Knowledge Base Overview

← User Guide · 简体中文

A knowledge base (RAG) gives the AI access to your own material: imported documents are split into segments and indexed; when you ask, the system retrieves the most relevant segments and injects them into the model, grounding answers with traceable citations.

Good use cases: policy Q&A, product-manual support, personal notes search, domain assistants.

Overall Flow

Create knowledge base → Import documents (pick segmentation mode) → Index (vectorize / graph) → Q&A or link to a character

Step	What	See
1. Create	Retrieval and segmentation settings	Create & Configure
2. Import	Form entry or file upload, choose segmentation mode	Import Documents
3. Index	Vectorization (required) and graph extraction (optional)	Import Documents
4. Maintain	Segment enable/disable, QA maintenance, retries	Segment Management, Q&A Import & Generation
5. Use	Knowledge base Q&A, or link the base to a character	Knowledge Base Q&A, Character Settings

Knowledge base Q&A page

My knowledge bases page

Two Entry Points

• Knowledge base Q&A: the Knowledge Base menu item — pick a base and ask;
• Knowledge base management: the Knowledge base management button at the bottom of the Q&A sidebar (or visit the "My knowledge bases" page directly) — creating and maintaining.

Three Segmentation Modes

The segmentation mode decides how documents are segmented and retrieved. It is a document-level property — chosen when importing or editing a document; one knowledge base can mix modes.

Mode	Best for	Retrieval
Plain segmentation	Continuous text: manuals, articles	Auto-segmented per the base's segmentation strategy; semantic similarity recall
Q&A	Paired content: FAQs, support scripts	Question-answer pairs; hitting any question recalls the answer
Parent-child	Long, hierarchical documents	Small chunks (children) match precisely; the hit returns its parent (larger context) to the model

How to choose:

• Material is naturally question-answer → Q&A (highest accuracy, but requires reformatting);
• Continuous prose with clear paragraphs → plain (easiest);
• Long documents needing both precise hits and full context → parent-child (e.g. a manual: child hits a spec item, parent returns the section).

NOTE

Q&A mode only imports XLSX / XLS / CSV files; parent-child mode has its own child chunk max-length setting (default 200).

Segmentation Settings vs Segment Mode

Segmentation parameters and the segment mode live on two orthogonal levels; the pipeline is "cut segments first with the base-level segmentation settings, then organize the segments per the document-level mode":

1. Segmentation settings (knowledge-base-level) — the Document Segmentation Settings configured on the base (segmentation strategy, max segment length, segment overlap length, custom separator, token estimator) decide where the cut points are; they apply to every document in the base;
2. Segment mode (document-level) — each document chooses how to organize the segments:
   • Plain segmentation: each segment is vectorized and retrieved directly;
   • Parent-child: the segment serves as the parent; smaller child chunks (at the document's child chunk size) are split inside the parent, only children are vectorized, and a hit returns its parent;
   • Q&A: pairs are generated from the body text by AI (questions vectorized, answers returned as-is), with the generation window split at the base-level max segment length. Importing pair files (XLSX/CSV) skips splitting entirely — the only exception.

TIP

Naming convention: the top-level split output is always a segment (plain segment / parent / answer segment); only parent-child mode splits further into child chunks — which is why the child chunk max-length setting exists only on parent-child documents. The "Segmentation config" block on the document edit page lists the effective chunking params per mode (paragraph / parent / QA segment max length, segment overlap length, chunking strategy or custom separator; parent-child adds the child chunk max length) — all sourced from the knowledge base's segmentation settings (an "inherited from knowledge base settings" note beside the block title; the settings words link there).

Two Index Types

Index	Purpose	Note
Vector index	Embeds segments for semantic similarity search	The foundation; prerequisite for Q&A
Graph index	Extracts entities and relations into a knowledge graph	Optional; extraction uses an LLM and consumes tokens

Both can coexist; answers can be traced via citations and the citation graph (see Knowledge Graph). They complement each other: vectors are best at "semantically similar" content, graphs at "multi-hop relations" (who is the head of A's department).

Strict vs Lenient Mode

A knowledge-base-level answering strategy (set at creation, see Create & Configure):

• Strict: with no relevant hits, returns "[No answer]" instead of making things up — for policies and compliance;
• Lenient: with no hits, the question goes to the LLM for a free-form answer — for assistant scenarios.

Public vs Private

• Private: visible and usable only by the creator;
• Public: every user can see and use it (read-only) under the "Public" tab.

Where to Go Next

Goal	Page
Build a base from scratch	Create & Configure
Import Word / PDF	Import Documents
Already have an FAQ sheet	Q&A Import & Generation
Keep one segment out of retrieval	Segment Management
Try graph retrieval	Knowledge Graph
Ask the base directly	Knowledge Base Q&A

---

Previous: Gallery · Next: Create & Configure