Struct FastembedEmbedder 

pub struct FastembedEmbedder { /* private fields */ }

ONNX-backed EmbeddingGemma 300M embedder. See module-level doc for the adapter strategy and failure-mode table.

Implementations

impl FastembedEmbedder

pub fn new(opts: FastembedEmbedderOpts) -> Result<Self, LunarisError>

Construct a real ONNX-backed embedder. On first call this triggers an HF Hub download of the EmbeddingGemma 300M weights (~600 MB) into opts.cache_dir; subsequent calls hit the cache.

Construction is synchronous because fastembed’s try_new is itself synchronous — the I/O happens inline. Callers that need to avoid stalling the runtime should wrap this in tokio::task::spawn_blocking at the call site; we deliberately do NOT wrap inside new so the error mapping stays straightforward and the caller controls the spawn context.

pub fn from_user_defined( opts: FastembedUserDefinedOpts, ) -> Result<Self, LunarisError>

Bring-your-own ONNX model (Phase 20 Plan 20-01). The operator supplies the model bytes + tokenizer bytes in FastembedUserDefinedOpts and declares the output dimensionality (dim); the constructor wires fastembed’s UserDefinedEmbeddingModel / InitOptionsUserDefined and returns a ready embedder.

Trust requirement

The ONNX bytes execute in-process through ONNX Runtime. They MUST come from a trusted source (operator-controlled model registry, not user-uploaded content) — Lunaris performs no graph validation. See .planning/phases/20-fastembed-adoption/20-01-PLAN.md threat T-20-01-01.

Storage-dim constraint

Lunaris’s default storage schema is 768-d (Moon FT index + Postgres vector(768) column). Operators bringing a model whose dim != 768 MUST also reindex storage — this is the storage-side migration covered by Plan 20-03. Lunaris does NOT enforce dim parity between embedder and storage on the hot path; a mismatch surfaces as a backend insert error at first ingest.

Example

use std::sync::Arc;
use lunaris_embed::fastembed::{
    FastembedEmbedder, FastembedUserDefinedOpts, PoolingMode, ExecutionPreference,
};

let onnx = std::fs::read("models/helios-finetuned.onnx")?;
let tok = std::fs::read("models/helios-finetuned/tokenizer.json")?;
let embedder = FastembedEmbedder::from_user_defined(FastembedUserDefinedOpts {
    onnx_file: onnx,
    tokenizer_file: tok,
    tokenizer_config_file: None,
    special_tokens_map_file: None,
    config_file: None,
    dim: 1024, // MUST match the ONNX model's output dim
    pooling: PoolingMode::Mean,
    execution: ExecutionPreference::Cpu,
    max_length: 2048,
})?;
// let lunaris = Lunaris::open(url).await?.with_embedder(Arc::new(embedder));
let _ = Arc::new(embedder);

Trait Implementations

impl Clone for FastembedEmbedder

fn clone(&self) -> FastembedEmbedder

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for FastembedEmbedder

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Embedder for FastembedEmbedder

fn dim(&self) -> usize

fn embed_batch<'life0, 'life1, 'life2, 'async_trait>( &'life0 self, inputs: &'life1 [&'life2 str], ) -> Pin<Box<dyn Future<Output = Result<Vec<Vec<f32>>, LunarisError>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait, 'life2: 'async_trait,