How to Improve Botpress Knowledge Base Accuracy

October 26, 2026

Botpress knowledge base accuracy depends on document structure and chunk size far more than on which language model you select. Rewriting five help articles into question-answer format and cutting chunk size in half took one support bot from 54% correct answers to 81% in a week.

That’s the whole argument. Model swapping is the first thing people try and the last thing that helps.

Why does my Botpress bot give wrong answers?

Mostly because the retrieval step fetched the wrong document. The bot isn’t hallucinating — it’s reading a paragraph about enterprise pricing when the user asked about the free tier, then answering confidently from it.

Botpress splits uploads into chunks, embeds them, and at query time pulls the closest matches. Three things break that pipeline:

  • Documents that cover many topics. A single “Billing FAQ” page with 20 unrelated answers produces chunks that straddle topics and match loosely.
  • Chunks too large. A 900-word chunk gets averaged into an embedding that represents nothing precisely.
  • Contradictions across files. Two docs with different prices produce two high-scoring chunks and an arbitrary pick.

How do I restructure documents for better retrieval?

Write for retrieval, not for reading. The single biggest win: one question per heading, with the answer directly under it.

Before:

Our pricing is designed to scale with your needs. We offer several tiers starting at $19, with enterprise options available for larger teams…

After:

What does the Starter plan cost? Starter is $19 per month for one workspace and includes 2,000 messages.

That second version embeds cleanly because the chunk is about exactly one thing. Do this for your top 30 questions and reupload. It’s tedious. It works better than any setting.

What chunk settings should I use?

Smaller than default. For support content, target chunks that hold one complete answer — roughly 200 to 400 words. Anything past 600 words dilutes the embedding.

Test it directly: after uploading, open the knowledge base explorer and search a real customer question. Look at the top three returned chunks. If none of them actually contains the answer, no amount of prompt tuning will save the bot.

How do I test accuracy without guessing?

Build a test set of 50 real questions from your support inbox, with the expected answer noted next to each. Then run them through the console manually, weekly. Score each as correct, partial, or wrong.

You’ll find a pattern fast. For most teams it breaks down like this:

Failure type Share of wrong answers Fix
Wrong chunk retrieved ~60% Restructure docs
Correct chunk, wrong synthesis ~25% Tighten agent instructions
No relevant chunk at all ~15% Add missing content

That table is worth more than any model benchmark. Sixty percent of your errors are a document problem.

Do better models fix accuracy?

Somewhat, and less than vendors imply. Moving from a smaller model to a frontier one buys you better synthesis on ambiguous questions, which is that 25% band. It does nothing for missing content or bad chunking, and it costs more per message.

On Botpress, model choice lives in the agent configuration. If you’re on the free or Plus tier, check which models are available before you plan around one — availability shifts with plan level.

What about agent instructions?

Keep them short and negative-leaning. Tell the bot what to do when it doesn’t know:

  • Answer only from the knowledge base. If nothing relevant appears, say so and offer a human handoff.
  • Never quote prices not present in the retrieved content.
  • Keep replies under 80 words.

That third rule does more for perceived quality than you’d think. Long bot answers read as padding.

How long does this take?

Budget two days of real work for a 100-article knowledge base. Day one to restructure and reupload. Day two to build the 50-question test set and iterate on the worst ten failures.

Skip it and you’ll spend those two days anyway — just spread over a month of confused customers and a support team that stops trusting the bot.

Key Takeaways

  • Botpress answer quality is driven by document structure, not model selection.
  • One question per heading with the answer directly beneath it embeds far better than long FAQ pages.
  • Target chunks of 200-400 words; above 600 words retrieval accuracy drops.
  • Roughly 60% of wrong answers come from retrieving the wrong chunk.
  • Build a 50-question test set from real tickets and re-run it weekly.
  • Restructuring a 100-article base takes about two working days.

Platform limits and plan details are on the Botpress tool page. If you haven’t built the agent yet, start with the basic build walkthrough first.