Skip to main content
For the complete documentation index for agents and LLMs, see llms.txt.

Release 2026.10

Check October's updates.

Ship It Through Gateway​

You can now publish pipelines through Gateway, giving each one a stable endpoint that stays the same even as you keep improving the pipeline behind it. Create a service for a pipeline, publish new versions whenever you're ready, and your callers stay connected throughout. Learn more in Gateway.

Skills for Your Agents​

You can now give your Agents skills. Skills are versioned folders of files, such as instructions and reference material, that an Agent reads while it runs to work through a task the way you want.

On the new Skills page under Resources, you can browse skills made by deepset, create your own skills, and scope them to a workspace or your whole organization. You can also connect your organization's GitHub or GitLab repository so its skills appear in the library. Once a skill is available, you can attach it to an Agent from the Agent's settings.

For details, see Agent Skills and Manage Agent Skills.

File Metadata Improvements on the Files Page​

Manage file metadata more easily right from the Files page.

  • Download Metadata: Click More actions next to a file and choose Download Metadata to save its metadata as a <filename>.metadata.json file.
  • Upload with metadata: When uploading files from Files > Upload Files, expand Add Metadata (optional) and type a JSON object to attach metadata to all files in the batch.

For details, see Add Metadata to Your Files.

Claude 5.5 Haiku in Your Pipelines​

You can now use Claude 5.5 Haiku in your pipelines by choosing it from the model list on the LLM component, through either Anthropic or Amazon Bedrock. For setup and thinking settings, see Use Anthropic Models and Use Amazon Bedrock Models.

See How Claude Thinks​

You can now choose whether Claude 5 models include a summary of their extended thinking in responses, using the settings on the Advanced tab of the LLM component. For details, see Use Anthropic Models.

New Claude, OpenAI, and Gemini Models​

You can now use the latest OpenAI, Anthropic, and Google models in your pipelines, including the GPT-6 and GPT-5.6 families, and your existing pipelines keep working without changes. For the available OpenAI models, see Use OpenAI Models.

Advanced Filter Mode in Playground​

You can now test complex filters in Playground, including nested AND, OR, and NOT conditions. Switch the filter panel to Advanced and type or paste a raw JSON filter. We validate it before applying it. Simple mode keeps the field list you already know. For details, see Filter Syntax.

Delete Indexes That Pipelines Still Use​

You can now delete an index at any time, even if pipelines still use it. Before you confirm, the dialog lists the deployed and undeployed pipelines that the deletion affects. Deleting an index is permanent, and those pipelines lose access to its data. In the API, the delete index endpoint has a new optional force parameter that lets you delete an index that pipelines still use. For details, see Delete an Index and Indexes.