The Voiceflow knowledge base: how it works, and how to load your knowledge base into it

Voiceflow's agents search the knowledge base as a built-in tool. With the knowledge base switched on in the agent's system tools, the agent searches it whenever it needs information, using the caller's last message as the query by default, and gets back 3 chunks unless you raise the limit.

This page is written from Voiceflow's own documentation, checked on 25 September 2026. KB Builder is not affiliated with Voiceflow. Our own live calls so far ran on a different platform, so the advice below applies what we learned there; it has not been tested on Voiceflow yet.

How the Voiceflow knowledge base works

	What the docs say
Where it lives	Knowledge base in the left menu (Shift + K)
How the agent uses it	A system tool the agent uses when it needs information; playbooks can query it with their own settings
What you can add	Files (.pdf, .txt, .docx, up to 10 MB), tables (.csv, .xlsx; each row is a chunk), URLs, sitemaps, plain text, Zendesk, Shopify, API
Markdown	Not listed; paste as plain text or upload as .txt
Chunks per query	3 by default, 1 to 10
Chunk size	About 1,000 tokens by default (500 to 1,500), set through the API
Search query	The caller's last message by default; optional query rewriting and a custom query
Refresh	URL and integration sources: never, daily, weekly or monthly
Going live	Knowledge base changes go live when you publish the environment
Cost	Plan prices are shown inside the product; LLM chunking strategies use credits on every sync

Where it goes wrong on calls

The query is only the last message. "And how much is that one?" is searched as it is, with no product name. Voiceflow's optional query re-writing has the model rewrite the last message before searching, which helps here.

Three chunks miss the right product. For a business with many similar products, 3 chunks may not include the one the caller means. Voiceflow suggests starting at 3 and raising it if the agent misses things.

Rewritten content. Voiceflow's LLM chunking strategies can regroup, summarise or prefix your content before it's searched. "Summarize" keeps only key points, which can drop the exact fact a caller asks about.

How to load a KB Builder file into Voiceflow

A KB Builder knowledge base is a Markdown file already written in short sections that each name one product or service, so it doesn't need Voiceflow to re-chunk or summarise it.

1. Open Knowledge base and click Add data sources.
2. Choose Plain text and paste the whole file, or save the file as .txt and upload it.
3. Leave the LLM chunking strategies off. The file's sections are already self-contained, and this keeps every fact word for word (and uses no credits).
4. On the agent, make sure the knowledge base toggle is on in System tools. Consider a chunk limit of 4 or 5, and switch on query re-writing.
5. Publish the environment, then test with follow-up questions that don't repeat the product name.

When the business's website changes, rebuild the file in KB Builder, replace the old source's text with the new one (or delete it and add the new one), and publish again.

Lines to add to the agent's instructions

Voiceflow's default trigger description already tells the agent never to answer from its own knowledge. Add these, from our own live calls:

Answer questions about the business only from the knowledge base. If it doesn't have the answer, say you will check, and give the business's phone number.
Give a price, time or rule only from a sentence that names the product or service the caller asked about.
Whenever you give a price, say the name of the product or service with it.
If the caller hasn't said which product or service they mean, ask before answering.
When the knowledge base says something has not been confirmed, say exactly that. Never guess.
Don't promise messages, callbacks or bookings unless the knowledge base says the business offers them.