Module embeddings 

Universal embedding client with config-driven provider cascade.

Supports any OpenAI-compatible embedding API (Ollama, vLLM, TEI, etc.) Providers are tried in priority order until one responds.

Example config.toml

[embeddings]
required_dimension = 2560
max_batch_chars = 32000
max_batch_items = 16

[[embeddings.providers]]
name = "ollama-local"
base_url = "http://localhost:11434"
model = "qwen3-embedding:4b"
priority = 1

[[embeddings.providers]]
name = "dragon"
base_url = "http://dragon:12345"
model = "Qwen/Qwen3-Embedding-4B"
priority = 2

Structs

DimensionAdapter

Adapter for cross-dimension embedding compatibility.

EmbeddingClient

Universal embedding client with provider cascade

EmbeddingConfig

Complete embedding configuration

MlxConfig

Legacy MLX configuration - deprecated, use EmbeddingConfig instead

MlxMergeOptions

Options for merging file config into MlxConfig

ProviderConfig

Single embedding provider configuration

RerankerConfig

Reranker configuration (optional, separate from embedders)

TokenConfig

Token estimation configuration

Constants

DEFAULT_MAX_TOKENS

Default safe ceiling for qwen3-embedding context window. Real model limit is 40 960 tokens; we leave ~6k margin for safety.

DEFAULT_OLLAMA_EMBEDDING_MODEL

DEFAULT_REQUIRED_DIMENSION

Functions

cross_dimension_search_adapt

Perform cross-dimension search by adapting query embedding

estimate_tokens

Estimate token count for text

safe_chunk_size

Calculate safe chunk size in characters for given token limit

truncate_to_token_limit

Truncate text to fit within token limit

validate_batch_tokens

Validate a batch of texts and return which ones exceed limits

validate_chunk_tokens

Validate that a chunk fits within token limits

Type Aliases

MLXBridge