Create a Knowledge Base
The two most important choices when creating are the name and the embedding model. Materials can be added later, but the name boundary and retrieval method will affect future maintenance.
Make two decisions before creating
The name should indicate the scope of the materials
Prefer “object + purpose,” such as [Employee Travel Policy], [Product After-Sales Manual], or [Market Research Materials]. Avoid names like [Materials] or [Test], which make it impossible to judge the content scope later.
Choose an embedding method
Do not use
First-time experience, clear keywords
BM25 keyword retrieval
None
Cloud embedding model
The user’s wording differs greatly from the original material
BM25 + vector hybrid retrieval
The corresponding model service can be called normally
Local embedding model
You want vectorization to be completed on this machine
BM25 + local vector retrieval
First download it in [Local Models]
Creation steps
Use a local embedding model
Open [Settings] → [Local Models], and download an available model in the [Embedding Models] section. The models and download sizes shown in the interface may vary depending on the installation environment; please refer to the current list.

Changing the model after materials already exist
When enabling an embedding model for a knowledge base that previously used only BM25, a vector index can be built. When changing the embedding model for a knowledge base that already has vectors, the interface will enter the [Rebuild Knowledge Base] process.
Configuration notes
Name
Empty
Object + purpose
Distinguish the scope of the materials
All knowledge bases
Materials with different permissions or lifecycles should be separated
Embedding model
Do not use
Do not use it for the first-time experience
Determines whether vector retrieval is included
Colloquial questions, many paraphrases
Billing and data processing for cloud models depend on the service provider
Local embedding model
Not downloaded
Download only when local processing is needed
Complete vectorization on this machine
Offline or high privacy requirements
You still need to separately check parsing, reranking, and the chat model
Expected results
The new knowledge base appears in the list, and its name can be distinguished from other knowledge bases.
You know whether the current retrieval is keyword retrieval or hybrid retrieval.
The selected cloud model can be called, or the local model has already been downloaded.
User case
Xiaolin created an [Employee Travel Policy] knowledge base for the first time. He first chose [Do Not Use] for the embedding model, imported three policies, and completed the recall test. After keyword queries became stable, he configured an embedding model and compared the results of colloquial questions using the same questions.
The completion criterion is: after upgrading the retrieval method, the original fixed questions do not regress, and colloquial questions can more stably find the same policy.
FAQ
Continue reading
Last updated
Was this helpful?
