Expand description
Embroider — Jina v5 text embeddings through ONNX Runtime.
Rust core with opt-in PyO3 bindings (the bobine pattern). The crate
layering: providers/probe/policy/acquire/error/diag
are the shared ONNX plumbing (session building, provider fallback, CUDA
probing, HF acquisition); jina is the frozen text-embedding contract.
Provenance: a clean move out of OKFgraph’s rust/okf-embed (itself an
exact port of EmbeddingEngine._encode, pinned by that repo’s parity
harness at ≤1e-5). Session/threading choices follow EmbedAnything’s
ort_jina.rs; device fallback semantics mirror bobine: CUDA is
opportunistic, never fatal.
§Module layout
error— anyhow-based error plumbing (oestringification)providers— provider-name matrix + clone-and-fallback applicationprobe— corrected CUDA availability check (OnceLock-cached)policy—DeviceReq+ explicitSessionPolicy(text vs defaults)acquire— validated HF hub acquisition (blocking, cached)diag—OrtReportobservation for logs and diagnosticsjina—JinaV5+TokenizerHandle(the frozen contract)vision—JinaV5Vision(the dynamic-grid image contract)
Re-exports§
pub use acquire::artifact_for;pub use acquire::artifact_for_task;pub use acquire::builtin_models;pub use acquire::builtin_tasks;pub use acquire::cache_info;pub use acquire::cache_info_files;pub use acquire::fetch_repo_file;pub use acquire::fetch_tokenizer_file;pub use acquire::lookup_model;pub use acquire::lookup_task;pub use acquire::parse_owner_name;pub use acquire::repo_for_precision;pub use acquire::Artifact;pub use acquire::CacheReport;pub use acquire::ModelSpec;pub use acquire::TaskKind;pub use acquire::TaskSpec;pub use acquire::FP16_TEXT_MODEL;pub use acquire::FP32_TEXT_MODEL;pub use acquire::GLICLASS_SMALL;pub use acquire::GLICLASS_SMALL_MODEL;pub use acquire::GLICLASS_V4_EDGE;pub use acquire::GLICLASS_V4_EDGE_MODEL;pub use acquire::LAYA_BASE;pub use acquire::LAYA_BASE_MODEL;pub use acquire::LAYA_MULTI;pub use acquire::LAYA_MULTI_MODEL;pub use acquire::LAYA_TD;pub use acquire::LAYA_TD_MODEL;pub use acquire::NANO_TEXT_MODEL;pub use acquire::RERANK_MMARCO;pub use acquire::RERANK_MMARCO_MODEL;pub use acquire::RERANK_MINILM;pub use acquire::RERANK_MINILM_MODEL;pub use acquire::TEXT_NANO;pub use acquire::TEXT_SMALL;pub use acquire::DECIDER2B_MODEL;pub use diag::configure_runtime;pub use diag::report;pub use diag::runtime_report;pub use diag::runtime_report_for;pub use diag::save_runtime_report;pub use diag::OrtReport;pub use jina::JinaV5;pub use jina::TokenizerHandle;pub use jina::MAX_LENGTH;pub use jina::MODEL_MAX_TOKENS;pub use jina::NATIVE_DIM;pub use rerank::Reranker;pub use rerank::RERANK_ABSOLUTE_MAX;pub use rerank::RERANK_DEFAULT_MAX;pub use decide::Decider;pub use decide::Decision;pub use decide::Question;pub use decide::DECIDE_ABSOLUTE_MAX;pub use decide::DECIDE_DEFAULT_MAX;pub use decider2b::Decider2B;pub use decider2b::Decider2BDecision;pub use decider2b::Decider2BOption;pub use decider2b::Decider2BQuestion;pub use laya::LayaDecider;pub use laya::LayaDecision;pub use laya::LayaLongDecision;pub use laya::LayaOption;pub use laya::LayaQuestion;pub use laya::LayaTemperatures;pub use laya::BinningMap;pub use vision::vision_target_size;pub use vision::JinaV5Vision;pub use vision::VISION_MAX_PIXELS;pub use vision::VISION_MIN_PIXELS;pub use acquire::VISION_NANO;pub use acquire::VISION_NANO_MODEL;pub use acquire::VISION_NANO_REPO;pub use policy::DeviceReq;pub use policy::Precision;pub use policy::SessionPolicy;pub use probe::cuda_available;pub use providers::apply_providers;pub use providers::apply_providers_with_arena;pub use providers::map_provider;pub use providers::ProviderMapping;
Modules§
- acquire
- HuggingFace acquisition (blocking, cached).
- decide
- The typed-decision contract (GLiClass uni-encoder).
- decider2b
- Slot-layout typed decisions: serving Mapika/decider-2b.
- diag
- Runtime observation shared by consumers’ logs and diagnostics.
- error
- Error plumbing.
- jina
- The Jina v5 text-embedding contract.
- laya
- Laya typed-decision serving: the marker-layout contract.
- policy
- Device requests and session tuning policy.
- probe
- CUDA availability probing against the loaded ONNX Runtime library.
- providers
- ONNX Runtime execution-provider plumbing.
- rerank
- The cross-encoder rerank contract.
- vision
- The Jina v5 vision contract: dynamic-grid image embeddings (Phase 6).