Chat and Retrieval

The knowledge base module has two tabs at the top: Knowledge Bases and Chat. The Chat page is used to ask questions within selected knowledge bases, ask multi-turn follow-up questions, and view source citations.

Before looking at specific operations, first understand the role of this entry point—it is a validation layer, not a publishing layer:

	Knowledge base chat	Agent
Purpose	Validate knowledge scope and answer ad hoc questions	An application with configurable skills / tools / publishing
Knowledge scope	Knowledge bases selected for the current conversation	Knowledge bases and capabilities bound to the agent
Permissions and publishing	Lightweight entry point	Complete agent lifecycle

In other words, knowledge base chat is suitable for validating retrieval results and making ad hoc queries. Answer quality depends on the selected knowledge bases, whether data source processing is complete, semantic configuration, and your permissions on the underlying data. When an answer does not meet expectations, the issue is usually with one of these areas rather than with the chat function itself. Do not treat it as the entry point for a production application. To build an application that can be published to end users, use an agent.

Corresponding UI: Knowledge Bases · Chat (multiple sessions + streaming Q&A).

Tabs and Layout

Area	Content
Session management on the left	Create chats, search session titles, pin / rename / delete
Main area	Messages in the current session, knowledge base selection, model selection, input box
Message capabilities	Streaming generation, stop, retry, like / unlike, delete messages, export reports, and more

Starting a Chat

1. Open the Chat tab and create a chat or select an existing session.

2. Select one or more knowledge bases for the current conversation (at least one is required).

3. Select a model (which may include the system default; the list is provided by the platform).

4. Enter and send a question. Multi-turn follow-up questions are supported, and the system uses the conversation history to understand subsequent questions.

5. Review the final answer and references. You can open source details and preview some file-based sources.

Note A session can save the knowledge bases and model configuration used for the current conversation. Switching sessions does not mix their respective histories or configurations.

Managing Multiple Sessions

Operation	Description
Create chat	Start an independent session
Search titles	Find sessions by title
Pin / unpin	Keep frequently used sessions at the top
Rename	Change the session title
Delete	The session history cannot be recovered after deletion

Use separate sessions to isolate topics, for example, keeping “Product documentation Q&A” separate from “Sales data analysis.”

Source Citations and Process Information

• References: Distinguishes document sources, data table sources, and others. File-based sources may include source excerpts, page numbers, relevance scores, and other information, and some can be previewed.

• Reasoning process / execution process (when enabled for viewing): May show tool steps such as retrieved evidence excerpts, SQL execution, and semantic parsing to help investigate why an answer was produced.

• Export: Exports conversation results as a report (such as Word) for sharing.

For business rules, data definitions, or important conclusions, prioritize checking the sources rather than relying only on generated text. When the knowledge base has no relevant content, the answer still requires manual verification.

Image-related Capabilities

The chat input area supports uploading images as part of a question for queries that incorporate visual content. Whether image assets in the knowledge base can be retrieved depends on whether image indexing was enabled at creation time and on the quality of image parsing and segmentation.

When image results are retrieved, you should still return to the original material through the source for precise verification.

Feature Controls

If “Feature controls” are available in the input area, you can enable options that may include:

Control	Description
Deep attribution	Strengthens the displayed association between answers and supporting evidence
Quality evaluation	A quality-related control for the current conversation at runtime, not a knowledge-base-level “Evaluation and optimization” page
Hallucination evaluation	As above, a session-level control

These switches are not Golden Set / automatic tuning product capabilities. The knowledge base module does not have a separate evaluation and optimization configuration page.

Related Documents

• Creation and Data Sources

• Data Content Management (validate only after processing is complete)

• Segments and Semantics

• Agents