Module embeddings 

Semantic search via embeddings.

This module provides embedding generation and vector similarity search for semantic code search. It supports multiple embedding providers:

• OpenAI: Uses text-embedding-3-small for high-quality embeddings
• Local: Uses fastembed for local, offline embeddings

Architecture

Embeddings are stored in SQLite as JSON-encoded float arrays. Similarity search is performed in Rust using cosine similarity for accuracy and portability (no native vector extensions required).

Usage

let provider = OpenAIProvider::new("sk-...")?;
let embedding = provider.embed("fn authenticate(user: &str)")?;

let results = search_similar(&db, "authentication functions", 10)?;

Re-exports

pub use local::LocalProvider;

pub use ollama::OllamaProvider;

pub use openai::OpenAIProvider;

Modules

local

Local embedding provider using fastembed.

ollama

Ollama embedding provider.

openai

OpenAI embedding provider.

Structs

Embedding

A vector embedding.

SearchResult

Search result from similarity search.

Enums

Provider

Which embedding backend to use.

Constants

LOCAL_EMBEDDING_DIM

OPENAI_EMBEDDING_DIM

Embedding dimension for different models

Traits

EmbeddingProvider

Trait for embedding providers.

Functions

build_provider

Build the embedding provider for the given backend, applying any provider-specific settings from .ctx/config.toml (embedding). This is the single place providers are constructed, so a new backend wires in once.

cosine_similarity

Compute cosine similarity between two vectors.

dot_product

Compute dot product similarity between two vectors.

embed_missing_symbols

Embed all symbols that don’t have embeddings yet.

normalize

Normalize a vector to unit length.

semantic_search

Perform semantic similarity search using embeddings.

warn_index_mismatch

Warn (to stderr) when the query provider/dimension differs from what the index was embedded with. Embeddings from different providers/models occupy different vector spaces, so mixing them yields meaningless similarities — the fix is to re-embed. No-op when the index is empty or consistent.