Feature Flags

The laurus crate ships with no default features. Enable embedding support as needed.

Available Flags

Feature	Description	Key Dependencies
`embeddings-candle`	Local BERT embeddings and ColBERT token vectors via Hugging Face Candle	candle-core, candle-nn, candle-transformers, hf-hub, tokenizers
`embeddings-openai`	OpenAI API embeddings	reqwest
`embeddings-multimodal`	CLIP multimodal embeddings (text + image)	image, embeddings-candle
`embeddings-all`	All embedding features combined	All of the above

Enables CandleBertEmbedder for running BERT models locally on the CPU, and CandleColbertEmbedder (the candle_colbert schema embedder) for ColBERT token vectors used by late-interaction rescoring. Models are downloaded from Hugging Face Hub on first use.

[dependencies]
laurus = { version = "0.12", features = ["embeddings-candle"] }

`embeddings-openai`

Enables OpenAIEmbedder for calling the OpenAI Embeddings API. Requires an OPENAI_API_KEY environment variable at runtime.

[dependencies]
laurus = { version = "0.12", features = ["embeddings-openai"] }

`embeddings-multimodal`

Enables CandleClipEmbedder for CLIP-based text and image embeddings. Implies embeddings-candle.

[dependencies]
laurus = { version = "0.12", features = ["embeddings-multimodal"] }

`embeddings-all`

Convenience flag that enables all embedding features.

[dependencies]
laurus = { version = "0.12", features = ["embeddings-all"] }

TLS and Network Behavior

The embedding features use two independent TLS stacks with different trust sources:

Feature	HTTP client	TLS backend	Trust source
`embeddings-candle`, `embeddings-multimodal`	`hf-hub` (`ureq`)	rustls	Bundled Mozilla root certificates (`webpki-roots`)
`embeddings-openai`	`reqwest`	rustls	OS trust store (via `rustls-platform-verifier`)

Model downloads from Hugging Face Hub (embeddings-candle / embeddings-multimodal) use certificates bundled into the binary rather than the operating system’s trust store. This is deliberate: it lets a fully static musl binary download models inside a scratch or distroless container with no ca-certificates package installed. The tradeoff is that SSL_CERT_FILE / SSL_CERT_DIR are not honored on this path, and a custom CA installed only in the OS trust store (for example behind a corporate TLS-inspecting proxy) will not be trusted. If you need to route Hugging Face downloads through such a proxy, pre-populate the cache and point HF_HOME at it, or set HF_ENDPOINT to an internally trusted mirror.

embeddings-openai reads the OS trust store, so containers using it still need ca-certificates installed.

Feature Flag Impact on Binary Size

Enabling embedding features adds dependencies that increase compile time and binary size:

Configuration	Approximate Impact
No features (lexical only)	Baseline
`embeddings-candle`	+ Candle ML framework
`embeddings-openai`	+ reqwest HTTP client
`embeddings-multimodal`	+ image processing + Candle
`embeddings-all`	All of the above

If you only need lexical (keyword) search, you can use Laurus with no features enabled for the smallest binary and fastest compile time.

Laurus Documentation