Default models

Set the company default chat model, the default small model, the default embeddings model, the image reading model and the image generation model, and what each one is used for.

For: Administrators (the admin ability); everyone can see the current defaults · Last updated

Below the provider table, the Model providers page has a panel for each company-wide model choice. Each holds one choice for the whole company.

Each change to a default model is recorded in the audit log, with who made it and the previous and new model.

Company default model#

The Company default model pre-selects the model step when someone hires a new Coworker. It also backs the platform's own background work that runs outside any single Coworker's model choice: job description parsing, summarization and the weekly feedback synthesis.

Default small model#

The Default small model is used for tasks that need little intelligence, such as classification, summarization and ranking the results of the Knowledge Base search box. If it is not set, the company default model is used. Any chat-capable model from an active provider is accepted, whatever its context window.

The same model screens Coworker actions that change something in another system, when Review risky writes is on in Features. One quick check runs before each such action set to Review (a connector function, a custom connector or a Coworker action), and is recorded under the job write-reviewer in usage and cost. The check sees the function, the details of the action with secret values removed, and the person's latest message. If no model is available or the check fails, the person is asked to approve instead. See Reviewing and editing connectors.

Relevance ranking in the Knowledge Base search box runs on this model. Coworkers and MCP tools do not use it: they look things up in Knowledge Bases directly, with no ranking model. Each search box search makes one small model call, which appears as Search rerank in usage and cost. If neither the default small model nor the company default model is set, relevance ranking is off and results appear in plain similarity order. You can also turn it off for the whole company with the Search box relevance ranking switch on the Features page (see Features).

Company default embeddings model#

The Company default embeddings model backs Knowledge Base embedding. Only models that the provider reports as embedding-capable are offered.

A Knowledge Base records the embedding model it was created with. Changing this default does not change existing Knowledge Bases.

Image reading model#

The Image reading model reads scanned PDF pages, standalone images, and large pictures inside Word, PowerPoint and OpenDocument files when they are added to a Knowledge Base. Text that a file already contains is read directly and never sent to a model. The panel reads "Used to read scanned pages and images when documents are added to a Knowledge Base." Only administrators see this panel.

The panel shows Currently used for image reading with the model that will be used. The model used is the first one available from this list:

  1. The model you choose in this panel.
  2. The Default small model, if it can read images.
  3. The Company default model, if it can read images.

When the model comes from step 2 or 3, a note beside it says the small model or the default model "is used because no image model is chosen". If none of these is available, the panel says "No model is available for reading images." Text pages are then still indexed, and scanned pages and pictures are flagged in the document's notes as not read; a file made only of scanned pages or pictures fails with a message asking for an image-reading model to be chosen.

If the provider behind the model you chose is not active, or no longer lists that model, the fallback order continues with the next step. A document read with a fallback model carries a note, "Read with a fallback model because no image-reading model is chosen." or "Read with a fallback model because the chosen image-reading model is not available." Documents read with the chosen model show Read from page images and no such note. A model you choose in this panel is used even if the product cannot tell whether it reads images. Run the test to be sure. Models chosen in steps 2 and 3 are used only when the product recognizes them as able to read images.

Changing the image reading model affects files read after the change. Documents that are already indexed keep their previous result, including any note about the model that read them, until someone who manages the Knowledge Base chooses Read again on them. That reads the file again with the model in force at that moment, and it is billed like the first read.

Choose the model#

  1. Open the model picker. It lists chat models from active providers. None clears the choice ("Use the company default model").
  2. Choose a model.
  3. Select Set image model, or Clear image model when you chose None. A "Saved" note confirms it.

A model is not checked against a context-window minimum, because each request is a few pages. Before you save a model that has not passed the test, the panel says "Run the test to confirm this model can read images."

Test image reading#

Test image reading sends the model a generated picture split into a red half and a blue half and checks that it reports red on the left and blue on the right. The test is metered like any other model call (see The Usage & cost view).

  • With a model selected in the picker, the test checks that model before you save it. The result is shown but not stored.
  • With nothing selected, it tests the model currently in use, including one chosen by the fallback order.
  • The result appears under the buttons. A pass reads "Passed — the model correctly saw red on the left and blue on the right." A failure reads Failed with the reason: the model answered the wrong colors, its reply held no usable answer, it rejected the image (it may not support image input), it is no longer available from its provider, it did not answer within 60 seconds, or the provider returned an error.
  • When the model you saved passes, the panel shows Verified with the date. When it fails, the panel shows "Last test failed:" with the reason. Changing the saved model clears the record, and a record only ever describes the model it was run against.

If no model is available, selecting Test image reading with nothing chosen reports that no image model is configured to test.

The test confirms that the model accepts images and can tell two colors apart. It does not measure how well the model reads text. Smaller models, and models run locally, can misread text that a larger model reads correctly, so after you choose one, add a real document to a knowledge base and check the result. Pages and pictures are enlarged to a readable size before they are read, so a small image is not sent at its original low resolution.

Image generation model#

The Image generation model creates new images from a text description, such as a Coworker's profile image. It is separate from the Image reading model, which reads images that already exist. Only administrators see this panel.

The picker lists only models that the provider reports as image-generation capable, and each model is shown with "(image)" after its name. Chat and embedding models are not offered. A provider that was added before it reported image-generation models needs Refresh models on its row in the provider table before those models appear; the panel then says "No image-generation models are available. Add a provider that offers one, then use Refresh models." until one does.

  1. Open the model picker. None clears the choice ("not set — image generation is off").
  2. Choose a model.
  3. Select Set default, or Clear default when you chose None. A "Saved" note confirms it.

Each generated image is recorded at the provider's list price when the provider's price table covers that model. A model the table does not cover is recorded as unpriced in usage and cost.

Unlike the other defaults, removing or deactivating the provider behind the image generation model does not need the default to be cleared first. The choice is cleared for you.

Set a default#

You need the admin ability. Other people see the current choice as read-only text, or "No default model set." / "No default embeddings model set." For the small model the panel says "Not set — using the company default model."

  1. In the panel, open the model picker. It lists models from active providers only, each with its provider's name.
  2. Choose a model.
  3. Select Set default. A "Saved" note confirms it.

If there is nothing to choose from, the panel says "Add a provider above before setting a default." (for embeddings: "Add a provider with an embedding-capable model above before setting a default."). Everyone sees this message, not only administrators.

To remove a default, choose None ("clear the default") and select Clear default. You must clear a default (including the small model) before you can deactivate or remove the provider behind it.

Context-window checks#

  • If the product cannot confirm a chat model's context window, a warning reads "We can't confirm this model's context window — it may not support long conversations." You can still set it.
  • If a chat model's context window is too small, the panel shows "This model's context window is too small to support reliable Coworker conversations" and Set default is unavailable. Choose another model.
  • The default small model is not checked against the conversation context-window minimum.
  • Embedding models are not checked against the conversation context-window minimum. If an embedding model's context window cannot be confirmed, the warning reads "We can't confirm this model's context window."