources: a site without a sitemap, the wrong subpages captured or facts that are not on the website at all. Below are the most common cases and the solution in Kolio.

  • "No usable text found on the page." The site renders content only with JavaScript or blocks automated visits. Enter the text manually in the editor or add a subpage that returns the content as HTML.
  • The wrong pages were captured (blog, products). Add exactly the subpages you want as separate sources (/contact, /pricing, /terms) and remove the entry source with the site root.
  • "Build failed: no sources with text." No source has at least 50 characters of captured text. Check the status of the sources and fix the capture errors.
  • The agent claims a branch does not exist. Locations and the offering live in Settings → Services; the service list takes precedence over the knowledge base. Add the service with its location, not just text in the base.
  • The agent quotes an outdated price. The knowledge base is a snapshot at build time. After a change on the website, capture the source again (delete and add) and click Refresh, or correct the price by hand and save a new version.
  • Manual corrections disappeared. A refresh from sources created a new version. In Version history find the manually edited version and click Restore, then carry the new data into it.
FAQ

Frequently asked questions

How many pages of a website does Kolio capture into the knowledge base?

Kolio captures at most 15 most useful pages per website source with at most 12,000 characters of text per page; pages with facts (about, contact, shipping, terms, pricing, opening hours) take priority, while the cart, login, categories, individual products and foreign-language versions are skipped. Add further subpages as separate sources, up to 20 MB of text in total.

How many credits does a knowledge base build consume?

A knowledge base build in Kolio costs 50 credits per every started 10 sources after duplicate pages are removed and is charged once per build. Capturing a website, manual editing and restoring a version consume no credits.

Can I write the knowledge base by hand, without a website?

In Kolio you can type the whole knowledge base into the Edit knowledge base editor and save it with Save new version (20 to 200,000 characters). Such a version has the origin "manually edited" and the agent uses it exactly like an AI-generated one.

How do I restore a previous version of the knowledge base?

In Version history on the Knowledge base page in Kolio click Restore on the version you want. Its content is saved as a new, active version with the note "Obnovljena verzija z dne …" (Slovenian for "restored version from …"), and the history stays unchanged.

Can I upload a PDF to the knowledge base?

You can upload a PDF of up to 20 MB (price list, terms, catalogue) to Kolio's knowledge base. Kolio checks the file, extracts its text and adds the source to the list; then rebuild the knowledge base so the agent uses the new content. A scanned PDF without a text layer is rejected — copy such content into the editor as manual text instead.

Can't find the answer?

Write to us or open the chat in the app — the Help button at the bottom of the menu hands the conversation to the Kolio team.

We reply during working hours, usually the same day.

Related