Skip to content

Knowledge Bases (RAG stores)

A Knowledge Base lets an AI Agent answer from your own documents instead of only from what the model was trained on. You feed it your content - a product manual, a policy handbook, your support articles - and at question time the agent can pull back the passages that match what it was asked, so its answer is grounded in your material. This retrieve-then-answer approach is the pattern commonly called RAG, short for retrieval-augmented generation.

How it works

A Knowledge Base does three things: it stores the documents you give it, it indexes them so their meaning can be searched, and at question time it finds the pieces most relevant to what was asked. So an AI Agent can answer from those pieces instead of guessing from its general training.

The indexing is what makes a meaning-based search possible. When you add a document, FlowRunner does not keep it as one long block of text. It breaks the document into smaller pieces, called chunks, and turns each chunk into an embedding - a list of numbers that captures the meaning of that piece of text. Those embeddings live in a vector store, a kind of database built to compare meanings and return the closest matches.

So when an AI Agent asks the Knowledge Base a question, the question is turned into an embedding too, and the store hands back the chunks whose meaning is nearest - not the ones that happen to share the same words. That is how a Knowledge Base answers from your content: it finds what is relevant by meaning, and the agent writes its answer from those passages.

A Knowledge Base is a standing resource: you set it up once, then feed and manage its documents over time. How it slices and embeds your text is fixed when you create it.

When to use it

Reach for a Knowledge Base when you want an AI Agent to answer from facts you supply - a product manual, a policy handbook, a set of support articles - rather than from whatever the model already knows. It keeps answers grounded in your material and lets you update what the AI knows by changing documents instead of retraining anything. The trade-off is that it is a standing resource you configure and feed, not a single block you drop into a flow, so it is worth setting up when an agent needs real, current domain knowledge, and overkill when a short fixed instruction in the agent's prompt would do.

Creating a Knowledge Base

You create a Knowledge Base from the Knowledge Bases group in the workspace navigation - hover it and click the +.

The workspace navigation under Agent tools & Knowledge, hovering the Knowledge Bases group: a plus icon for creating a new Knowledge Base appears at the right, with MCP Servers below.

The setup screen

The setup screen takes a few things:

  • Title and Description - a name and summary for the Knowledge Base.
  • Embedding Model - the model that turns your text into vectors, and an AI API Key for it to run under.
  • Vector Store - the store that holds those vectors.
  • Chunk Size and Chunk Overlap - how each document is sliced into chunks.

Together these decide how your text is stored and searched.

In Memory is for testing only

The In Memory vector store does not persist - anything you add to it is held for a short window and then dropped. Use it only while you are trying things out; for anything you need to keep, pick a vector store that persists.

The Create AI Agents Knowledge Base dialog: a Title of Product Docs, a Description, an Embedding Model set to Text Embedding 3 Small, an AI API Key entered, a Vector Store set to In Memory (flagged as non-persistent), and Chunk Size and Chunk Overlap fields with explanatory help text.

Giving a Knowledge Base to an agent

Creating a Knowledge Base does nothing on its own - an agent has to be pointed at it. You do that on the AI Agent block: open Manage Capabilities, go to Knowledge, and add the Knowledge Base from the list of the workspace's knowledge bases. From then on the agent can search it and answer from its documents.

The Manage AI Agent Capabilities window with the Knowledge category selected. The right panel has two groups: Document Management (the Add Document, Delete Document, Delete by Filter, and List Documents action tools) and Search, which lists the workspace's knowledge bases - here "Test KB" - each with a plus button to add it to the agent.

Keeping a Knowledge Base up to date

A Knowledge Base is read at question time, but you also have to get documents into it and keep them current. A flow can do that with the Knowledge Base action blocks:

Because a flow manages the documents, the Knowledge Base can stay in sync with wherever your content lives. A flow that runs when a file lands in a Git repo, a row is added to a Google Sheet, or a record is created in Airtable can add that content on its own - and remove a document just as easily when the source goes away.

You can also hand those same action blocks to an AI Agent as tools. Then the agent decides for itself when to add or delete a document - saving something it was told to remember, or clearing an entry that is out of date - instead of you wiring every change into a flow.

Things to watch for

  • The embedding model, chunk size, chunk overlap, and vector store are set when you create the Knowledge Base and lock once it holds any documents. You cannot change how existing content was sliced or embedded after the fact, so choose these with care up front - to switch, you create a fresh Knowledge Base and add the documents again. The title, description, and AI API Key stay editable, so you can rename it or rotate the key at any time.
  • Adding a document is not instant. A new document shows as Processing while it is being chunked and embedded, and only becomes searchable once it reaches Completed. A query run right after you add will not find content that is still processing.