Embedding models for knowledge search
Embedding options turn a knowledge search into numeric values for matching by meaning, separately from the chat model.
What it is
Embeddings Model selects the model that turns a knowledge-search question into numeric values representing its meaning. Weaviate can compare these values with its stored content. The embedding model and credential are separate from the model that writes the conversation reply.
Who uses it and why
Operators with the Developer role, or the Settings role with AI access, configure embeddings to match their external knowledge index and service. Operators and visitors use the resulting knowledge answers without selecting an embedding model.
Where to find it
Open Settings › AI › Functions and edit a Knowledge Base function with Knowledge Base Type set to Weaviate. Its options include Embeddings Model, Embeddings Endpoint and OpenAI API Key.
You need the Developer role, or the Settings role with AI access, to configure this client-managed screen.
How it works
- The AI feature calls a knowledge function with a search question.
- For Vector, and for a working Hybrid configuration, INSIDE asks the embedding service to represent that question numerically.
- It sends the result to the external search service alongside the configured retrieval options.
- The retrieved content returns to the AI feature, which can use it when composing an answer.
Keyword search does not call the embedding service. It searches the words in the question directly.
Configure it
- Choose the appropriate Weaviate Query Type on the knowledge function.
- Enter Embeddings Model to choose a specific model. If it is unset, INSIDE uses an older embedding-model choice already stored on that function, where present, then text-embedding-ada-002.
- Enter Embeddings Endpoint to choose the embedding service explicitly. Otherwise INSIDE uses the feature account's API Endpoint, then the built-in OpenAI embedding service.
- Enter the embedding service credential in OpenAI API Key. INSIDE does not copy the chat account's key into this field.
- Save the function and check retrieval through the intended AI feature.
An explicit Embeddings Endpoint can be the service base or its full embeddings address. INSIDE adds the embeddings suffix when it is missing from that explicit value.
Settings
| Setting | Where | Applies to | Default | Who can change it | What it does |
|---|---|---|---|---|---|
| Embeddings Model | Settings › AI › Functions | website | text-embedding-ada-002 unless an older model choice is configured | Settings role, Developer role | Overrides the model used to embed a Weaviate search; without an override or an older configured model, the default is text-embedding-ada-002. |
| Embeddings Endpoint | Settings › AI › Functions | website | Not set | Settings role, Developer role | Overrides the embedding service address; otherwise the feature account's API Endpoint is used, followed by the built-in OpenAI embedding service. |
| Weaviate Query Type | Settings › AI › Functions | website | Keyword | Settings role, Developer role | Vector and Hybrid searches request embeddings; Keyword does not. New function forms initially select Keyword. |
Fields
| Field | Where | Required | What it is for |
|---|---|---|---|
| OpenAI API Key | Settings › AI › Functions | No | Supplies the embedding service credential independently of the account's chat credential; no key is supplied by default. |
Permissions
| Permission | Lets you |
|---|---|
| AI | With the Settings role, allows configuration of AI Functions and their embedding options. |
Limits and good to know
- The service must accept an OpenAI-compatible embedding request and return compatible numeric embeddings.
- Use a model compatible with the stored knowledge index. INSIDE does not establish that match for you.
- Changing the chat account's API Endpoint can also change embedding requests when Embeddings Endpoint is unset. Set it explicitly when the two services need different addresses.
- Changing the embedding model here does not regenerate the content already stored in the external index.
- Embedding calls can have their own service costs. Do not assume completion-token totals include every embedding request.
- Hybrid needs Powerfront verification in this release before you rely on it. Its saved selection can fail to match the accepted search mode.
Depends on and works with
View all connections as a list
This feature uses
- AI functions and required inputsRequires
Embedding options are configured on a Knowledge Base function.
- AI accounts and shared defaultsConfigured by
The feature's AI account supplies the fallback endpoint when no embedding endpoint is specified.
- Knowledge Base connections for AIConfigured by
The Weaviate Knowledge Base function holds the embedding model, endpoint, credential and search-mode choices.
Used by
- Knowledge Base connections for AIEnhances
Vector and hybrid Weaviate retrieval need an embedding service; keyword retrieval does not.
- Saved responses as AI knowledgeConfigured by
For Vector or Hybrid retrieval, the generated search function's embedding options control meaning-based matching.
FAQ and troubleshooting
Why does Keyword work while Vector fails?
Keyword does not call the embedding service. Check OpenAI API Key, Embeddings Endpoint and Embeddings Model, then confirm compatibility with the external index.
Does changing my chat model change the embedding model?
No. Embeddings have a separate model selection. The account endpoint can still be inherited when no embedding endpoint is set.
Why are searches going to my chat service?
Without an explicit Embeddings Endpoint, INSIDE first falls back to the feature account's API Endpoint. Set the embedding address explicitly if that service does not provide embeddings.
Will choosing a different embedding model rebuild my index?
No. Arrange any required reindexing with whoever maintains the external knowledge service.