Skip to content
Bedrova bedrova
FeaturesCompatibilityDevelopersCompareDocs Coming soon

Docs

Embeddings & multimodal search

Turn text, images, and supported audio/transcript artifacts into vectors locally — the foundation of document chat, search, rerank, and clustering.

Embeddings turn text, images, and supported audio/transcript artifacts into numeric vectors so you can search and compare by meaning instead of exact words. They power Documents & RAG, visual document retrieval, image search, audio/transcript search, and reranking. Everything runs on-device.

The request

/v1/embeddings takes one string or a batch:

curl http://127.0.0.1:1337/v1/embeddings \
  -H 'Content-Type: application/json' \
  -d '{"model": "nomic-embed", "input": ["first sentence", "second sentence"]}'

You get one vector per input, in the OpenAI embeddings shape, ready to store or compare. For image/audio/mixed inputs, use the native /api/v1/embeddings/multimodal route so the request can carry local artifacts and modality metadata.

Choosing a model

Embedding models are small and cheap — Nomic Embed peaks around 1 GB — so they sit alongside a chat model in memory comfortably. A few things to weigh:

  • Dimensions. Higher-dimension vectors capture more but cost more to store and compare. Match the dimension your vector store expects.
  • Domain. General models (Nomic, BGE, Gemma, MiniLM) cover most text; pick a code-or-domain-tuned model if your corpus is specialized.
  • Consistency. Embed your documents and your queries with the same model — vectors from different models aren’t comparable. Re-embed if you switch models.
  • Modality. Text, image, visual-page, and audio/transcript vectors may require different providers. Bedrova shows that in Settings and rebuilds indexes when you switch.

Set your default embedding/search model in Settings → RAG/Embeddings.

Batching and performance

Send inputs in batches rather than one request each — it’s markedly faster. The embedding lane runs in parallel with chat and other workloads (see Concepts), so indexing a folder doesn’t block a conversation.

See also