Use in Chat

After the retrieval test passes, you can select one or more knowledge bases in a regular conversation and let the model answer based on the retrieved snippets while showing the sources.

The conversation is responsible for organizing the answer, and the knowledge base is responsible for providing evidence. First confirm the correct snippets in [Retrieval Test], then decide whether the prompt or chat model needs adjustment.

Prerequisites for use

Check item	Qualified status
Chat model	Supports tool calls
Knowledge base materials	At least one material is ready
Current message	No attachments added at the same time
Retrieval quality	Key questions can find the correct source and complete snippet

When the current message has attachments, knowledge base selection will be disabled. Remove the attachments first, then select the knowledge base from the input area.

Complete a Q&A session with sources

1

1. Select a model that supports tool calls

Create or open a regular conversation and confirm in the model selector that the current model supports tool calls. If the knowledge base entry says the capability is insufficient, switch models first.

2

2. Open knowledge base selection

Click the add button at the lower left of the input area, choose [Knowledge Base], then check one or more target libraries.

3

3. Confirm the selection status

The knowledge base name should appear in the input area. When the question involves only one topic, prefer selecting only one library to reduce competition from irrelevant snippets.

Before sending, confirm that the selected knowledge base and the current question belong to the same scope of materials.

4

4. State the task, scope, and format clearly

For example:Answer only based on the selected knowledge base: what is the accommodation cap for first-tier cities in China? If standards differ by rank, list them item by item and mark the source after each item.

5

5. Open source verification

Check the source name, snippet content, and applicable conditions. Anything not stated in the materials should not be filled in as fact.

6

6. If it fails, go back to the retrieval test

Use the same question to check the snippets returned by the knowledge base. If retrieval is wrong, fix the materials, parsing, or search first; if retrieval is correct, then adjust the prompt and chat model.

When the conversation answer is unsatisfactory, the retrieval results can help determine whether the problem is in the retrieval layer or the answer layer.

How the answer is formed

The chat model sees the final retrieved snippets, not the entire knowledge base automatically.

Recommended question templates

Ask about a clear rule

Answer only based on the selected knowledge base: what is the accommodation cap for first-tier cities in China? If standards differ by rank, please list them item by item, and mark the source after each item.

Compare multiple materials

Compare the approval differences between domestic business trips and overseas business trips. Put them in a table by 'trigger conditions, approver, pre-departure materials'; mark anything not specified in the materials as 'Not stated'.

Require distinguishing facts from suggestions

First list the facts supported by the original text of the policy, then give operational suggestions separately. Suggestions must not be written as policy requirements, and each fact should retain the source name.

A good question contains four things at the same time: the task to be completed, the permissible scope of materials, the expected output format, and what to do when the materials are missing.

Choose one knowledge base or multiple

Situation	Recommendation	Reason
Single policy or product issue	Select only one library	Reduce competition from irrelevant snippets
Cross-department or cross-product comparison	Select multiple libraries and explain each one's purpose	Help the model keep source boundaries clear
Mixed results from multiple libraries	Split into multiple questions and verify separately	First confirm that each library can retrieve independently
Need long-term multi-step research	Switch to an Agent bound to a knowledge base	More suitable for continuous search, organization, and file delivery

Configuration notes

Configuration item	Recommended starting point	Effect	Notes
Number of knowledge bases	1	Control the scope of materials	Increase only when there is a real cross-library need
Question scope	Explicitly write 'only based on the knowledge base'	Reduce common-sense filling-in	Important conclusions still need source verification
Output format	Table or itemized list	Convenient for item-by-item acceptance	Ask for 'Not stated' instead of guessing
Regression issue	Use the same question as in the retrieval test	Distinguish retrieval problems from answer problems	Change only one variable per round

Consolidate conversation content back into the knowledge base

Cherry Studio can save messages, topics, or notes to the knowledge base. Before saving, delete model guesses, duplicate content, and temporary discussion, and use a title that clearly indicates the topic and version.

After saving, a new material snapshot is created and will not sync in real time with the original conversation or notes. When the content changes, it needs to be saved again or replaced.

User case

Xiaolin asked about the accommodation cap in [Employee Business Travel Policy]. The first answer mixed in model common sense, so he changed the prompt to 'write Not stated when it is not specified in the materials' and required the source to be retained for each item. He then opened the citations one by one to verify the city tier, rank, and amount.

The acceptance criterion is: every amount can be directly supported by the cited snippet, and exceptions not written in the policy will not be filled in by the model on its own.

Common questions

Why is the knowledge base entry grayed out?

First select a model that supports tool calls and remove the current message attachments; then confirm that there is at least one knowledge base with ready materials.

Why is there no source in the answer?

Confirm that the input area still shows the selected knowledge base, then put the same question into the retrieval test. If there is no correct retrieval, fix the knowledge base first.

What should I do if the source is correct but the conclusion is inaccurate?

Require the model to answer only based on the citations, break the task into smaller factual items, and manually verify important conclusions. This is usually a prompt, model capability, or context organization issue.

Continue reading

Check materials and recall

First confirm that the correct snippets can be hit consistently.

Use with Agent

Let the knowledge base participate in multi-step tasks and file delivery.

Knowledge base application cases

Reuse policies, after-sales, and research cases.