Skip to main content
Embeddings are numerical representations of text that capture semantic meaning. They convert text into vectors (arrays of numbers) that can be used for various machine learning tasks. OpenRouter provides a unified API to access embedding models from multiple providers.

What are embeddings?

Embeddings transform text into high-dimensional vectors where semantically similar texts are positioned closer together in vector space. For example, “cat” and “kitten” would have similar embeddings, while “cat” and “airplane” would be far apart. These vector representations enable machines to understand relationships between pieces of text, making them essential for many AI applications.

Common use cases

Embeddings are used in a wide variety of applications: RAG (Retrieval-Augmented Generation): Build RAG systems that retrieve relevant context from a knowledge base before generating answers. Embeddings help find the most relevant documents to include in the LLM’s context. Semantic Search: Convert documents and queries into embeddings, then find the most relevant documents by comparing vector similarity. This provides more accurate results than traditional keyword matching because it understands meaning rather than just matching words. Recommendation Systems: Generate embeddings for items (products, articles, movies) and user preferences to recommend similar items. By comparing embedding vectors, you can find items that are semantically related even if they don’t share obvious keywords. Clustering and Classification: Group similar documents together or classify text into categories by analyzing embedding patterns. Documents with similar embeddings likely belong to the same topic or category. Duplicate Detection: Identify duplicate or near-duplicate content by comparing embedding similarity. This works even when text is paraphrased or reworded. Anomaly Detection: Detect unusual or outlier content by identifying embeddings that are far from typical patterns in your dataset.

How to use embeddings

Basic request

To generate embeddings, send a POST request to /embeddings with your text input and chosen model:
Output

Batch processing

You can generate embeddings for multiple texts in a single request by passing an array of strings:
Output

Image input

Some embedding models support image inputs, enabling multimodal embeddings that capture visual content alongside text. This is useful for image search, visual similarity, and cross-modal retrieval tasks. To send an image, wrap your input in the multimodal format with a content array containing image_url objects. You can also combine text and images in a single input block.
Output
You can also combine text and images in a single input to generate a joint embedding:
Output

Contextual embeddings

Some embedding models embed every chunk of a document with the rest of that document as context, so each chunk’s vector carries the meaning of the document it came from. This improves retrieval when chunks are short or ambiguous on their own, such as a paragraph that refers to “it” or “the company” without naming the subject. To use contextual mode, send input as an array of documents, where each document is an ordered array of chunk strings, and set input_type to "document". You get one embedding back per chunk, in document order.

Response shape

The response is the same flat data array as every other embeddings request, with one extra field. index counts across all chunks in the request, and document_index is the position of the chunk’s document in your input:
Providers that offer contextual embeddings natively, such as Perplexity, return a nested response: one list per document, each with its own data array and a chunk index that restarts at 0 in every document. OpenRouter flattens that into the single data array that existing embeddings clients expect. If you are porting code from a provider’s SDK, read document_index instead of nesting, and treat index as global rather than per document.
Chunks of the same document are contiguous and in the order you sent them, so grouping is a one-liner:

Querying a contextual index

Embed search queries as a plain string or string array with input_type: "query". Queries have no document context, so they never use the nested shape:

Constraints

Only models that advertise contextual embeddings accept nested input; other models accept strings and string arrays, and token arrays only where the provider supports them. Contextual input is text only.

API reference

For detailed information about request parameters, response format, and all available options, see the Embeddings API Reference.

Available models

OpenRouter provides access to various embedding models from different providers. You can view all available embedding models at: https://openrouter.ai/models?fmt=cards&output_modalities=embeddings To list all available embedding models programmatically:
Output
Here’s a complete example of building a semantic search system using embeddings:
Expected output:

Best practices

Choose the right model: Different embedding models have different strengths. Smaller models (like perplexity/pplx-embed-v1-0.6b or openai/text-embedding-3-small) are faster and cheaper, while larger models (like openai/text-embedding-3-large) provide better quality. Test multiple models to find the best fit for your use case. Batch your requests: When processing multiple texts, send them in a single request rather than making individual API calls. This reduces latency and costs. Cache embeddings: Embeddings for the same text are deterministic (they don’t change). Store embeddings in a database or vector store to avoid regenerating them repeatedly. Normalize for comparison: When comparing embeddings, use cosine similarity rather than Euclidean distance. Cosine similarity is scale-invariant and works better for high-dimensional vectors. Consider Context Length: Each model has a maximum input length (context window). Longer texts may need to be chunked or truncated. Check the model’s specifications before processing long documents. Use Appropriate Chunking: For long documents, split them into meaningful chunks (paragraphs, sections) rather than arbitrary character limits. This preserves semantic coherence.

Provider routing

You can control which providers serve your embedding requests using the provider parameter. This is useful for:
  • Ensuring data privacy with specific providers
  • Optimizing for cost or latency
  • Using provider-specific features
Example with provider preferences:
For more information, see Provider Routing.

Error handling

Common errors you may encounter: 400 Bad Request: Invalid input format or missing required parameters. Check that your input and model parameters are correctly formatted. 401 Unauthorized: Invalid or missing API key. Verify your API key is correct and included in the Authorization header. 402 Payment Required: Insufficient credits. Add credits to your OpenRouter account. 404 Not Found: The specified model doesn’t exist or isn’t available for embeddings. Check the model name and verify it’s an embedding model. 429 Too Many Requests: Rate limit exceeded. Implement exponential backoff and retry logic. 529 Provider Overloaded: The provider is temporarily overloaded. Enable allow_fallbacks: true to automatically use backup providers.

Limitations

  • No Streaming: Unlike chat completions, embeddings are returned as complete responses. Streaming is not supported.
  • Token Limits: Each model has a maximum input length. Texts exceeding this limit will be truncated or rejected.
  • Deterministic Output: Embeddings for the same input text will always be identical (no temperature or randomness).
  • Language Support: Some models are optimized for specific languages. Check model documentation for language capabilities.