Segments and Semantics¶

Whether a knowledge base can answer accurately ultimately depends on two specific things: whether unstructured content has been divided into suitable retrieval units, and whether the business definitions for structured data have been clearly described. The former is intelligent segmentation, corresponding to the unstructured path in Data Content Management; the latter is semantic configuration, corresponding to the structured path. Both serve the same goal: enabling the system to truly understand your data rather than merely storing it. This is also why answers may remain inaccurate after the processing status shows “Completed”: segmentation and semantic configuration themselves require maintenance and are not one-time tasks completed upon ingestion.

This page has two parts:

Intelligent segmentation: Retrieval units produced by parsing unstructured files (text, images, tables, transcripts, and so on).
Semantic configuration: Ten types of semantic entries for structured tables, used to improve NL2SQL and data analysis comprehension.

Corresponding UI: the segment area in document details; Advanced Configuration · Semantic Configuration tab.

Intelligent Segmentation¶

A segment is the basic unit of retrieval and citation. After the system parses a file into segments, chats and agents retrieve relevant segments to construct answers.

Segment Quality and Answer Accuracy¶

When an answer is inaccurate, the problem may lie in segmentation rather than only in agent configuration. For example:

A complete rule is split too finely, causing conditions to be omitted
Unrelated paragraphs are merged, mixing retrieval results
Titles are separated from body text, removing context
Image OCR / descriptions are inaccurate
Table structures are recognized incorrectly

Segment Types¶

Common types in the UI:

Type	Description
Text	Body text excerpt
Image	Searchable fields including OCR and image descriptions
Table	Table structure / cell-related excerpt
Image and text	Mixed image and text content
Transcript	Transcribed audio or video content

Common Operations¶

Operation	Description
Search / filter / sort	Search by keyword; sort by original order or retrieval priority
Create segment	Add content that parsing missed but that needs to be retrievable
Edit segment	Modify body text, OCR, image descriptions, and so on
Delete segment	Remove content that should not be cited (this creates a new segment version; the historical version still shows the content from before deletion)
Enable / disable segment	Temporarily exclude a segment from retrieval
Generate initial segments	Generate initial segments when none exist
Re-embed	Rebuild the vector index after segment changes (subject to the UI)

Segment cards generally show the word count and retrieval count, helping identify frequently retrieved segments and segments that have never been retrieved over a long period.

Note Segments can be edited only in the currently active version. Segment changes are recorded in the version history described in Data Content Management. After making changes, validate them with real questions in Chat and Retrieval.

Key Points for Maintaining Image and Table Segments¶

Images: Whether OCR is accurate; whether the image description clearly explains the subject and context; whether irrelevant decorative images should be deleted. Tables: Whether headers are complete; whether row, column, and cell numbers / dates / statuses are correct; whether context is preserved together with the table.

Semantic Configuration¶

Semantic configuration captures field meanings, metric definitions, table relationships, and query experience known by business users as maintainable entries, helping the system perform data Q&A and NL2SQL more accurately.

Note Semantic configuration is only for structured data tables and is not used to retrieve content from ordinary documents, images, audio, or video. Configure data tables (or files) in the data sources first, and then add semantic entries.

Ten Types of Semantic Entries¶

Data tables contain only table names, field names, and data records; by themselves, they cannot express business concepts such as “sales revenue,” “valid orders,” or “this month.” These meanings, definitions, and query practices usually exist only in the minds of business users. Semantic configuration turns that experience into maintainable entries that help the system understand the data more accurately. The ten types below match the “Semantic Configuration” types in the product UI:

Type	Purpose
Dimension column	A field that can be used for categorization and filtering
Fact column	A numeric field that can be used for statistics and calculations
Business metric	A reusable statistical definition (including aggregate expressions and so on)
Table relationship	A JOIN relationship between multiple tables
Column preference	The preferred column for a business concept, or a column that should be deprecated
Named filter	A frequently used filter condition captured as a reusable rule
Standard Q&A	A frequently asked question + verified SQL
Term explanation	A definition and interpretation of a business term
Rule injection	A business rule injected during planning, SQL generation, execution, rendering, or another stage
SQL result set	Governed read-only SQL encapsulated as a reusable query capability

The list supports searching by Key and creating / editing / deleting entries. It can display associated tables, summaries, update times, and other information.

Configuration Recommendations¶

Scenario	Use first
Unclear metric definition	Business metric, term explanation
Multi-table analysis	Table relationship
Frequently used filter conditions	Named filter
The same type of question repeatedly fails	Standard Q&A
Ambiguous field / deprecated column	Column preference, dimension column / fact column descriptions
Need to reuse a stable query	SQL result set, rule injection

Use business language wherever possible rather than writing only technical field names. If the same concept has multiple names, document them together. Metric expressions should be executable (such as SUM(amount)) rather than only natural language such as “total sales revenue.”

Import, Export, and Validation¶

The semantic configuration toolbar supports:

Operation	Description
Import	Upload or paste JSON; this replaces all existing semantic entries in the knowledge base and cannot be undone
Export	Export the current semantic configuration
Validate	Check semantic configuration completeness and return whether it passes and the number of issues