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

Release 2026.5

Check what's coming in May.

Builder Updates​

Semantic Search in Builder​

Search for components in Builder using natural language queries. This is a powerful feature that helps you find the right components for your pipeline quickly and easily. To find a component:

  1. In Builder, open the Component Library.
  2. Type your query in the search bar using natural language. For example, "generate text from documents" or "rank search results by relevance". The library will suggest components that match your query.

Improved File Preview in Playground​

Uploaded file previews in Playground now work correctly for all supported file types. Image files (PNG, JPG, JPEG, GIF, WebP, SVG) now display inline as images instead of being treated as binary content. Other file types also benefit from improved MIME type detection, so PDFs, text files, and other formats render reliably after upload.

For details on supported preview formats, see Upload Files.

Organizing Components with Groups​

Group components to organize them and make your pipelines more readable. Groups help you keep related components together and you can collapse them to reduce visual clutter.

To create a group, click the Groups button (group icon) in the bottom right corner of the canvas. Then click and drag on the canvas to draw a rectangle around the components you want to group.

Keyboard shortcut: Press Cmd+G (Mac) or Ctrl+G (Windows/Linux) to activate group drawing mode, then draw the rectangle. Press Esc to cancel drawing mode.

Groups are saved with your pipeline and restored when you reload the builder.

The Builder canvas with a group rectangle drawn around the components

Download Answer from Playground​

You can now download the answer text from a search result directly in Playground. Click the download icon next to an answer to save it as a Markdown file. The download icon appears alongside the existing copy icon and is available whenever an answer is displayed.

For details, see Test Your Pipeline in Playground.

MultiRetriever Component​

You can now use the MultiRetriever component in Pipeline Builder to query multiple indexes and document stores in a single pipeline step. Configure each knowledge source with its own document store, index, and retriever strategy (BM25, embedding, or hybrid). Results from all sources are combined into a single list of documents passed to downstream components.

Key capabilities:

  • Add, edit, and remove knowledge sources directly from the component drawer.
  • Mix retriever strategies across sources: use keyword-based BM25 for one index and embedding-based retrieval for another.
  • Select a query embedder (deepset NVIDIA, SentenceTransformers, or FastEmbed) for sources that use embedding or hybrid retrieval.
  • The embedding model syncs automatically from the selected index when compatible.

For details, see MultiRetriever.

Prompt Explorer: Updated Prompt Saves as a New Version​

When you click Update in Prompt Explorer, the updated prompt is now saved as a new pipeline version instead of being deployed automatically. Your current draft and the deployed version remain unchanged. To make the updated prompt live, deploy the new version manually.

For details, see Using Prompt Explorer.

Updated Default Reranker Model for DeepsetNvidiaRanker​

New pipelines that use DeepsetNvidiaRanker now default to the tomaarsen/Qwen3-Reranker-0.6B-seq-cls model. The previous default models (intfloat/simlm-msmarco-reranker, BAAI/bge-reranker-v2-m3, and svalabs/cross-electra-ms-marco-german-uncased) are decommissioned and no longer selectable for new pipelines. Existing pipelines that use one of these models continue to work without any changes, and the model has a legacy label in the model list.

Updated Editor Role​

The Editor role can now manage workspace members, including adding and removing them, and updating their roles. For details, see User Roles and Permissions.

Filter Indexed Files by Metadata​

You can now control which files your index processes based on their metadata. Set a list of conditions on a pipeline index — each specifying a metadata field, an operator (== or !=), and a value — and only files whose metadata satisfies all conditions are indexed.

This gives you a more flexible alternative to the existing filename prefix filter, letting you scope an index to a specific document category, language, source, or any other metadata attribute your files carry.

For details, see Choose Files to Index.

Simpler Pipeline Versions​

Creating a new draft version no longer fails when a draft already exists. If you create a new draft and one already exists, the existing draft is automatically finalized and a new draft is created in its place. You only see a conflict error if another user is actively editing the existing draft. For details, see Manage Pipeline Versions.