Create & Configure a Knowledge Base

← User Guide · 简体中文

Open Knowledge Base → Knowledge base management — the "My knowledge bases" overview. The table lists title (link to detail), description, public flag, strict flag; the actions column offers Docs / API / Edit / Delete.

• The search box filters by title (Enter or click the search button);
• The table paginates at the bottom ("{n} items in total").

My knowledge bases

Creating a Knowledge Base

1. Click Add to open the form (five settings cards);
2. Fill in each group as described below;
3. Save and return to the list.

TIP

For a first base, only the title is required — start importing with defaults; segmentation and retrieval settings can be changed later (segmentation changes take effect for subsequent indexing only; see below for rebuilding existing documents).

1. Basics

Field	Required	Description
Title	yes	Name of the knowledge base
Description	no	Purpose, visible to others when public
Public	no, default private	Public: visible and usable by everyone; private: creator only
Strict mode	no, default lenient	Strict: returns "no answer" directly when nothing relevant is retrieved, never invents one; Lenient: hands the question to the LLM for a free-form answer when nothing is retrieved. See Overview

2. Document Segmentation Settings

These parameters are knowledge-base-level; they apply to the segmentation of every document in the base (the document edit page shows the effective segmentation config read-only, per the document's segmentation mode).

Field	Description
Segment overlap length	Content shared by adjacent segments (marked "applies to new indexing after change") — keeps boundary information intact; validated on save to be smaller than the max segment length (the effective backend cap is half of it)
Segmentation strategy	Recursive / by paragraph / by line / by sentence / custom separator (required when custom, e.g. \n\n, \n)
Max segment length	Upper bound of a segment, range 100–10000; longer content keeps splitting; Q&A generation also splits the source text at this size
Token estimator	OpenAI / Qwen / Huggingface counting conventions — match your embedding model family

NOTE

Saving a change to these parameters takes effect for subsequent indexing only (newly uploaded documents and later re-splits follow the new parameters) and never affects existing segments; generated Q&A pairs pick up the new parameters only when regenerated. Rebuild Vectors only recomputes vectors and never re-splits segments (see Import Documents · Full Rebuild); to re-split one document with the new parameters, edit that document (change its content or switch its mode). How the segmentation settings combine with the document-level segment mode: Overview · Segmentation Settings vs Segment Mode.

TIP

"By paragraph" suits well-structured documents; "recursive" is the general default; tables and logs suit "by line" or a custom separator.

3. Indexing Model Settings

Model: the model shared by graph extraction and Q&A-pair generation; if empty, the first available model is used.

WARNING

Graph extraction and QA generation both consume this model's tokens. Pick a cost-effective model; leaving it empty falls back to the first available model.

4. Retrieval Settings

Field	Description
Max retrieved	Maximum segments recalled per query — higher is more thorough but costs more
Min score	Segments below this similarity are excluded — higher is stricter and filters out more noise

5. LLM Parameters

Field	Description
System prompt	Persona for knowledge base Q&A (e.g. "You are the company policy assistant; answer only from retrieved content")
Creativity / randomness	Answer temperature: lower = more conservative, higher = more varied

Create form

Editing

Click Edit in the actions column. Note the consequences:

Change	Effect
Retrieval settings, LLM parameters	Immediate, no rebuild
Segmentation parameters (overlap / strategy / max length…)	Takes effect for subsequent indexing only — existing segments are untouched; re-split a single document by editing it (see Full Rebuild)

API

The API action (actions column, or the top-right of the knowledge base detail page) generates this base's API key and endpoint docs — see API Reference · Knowledge Base Q&A.

Deleting

1. Click Delete in the actions column;
2. Confirm "The data cannot be recovered after deletion. Delete knowledge base [{title}]?".

WARNING

Deleting removes documents, segments, vectors and graphs — irrecoverable. Characters linked to the base lose its retrieval.

---

Previous: Overview · Next: Import Documents