Knowledge base basics

A knowledge base organizes files, notes, directories, or web pages into a reusable collection of materials that can be searched repeatedly. First use recall testing to confirm the system can find the correct passages, then hand the knowledge base over to a conversation or Agent.

If you only need to handle a small piece of text temporarily, it is faster to paste it directly into the conversation. Build a knowledge base when the materials will be reused repeatedly and the answer must be based on internal sources.

When is it suitable to use

Need	Recommended approach	Reason
Query employee policies, product manuals, and project materials	Build a knowledge base	The materials will be reused, and you need stable citation of the original text
Temporarily analyze an attachment	Upload directly in the conversation	No long-term maintenance or indexing needed
The materials are still being organized	First organize it with [Notes]	Avoid unverified content being treated as an official answer
Need to automatically process materials over the long term	Bind to an Agent after building the knowledge base	The Agent can continuously use the same material scope during tasks

What happens in a single answer

The materials are first parsed and split into chunks, then candidate passages are found using keywords or semantic search; the chat model is only responsible for forming an answer based on the recalled content.

Get to know these terms first

Name	Meaning in this task
Knowledge base	A set of materials and retrieval settings organized around the same topic
Material entry	An imported file, note, file in a directory, or web page snapshot
Chunk	A small passage used for retrieval after the materials are split
Recall	The process of finding relevant passages based on a question
Embedding model	Converts text into vectors to match content with different wording but similar meaning; optional
Reranking model	Scores and reorders candidate passages again; also optional

A knowledge base can still be used without an embedding model, in which case it mainly relies on BM25 keyword retrieval. For your first try, you can choose [Do not use] and run through the creation, import, and recall workflow first.

Complete your first use in 5 minutes

1

1. Create a knowledge base with a clear scope

Open the left navigation [Knowledge Base] → click the Add button above the knowledge base list. Use the format “object + purpose” for the name, for example [Employee Travel Policy].

2

2. Choose a retrieval method

For your first try, you can set [Embedding Model] to [Do not use]. Configure an embedding model later when you need to match conversational wording or synonyms.

3

3. Add materials

After entering the knowledge base, click the Add Material button and choose [File], [Note], [Directory], or [Link].

Choose the entry point according to the source of the materials; do not import irrelevant directories just to reduce the number of steps.

4

4. Wait for the materials to become ready

After processing is complete, the materials will appear in the list. Spot-check the main text and Chunks to make sure there is no garbled text, missing pages, or obvious misordering.

After the materials are ready, you still need to spot-check the content; import completion does not mean the retrieval quality is already qualified.

5

5. Complete the recall test

Open [Recall Test], enter a real question whose answer you already know, and check whether the correct source appears among the first few results.

6

6. Enter chat or bind an Agent

Once recall is stable, select the knowledge base in the chat input area; when you need a long-term workflow, bind the knowledge base on the Agent editing page.

Recommended starting point

Configuration item	Product default value	Recommended starting point	Function	Applicable scenarios	Notes
Knowledge base scope	—	A clear topic	Controls which materials participate in retrieval together	Policies, products, project materials	Content with different permissions or lifecycles should be separated
Embedding model	Do not use	Do not use for now	Determines whether vector retrieval is added	Conversational wording differs greatly from the original text	Billing and data handling for cloud models depend on the provider
Test questions	—	3 to 5 real questions	Establish a long-term regression baseline	After each materials or settings update	Do not test using only the material title and original sentences
Enter formal use	—	Materials, Chunks, and recall all pass	Avoid bringing parsing or retrieval problems into the conversation	All knowledge bases	The chat model cannot make up key information that was not recalled

How to tell when it is ready to use

• All required materials show as available, with none stuck in processing or error status for a long time.
• Randomly open one or two materials; the main text and Chunks should have no garbled text, missing pages, or obvious misordering.
• Use fixed questions for recall testing, and the correct source should consistently appear among the first few results.

User case

Xiaolin wants colleagues to be able to query the travel policy. He creates an [Employee Travel Policy] knowledge base, imports three materials covering lodging, transportation, and approvals, and leaves the embedding model unconfigured for now. After the three materials are ready, he tests with questions like “What is the lodging limit in Beijing?” and “Who approves amounts over 5,000 yuan?”

The completion standard is: for each question, the correct policy source can be found, and the passage includes both the applicable conditions and the conclusion. Only after meeting this standard does he bind the knowledge base to the Agent responsible for employee Q&A.

Common questions

What is the difference between a knowledge base and a chat model?

The knowledge base is responsible for finding passages from your materials, while the chat model is responsible for understanding the question and forming the answer. If key information is not found during recall, simply switching the chat model usually will not solve the problem.

Can it be used directly after materials are imported successfully?

You still need to spot-check the main text and Chunks, and complete a recall test. Successful import only means the processing flow has finished; it does not mean the results are complete or correctly ranked.

Will old materials change automatically after modifying the chunking settings?

No. To make existing materials use the new chunking settings, you need to perform [Reindex], then retest with the same set of questions.

Continue reading

Create a knowledge base

Determine the name boundary and retrieval method.

Add and organize materials

Import files, notes, directories, and web pages.

Check materials and recall

Verify retrieval quality with fixed questions.

Model and retrieval settings

Continue adjusting embeddings, reranking, and chunking.