Embedding providers and setup
Choose a search index model, configure remote or local embeddings, and verify the rebuilt index
Embeddings turn content into vectors for semantic search. In Nowledge Mem, this setting is called the Search Index Model. Configure it separately from the chat model used by AI Now.
Keep the built-in local model if local search meets your needs. To move vector computation to a remote service, get its API key and an embedding model ID, then open Settings → Providers → Search Index Provider.
Chat subscriptions do not provide an embedding API
ChatGPT Subscription and SuperGrok sign-in cannot supply the search index
model. OpenAI embeddings require a separate OpenAI API key. Chat models such
as gpt-*, grok-*, glm-*, and Gemini Flash cannot replace an embedding
model.
Choose a provider and model
These are the main entries available in the search index provider selector. Model names are examples, not a quality ranking. Check the provider's account and documentation for model access, quotas, and regions.
| Provider | Example model or where to find one | Setup note |
|---|---|---|
| OpenAI | text-embedding-3-small, text-embedding-3-large | Add an OpenAI API key separately from your ChatGPT subscription |
| Gemini | gemini-embedding-001 | Use a Google AI Studio API key and an embedding model |
| Jina AI | jina-embeddings-v5-text-small | Add a Jina API key; check the model's task parameter requirements |
| Fireworks AI | Find an embedding model ID available to your account | Use the full model ID returned by the provider |
| OpenRouter | For example, openai/text-embedding-3-small | Keep the provider prefix in the model ID |
| Vercel AI Gateway | Select an embedding model from the gateway catalog | Use gateway credentials and the full model ID |
| Ollama, LM Studio, Lemonade | An embedding model loaded by the local service | The service must expose an OpenAI-compatible embedding endpoint; see local model setup |
| Magpie | An embedding model supplied by its upstream service | Both the gateway and upstream must support embeddings; a working chat subscription route is not sufficient |
NVIDIA NIM and GitHub Copilot need additional verification. They can appear in the selector without every model being usable. Some NVIDIA models require input_type to distinguish queries from documents. The current generic remote request does not send that parameter. GitHub Copilot embedding authentication and endpoint support also need separate verification. A model listing alone is not enough to start a large rebuild.
Nowledge AI manages its search index model, so you do not select the underlying model. The bring-your-own-key steps here apply to local or self-hosted workspaces where you can manage providers. Nowledge Cloud manages its search index model.
Configure and apply
- Open Settings → Providers, add a provider from the table, and save. Enter an API key when the service requires one; local services can leave it blank if authentication is disabled.
- Select it under Search Index Provider. Chat and search indexing can use different providers.
- Select an embedding model. If the list cannot load, paste the official model ID. Do not enter a chat model name.
- Click Apply and follow the embedding test and index rebuild confirmation. A successful chat connection test does not test embeddings.
- Wait for the search index rebuild to finish before evaluating results. Check index status under Settings → Memory Processing → Search. Server users can also run
nmem statusandnmem models status.
Success means the selected index provider and model are visible, the index no longer needs a rebuild, and search finds saved content with different wording. For example, save a memory about reviewing project progress every Friday, then search for a weekly work review schedule. One successful test covers only that request; complete the initial batch rebuild and test real searches too.
Remote embedding sends text being indexed and search queries to the selected provider and uses its quota. Check pricing, data handling terms, and request limits before use.
Add another service with Custom
Services such as Alibaba Cloud Model Studio, SiliconFlow, or a private gateway can use Add Provider → Custom if they expose a compatible embedding API:
- Select Embedding as the purpose when creating the provider, rather than chat.
- Enter a name, API key, API base URL, and one embedding model ID.
- Enter the embedding API root as the base URL. Mem appends
/embeddings: for an endpoint athttps://example.com/v1/embeddings, enterhttps://example.com/v1. - Check the test result after saving. A failed test can leave the configuration saved; correct the address, model, or permissions before selecting it for search.
- Confirm the provider under Search Index Provider, then complete Apply and the rebuild flow.
See Alibaba Cloud text-embedding-v4 and the SiliconFlow embedding catalog for examples. The service must accept model and input and return an OpenAI-compatible vector response. Chat API compatibility alone is insufficient. Services that require extra task parameters, special authentication, or a different request format need adapter verification before use.
Zhipu's mainland China service provides an embedding API, but its endpoint and account access differ from the Z.AI international service. A working Z.AI chat configuration does not prove embedding access on the same configuration.
Why a configured provider is missing
The search index selector includes providers that are configured, have a saved model, and declare embedding support. Chat providers do not enter this list automatically. Create Custom entries with the embedding purpose. Seeing only a few choices such as Gemini can be expected even when several chat providers are configured.
Follow the rebuild prompt after changing the provider, model, vector dimensions, or model backend. Equal vector dimensions do not prove model compatibility. Query vectors from a new model must not search an index generated by the old model. See Search and relevance for more about search.