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

Retell AI searches the knowledge base automatically before every answer. It turns the recent conversation into a short search query, finds the closest chunks of your documents, and adds them to the agent's prompt under a heading. You don't need to prompt the agent to search, which removes one common failure. What it retrieves, though, is 3 chunks by default, so each chunk has to stand on its own.

This page is written from Retell AI's own documentation, checked on 25 September 2026. KB Builder is not affiliated with Retell AI. 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 Retell calls yet.

How the Retell AI knowledge base works

	What the docs say
Where it lives	The Knowledge Base tab; linked to an agent in its Knowledge Base section. An agent can have several
How the agent uses it	Automatic retrieval before every response, from the transcript so far (not the prompt)
What you can add	Files including .md, .txt, .pdf, .docx and spreadsheets; single pages or whole websites; typed text; Google Drive, OneDrive and Notion
Limits per knowledge base	25 files of up to 50 MB each, 500 URLs, 50 text snippets
Chunk size	Up to 2,000 characters by default (600 to 6,000), set through the API at creation only
Chunks per turn	3 by default (1 to 10), similarity threshold 0.6
Knowledge base instruction	Up to 500 characters; shapes the search query
Extra delay	Generally under 100 milliseconds
Cost	First 10 knowledge bases free, then $8 a month each; $0.005 a minute on calls with a knowledge base

Retell's call history shows, for each turn, the exact text retrieved and passed to the model, which makes debugging much easier than on most platforms.

Where it goes wrong on calls

Three chunks is not much. If the caller asks about "the monthly membership" and the gym has four membership plans, three chunks may not include the right one, and the agent answers with what it has.

A chunk loses what it's about. A chunk of 2,000 characters can start halfway through a product's details. If the product's name was in the previous chunk, the price in this one has no owner. Retell's own advice is the same as ours: include names in the text and avoid words like "it" or "this", "because prior chunks may not be present".

Tables break. Retell notes that retrieval is less reliable for tabular content. A rate card pasted as a table loses its column headings when chunked.

How to load a KB Builder file into Retell AI

A KB Builder knowledge base is a Markdown file written in short sections that each name the product or service they are about. It already follows Retell's structuring advice: focused ## sections, names in every sentence, no tables.

1. Open the Knowledge Base tab, click Add, choose File and upload the .md file.
2. Wait until the knowledge base shows as ready.
3. In the agent editor, open the Knowledge Base section and select it.
4. Optional: raise chunks to retrieve to 4 or 5 for businesses with many similar products, and add the knowledge base instruction below.
5. Make test calls, then open Knowledge Base Retrieval in the call history to see what was retrieved on each turn.

When the business's website changes, rebuild the file in KB Builder, delete the old file source and add the new one. Uploaded files are not refreshed automatically.

Knowledge base instruction

Retell lets you shape the search query in up to 500 characters. This one helps the search carry the product's name even when the caller doesn't repeat it:

Search using the names of the products and services mentioned so far in the call, and the caller's own words for prices, times, what is included and rules.

Lines to add to the agent's prompt

These come 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.