Creating a Knowledge Base

Use this article to create a knowledge base: name it, choose its embedding model, and set how its content is processed, chunked and retrieved. It also covers which of those settings you can change later. For what a knowledge base is and how an AI agent uses one, see Knowledge Base. To fill it with content, see Adding Documents and Adding a Website.

The steps below build a knowledge base called ProductManuals that holds installation and troubleshooting guides for a range of appliances.

The Knowledge Bases list

Open a project and choose Knowledge Bases from the project menu.

1. Tenant Knowledge Base Storage: How much of the tenant's knowledge base storage is in use, shown as "N% storage is used". It counts the document files uploaded to the Draft version of every project in the tenant. Websites do not count.

2. Search name: Filters the list by knowledge base name.

3. + New Knowledge Base: Opens the creation wizard. It is shown on the project's Draft version only.

4. Status: An icon showing whether the knowledge base is ready to use.

Creating the knowledge base

The wizard has five steps. Next stays disabled until the current step is valid, and Save on the last step creates the knowledge base. If you click Cancel after filling anything in, the wizard asks you to confirm before it discards your changes.

Step 1: General Settings

1. Name: Required. Letters, digits and underscores only. Spaces and hyphens are not allowed, so ProductManuals is accepted and Product Manuals is not. It must be unique among the project's knowledge bases.

2. Description: Required. For this example:

Installation and troubleshooting guides for our appliances.

3. LLM Configuration: Required. The LLM configuration that provides the embedding model.

4. Model: Required. It stays empty until you choose an LLM Configuration, then lists that configuration's embedding models (for example text-embedding-3-small). You can also choose a custom model and enter its name.

5. Token Per Minute: Required. The most tokens sent to the embedding model per minute. Default: 30000 (range 10,000 to 50,000,000). The real ceiling depends on your model and provider, so keep the value within their limits.

6. Embedding Dimension: Required. A higher dimension gives better semantic understanding but needs more compute and storage. If you choose a custom model, this field is hidden and the model's own dimension is used.

The embedding settings are fixed once the knowledge base exists. After creation you cannot change LLM Configuration, Model, Token Per Minute or Embedding Dimension. To use a different embedding, create a new knowledge base.

Step 2: Document Processing

This step is optional. It sets an LLM that reads visually complex documents: images, tables, charts and intricate layouts. The Advanced Documents and Data Table source types only become available once it is configured. Without it, both appear greyed out when you add a data source.

1. LLM Configuration: The LLM configuration that provides the document processing model.

2. Model: Only vision-capable models are offered.

3. Token Per Minute: Default: 30000.

4. Instructions: Prefilled with a prompt that converts each PDF page to markdown. It applies only to Advanced Documents. Edit it if you need pages converted differently. The expand icon opens a larger editor.

You can skip this step now and set it later on the knowledge base's detail page. Once the knowledge base holds Advanced Documents or Data Table sources, though, you cannot remove the Document Processing LLM.

Step 3: Chunking Strategy

1. Chunking Method: Recursive Chunking, the only option. It splits content into overlapping sections and keeps sentence and paragraph boundaries intact.

2. Chunk Size: Tokens per chunk. Default: 500 (range 128 to 4096).

3. Chunk Overlap: Tokens shared by consecutive chunks. Default: 50 (range 0 up to the Chunk Size).

Step 4: Retrieval Strategy

This step is optional. A reranking model reorders the retrieved chunks before they reach the model that writes the answer. Without one, chunks are passed on in their initial order. For when reranking helps, see Reranking.

1. LLM Configuration: Only Azure OpenAI and Sestek LLM configurations can be selected.

2. Model: The reranking model.

Every knowledge base uses Hybrid Search (semantic and keyword matching together). It is always on and has no setting. See Hybrid Search.

Step 5: Tags

This step is optional. Use + New Tag to define tags that let a flow narrow retrieval to one brand, product or region. You can also add tags later. How to define them, tag content and filter on them is covered in Knowledge Base Tags.

Click Save. ProductManuals now appears in the list with the status Ready, because it has no data sources yet. Next, add content to it: Adding Documents or Adding a Website. Adding Documents also explains what makes a knowledge base Not ready, and what that blocks.

Changing settings later

Click a knowledge base in the list to open its detail page. The page opens on the Data Source section. The other sections hold the same settings as the wizard.

1. Sections: General Settings, Document Processing, Chunking Strategy, Retrieval Strategy, Tags and Data Source.

2. Embedding fields: Greyed out, because the embedding cannot change after creation.

3. Save: Saves the section you are editing.

You can edit Name, Description, Document Processing, Chunking Strategy, Retrieval Strategy and Tags. If you rename a knowledge base, the new name is updated in the flows that use it.

After you change Document Processing or Chunking Strategy, each data source shows an info icon next to its Indexed status. Sources keep working on the old settings until you run Refresh Index in the Data Source section, which re-indexes them with the new ones.

While any data source is Pending or Processing, the settings sections are read-only. Tags stay editable.

Limits and naming rules