Crate rlx_embed 

RLX-backed text and image embedding models.

Migrated from burnembed — compiles BERT / NomicBERT / NomicVision graphs via rlx-runtime and exposes tier-0 inference helpers.

use rlx_models::embed::{Pooling, RlxBertModel, BertTokenizer, embed_with_rlx};

let tok = BertTokenizer::from_dir(model_dir, 512)?;
let mut model = RlxBertModel::load(&config, &weights)?;
let vecs = embed_with_rlx(&mut model, &tok, &["hello", "world"], Pooling::Mean)?;

Structs

BertTokenizer

Wrapper around HuggingFace tokenizer configured for BERT-style encoding.

ImageModelInfo

Metadata for an image embedding model.

ModelInfo

Metadata for an embedding model.

RlxBertModel

RLX-compiled BERT model ready for inference.

RlxEmbed

High-level embedding model — auto-detects BERT / NomicBERT / NomicVision.

RlxNomicModel

RLX-compiled NomicBERT with shape-bucketed compile cache.

RlxVisionModel

RLX-compiled NomicVision encoder (patch preprocess host-side, trunk on RLX).

TokenizedBatch

Output of batch tokenization: token IDs, attention masks, and token type IDs.

Enums

Arch

Detected embedding architecture from config.json.

EmbeddingModel

Supported text embedding models.

ImageEmbeddingModel

Supported image embedding models.

ModelArch

Model architecture type.

Pooling

Pooling strategy for reducing token hidden states to one vector per sequence.

Functions

assemble_vision_hidden

Assemble encoder input [batch, seq, hidden] from NCHW pixels + preprocess weights.

compile_model

Compile an embedding graph for the given batch/seq on device.

compile_model_cpu

Compile on CPU (convenience for tests and default RlxEmbed::from_dir).

default_pooling

Default pooling heuristic from HuggingFace repo id.

detect_arch

Detect architecture from config.json fields.

embed_with_rlx

Embed texts with a compiled BERT model: tokenize, forward, pool, L2-normalize.

l2_normalize_in_place

L2-normalize a vector in place (matches fastembed: divide by norm + 1e-12).

models_map

Get the global model registry.

pool_embeddings

Pool [batch, seq, hidden] hidden states into [batch, hidden] and L2-normalize.