Model provider integrations
Weaviate integrates with a variety of self-hosted and API-based models from a range of providers.
This enables an enhanced developed experience, such as the ability to:
- Import objects directly into Weaviate without having to manually specify embeddings, and
- Build an integrated retrieval augmented generation (RAG) pipeline with generative AI models.
Model provider integrations
API-based
Enable all API-based modules
All API-based model integrations are available by default starting with Weaviate v1.33.
To opt out, for example in an air-gapped or otherwise restricted deployment, set the API_BASED_MODULES_DISABLED environment variable to true. Weaviate then loads only the modules that you list in ENABLE_MODULES. This variable was added in v1.33.
For releases before v1.33, enable all API-based modules by setting the ENABLE_API_BASED_MODULES environment variable to true. Weaviate stopped reading that variable in v1.33.
Locally hosted
| Model provider | Embeddings | Generative AI | Others |
|---|---|---|---|
| GPT4All (Deprecated) | Text (Deprecated) | - | - |
| Hugging Face | Text, Multimodal (CLIP) | - | Reranker |
| KubeAI | Text | - | - |
| Model2vec | Text | - | - |
| Meta ImageBind | Multimodal | - | - |
| Ollama | Text | Text | - |
Course: Embedding Model Evaluation & Selection
Embedding models are the heart of vector search. Learn how to evaluate and select appropriate embedding models for your use case.
Open Academy CourseHow does Weaviate generate embeddings?
When a model provider integration for embeddings is enabled, Weaviate automatically generates embeddings for objects that are added to the database.
This is done by providing the source data to the integration provider, which then returns the embeddings to Weaviate. The embeddings are then stored in the Weaviate Database.
Weaviate generates embeddings for objects as follows:
- Selects properties with
textortext[]data types unless they are configured to be skipped - Sorts properties in alphabetical (a-z) order before concatenating values
- Prepends the collection name if configured
For Weaviate versions before v1.27, the string created above is lowercased before being sent to the model provider. Starting in v1.27, the string is sent as is.
If you prefer the text to be lowercased, you can do so by setting the LOWERCASE_VECTORIZATION_INPUT environment variable.
The text is always lowercased for the text2vec-contextionary integration.
Rate limits for API-based embeddings
Model providers throttle how fast you can call them, so a large import often ends in 429 errors. To keep Weaviate inside your quota, tell it what the quota is with two request headers:
X-<Provider>-Ratelimit-RequestPM-Embedding: the requests-per-minute limitX-<Provider>-Ratelimit-TokenPM-Embedding: the tokens-per-minute limit
<Provider> is the same name as in the API key header on the integration's page. A collection vectorized by Cohere, for example, takes X-Cohere-Ratelimit-RequestPM-Embedding next to X-Cohere-Api-Key. Set the headers when you connect, so every batch request carries them.
The values must be whole numbers. Weaviate falls back to the integration's own default if a header is missing or cannot be parsed.
The headers are read by the Cohere, Databricks, DigitalOcean, Jina AI, Mistral, OpenAI, Azure OpenAI, Voyage AI, and Weaviate Embeddings integrations. Other integrations ignore them.
Troubleshooting: a missing or misconfigured API key
Every API-based integration reports a missing key the same way. If the key is absent, the request fails with:
no api key found neither in request header: X-<Provider>-Api-Key nor in environment variable under <PROVIDER>_APIKEY
The API key is never part of the collection configuration, so supply it as a request header when you connect, or set the environment variable on the Weaviate server. A key in the header takes precedence over the environment variable.
Questions and feedback
Have a question or feedback? Here's how to reach us.