Back to blogGuides

Structuring a knowledge base your bot can actually use

Chunking, collections and the mistakes that make retrieval worse. A practical guide from how Naviooai works under the hood.

An AI assistant is only as good as what it can retrieve. When answers go wrong, the model usually isn't the problem: the knowledge base is. After watching many bots get trained, here is what actually moves the needle.

How retrieval really works

When a visitor asks a question, Naviooai does not read your entire knowledge base. It splits your content into chunks of roughly a paragraph each, finds the handful of chunks most similar to the question, and hands only those to the model, which answers with citations.

That one fact explains most retrieval advice: whatever you upload will be read a paragraph at a time.

Write for the chunk, not the page

  • One topic per section. A wall of text about shipping, refunds and warranties in one breath becomes chunks that are about nothing in particular.
  • Lead with the answer. "Returns are accepted within 30 days" beats three sentences of preamble before the policy appears.
  • Repeat key terms naturally. Retrieval matches meaning, and a section that never names its own subject is hard to find.
  • Keep tables simple. Dense tables survive chunking badly. Short labelled lists work better.

Use collections for scope, not decoration

In Naviooai, knowledge lives in collections and each bot subscribes to exactly the collections it should know. That is a scoping tool, not a filing cabinet:

  • Give the support bot the product docs and policies; give the sales bot pricing and comparisons. Neither can leak the other's content.
  • Update a collection once and every subscribed bot is current immediately.
  • When two bots need the same content, share the collection instead of duplicating it. Duplicates drift.

Crawl, then curate

A whole-site crawl is the fastest way to get started: point it at your site and every page becomes a knowledge item. But crawling is a starting point, not the finish line. Delete the pages that answer nothing (career pages, press releases), and rewrite the ones your customers actually ask about.

Tip: after a week of real conversations, check the knowledge gaps report. It clusters the questions your bot could not answer well, which is exactly your writing to-do list.

Trust the citations

Every Naviooai answer can show the sources it used. If the bot cites the wrong document for a question, that is a structure problem you can fix: split the section, rename it, or move it to a better collection. The citations are your debugging tool.

Good knowledge bases are gardens, not warehouses. Plant deliberately, prune what confuses, and let the gap report tell you what to grow next.

Knowledge baseRAGGuides

Put the ideas to work.

Build an assistant on your own content, free.