Embeddings

Embeddings are numerical vector representations that capture the semantic meaning of text and code, enabling AI tools to understand similarity between code snippets that may look completely different syntactically but serve the same purpose. When an embedding model processes a function, it outputs a high-dimensional vector, typically 768 to 3,072 floating-point numbers, that encodes the function's purpose, patterns, and relationships.

The power of embeddings for coding lies in their ability to map semantically related code close together in vector space. A function called 'validateEmail' and another called 'checkEmailFormat' would produce similar embedding vectors even though they share no common words, because the embedding model understands they serve the same purpose. This enables semantic code search: you can search for 'authentication logic' and find relevant code even if no file contains those exact words.

Code embedding models are specifically trained to understand programming concepts. Models like OpenAI's text-embedding-3-large (3,072 dimensions), Voyage AI's voyage-code-2, and Nomic's nomic-embed-code are optimized for code understanding. They capture relationships between functions, understand type hierarchies, and recognize design patterns. The embedding process is relatively fast and cheap compared to running a full LLM: embedding a million tokens costs roughly $0.02-0.13 depending on the model.

In practice, embeddings are stored in vector databases like Pinecone, Chroma, or Qdrant, where they can be searched using cosine similarity or dot product operations. This infrastructure powers the RAG pipelines in AI coding tools, enabling sub-second retrieval of relevant code from repositories with millions of lines.

Cursor uses embeddings to index your entire project when you first open it. This embedding index powers its codebase-aware suggestions and the @Codebase reference in Chat and Composer. The indexing runs locally and updates incrementally as you edit files. Cody by Sourcegraph leverages enterprise-scale embedding infrastructure to index entire organizations' codebases, supporting cross-repository code search that finds relevant patterns across hundreds of repos.

Continue lets you choose your embedding provider, supporting OpenAI, Voyage AI, and local models through Ollama for privacy-sensitive environments. This flexibility means you can run embeddings entirely on your own hardware. Windsurf builds its codebase understanding through embeddings that power its intelligent context selection. Even tools without visible RAG features, like GitHub Copilot, use embedding-like representations internally to match your current code context with relevant training patterns.

What is Embeddings?

Numerical vector representations of text that capture semantic meaning, enabling similarity search and retrieval.

Why is Embeddings important in AI coding?

Embeddings are numerical vector representations that capture the semantic meaning of text and code, enabling AI tools to understand similarity between code snippets that may look completely different syntactically but serve the same purpose. When an embedding model processes a function, it outputs a high-dimensional vector, typically 768 to 3,072 floating-point numbers, that encodes the function's purpose, patterns, and relationships. The power of embeddings for coding lies in their ability to map semantically related code close together in vector space. A function called 'validateEmail' and another called 'checkEmailFormat' would produce similar embedding vectors even though they share no common words, because the embedding model understands they serve the same purpose. This enables semantic code search: you can search for 'authentication logic' and find relevant code even if no file contains those exact words. Code embedding models are specifically trained to understand programming concepts. Models like OpenAI's text-embedding-3-large (3,072 dimensions), Voyage AI's voyage-code-2, and Nomic's nomic-embed-code are optimized for code understanding. They capture relationships between functions, understand type hierarchies, and recognize design patterns. The embedding process is relatively fast and cheap compared to running a full LLM: embedding a million tokens costs roughly $0.02-0.13 depending on the model. In practice, embeddings are stored in vector databases like Pinecone, Chroma, or Qdrant, where they can be searched using cosine similarity or dot product operations. This infrastructure powers the RAG pipelines in AI coding tools, enabling sub-second retrieval of relevant code from repositories with millions of lines.

How do I use Embeddings effectively?

Choose a code-specific embedding model like Voyage Code 2 or OpenAI text-embedding-3-large over general-purpose text embeddings for significantly better code retrieval quality When using Continue with local embeddings, run a code-optimized model through Ollama for privacy while maintaining good retrieval accuracy Write clear function documentation and descriptive names to improve embedding quality, as models capture semantic meaning from identifiers and comments