Agent conversations

SAG Agent organizes retrieval, tool calls, and streamed generation into a traceable conversation path. An Agent can bind all sources by default, while @ can restrict the knowledge scope for an individual question.

How an answer is produced

user message
  -> resolve selected sources
  -> retrieve evidence
  -> stream model response
  -> attach citations
  -> open citation back to chunk

Citations are not detached summaries. Each one points to the original chunk behind a retrieval result. Open a citation to verify the title, source, and source text.

Select the knowledge scope

Enter @ in the composer to open source multi-select. An explicit scope reduces irrelevant candidates, lowers retrieval cost, and makes the answer's knowledge boundary clearer.

The default Agent binds all sources, which suits a unified personal knowledge base. Specialized Agents can save fixed bindings such as "Product support," "Research papers," or "Compliance handbook."

Attachments and conversation uploads

The composer supports pasted images and document attachments. Documents automatically enter the "Conversation uploads" source and participate in subsequent retrieval. Images can be sent as visual input to models that support vision.

Historical images are not repeatedly injected as complete base64 payloads, which prevents unbounded context growth. The composer estimates context usage so that you can adjust the question scope to the model window.

Streaming state and session actions

SSE streams both answer tokens and tool activity. A running state is restored in the conversation after navigating away. Messages support copy, retry, delete, and timestamp actions. Conversations can be archived; archive is not permanent deletion.

Use the Agent from another client

The same Agent is available through the OpenAI-compatible endpoint. Knowledge scope, streaming behavior, and citation data match the built-in conversation experience. MCP can expose a read-only knowledge-exploration surface instead.

• Model-client integration: OpenAI-compatible API
• Agent-tool integration: MCP and Agent Skill