AI Fusion 2.x Release Notes

Release notes of the aifusion extension, versions 2.x. Documentation for this version is in the legacy section.

For compatibility with Fabric and with the other AI Fusion extensions, see AI Fusion Versions.

Version 2.0.0

Released: February 2026 · Minimum Fabric version: 8.4.0

Re-designed backend for LLM invocation and embedding, for better maintainability and support for new features. Custom LLM and embedding interfaces can be implemented by extending the BaseIoLLMInvoke and BaseIoLLMEmbed classes, with the LLMResult, EmbeddingResult, UsageStats, ToolDefinition and LLMError records.
Reworked streaming: responses are written immediately to the output stream provided by the web service.
New Chat actor that wraps customer chat flows and manages the chat session context and state: Fabric GET of the session, STEP table updates, session monitoring and the output stream.
New ChatSessionParams actor that exposes the chat session parameters (session ID, customer ID, app ID, user query and custom parameters).
New WriteToOutputStream actor that writes JSON to the output stream at any point of a flow.
LLMInvoke and LLMAgent — new outputs: content (final text response), duration (ms), usage (input, output, cache read, cache write and total tokens), messages (full conversation) and tool_calls (flows or MCP tools requested by the LLM).
Embed — input accepts a string or an array of strings (batched when the interface supports it); new dimensionsOverride input; new embeddings output with the vectors, their string representation, the original text and dimensions.
Actor descriptions and remarks were updated.

The LLM actors no longer return a stream. The final result is returned in content. The stream output is kept for backward compatibility, but it is removed when a flow using LLMInvoke or LLMAgent is opened — reconnect its links to content.
Embed: flows that used the single embedding string must use the embeddingString key of the first element of embeddings, or iterate over the list.
The PrependParamsToStream actor is no longer needed and no longer works as expected with the new streaming.

Release notes of the aifusion extension, versions 2.x. Documentation for this version is in the legacy section.

For compatibility with Fabric and with the other AI Fusion extensions, see AI Fusion Versions.

Released: February 2026 · Minimum Fabric version: 8.4.0

Re-designed backend for LLM invocation and embedding, for better maintainability and support for new features. Custom LLM and embedding interfaces can be implemented by extending the BaseIoLLMInvoke and BaseIoLLMEmbed classes, with the LLMResult, EmbeddingResult, UsageStats, ToolDefinition and LLMError records.
Reworked streaming: responses are written immediately to the output stream provided by the web service.
New Chat actor that wraps customer chat flows and manages the chat session context and state: Fabric GET of the session, STEP table updates, session monitoring and the output stream.
New ChatSessionParams actor that exposes the chat session parameters (session ID, customer ID, app ID, user query and custom parameters).
New WriteToOutputStream actor that writes JSON to the output stream at any point of a flow.
LLMInvoke and LLMAgent — new outputs: content (final text response), duration (ms), usage (input, output, cache read, cache write and total tokens), messages (full conversation) and tool_calls (flows or MCP tools requested by the LLM).
Embed — input accepts a string or an array of strings (batched when the interface supports it); new dimensionsOverride input; new embeddings output with the vectors, their string representation, the original text and dimensions.
Actor descriptions and remarks were updated.

The LLM actors no longer return a stream. The final result is returned in content. The stream output is kept for backward compatibility, but it is removed when a flow using LLMInvoke or LLMAgent is opened — reconnect its links to content.
Embed: flows that used the single embedding string must use the embeddingString key of the first element of embeddings, or iterate over the list.
The PrependParamsToStream actor is no longer needed and no longer works as expected with the new streaming.