Crate rightkit_embed 

Local embeddings and a persistent cosine vector store.

Feature boundaries: inference (ONNX text embedder) and vector (SQLite-backed cosine store). Both are default-on; a vector-only build does not pull ort or tokenizers.

Structs

AcceleratorParity

Result of the load-time vector parity check an accelerator passed; see Embedder::accelerator_parity.

CoreMlUnits

The compute units of each Core ML function (embed_32 .. embed_512).

CoreMlUnitsParseError

An unreadable compute-unit or unit-map specification.

Embedder

EmbedderConfig

LoadTimings

Where an E40 load spent its time; see Embedder::load_timings. Phases are wall-clock durations of the work as Embedder::load ran it; they nest (session_creates happen inside the probe and the parity check) and do not add up to total exactly.

Nearest

ResidentIndex

An in-memory int8 snapshot of the vectors of one model. See the module documentation.

SessionCreate

One graph build during an E40 load.

VectorStore

Enums

CoreMlComputeUnit

Which Core ML compute units run one function (MLComputeUnits).

EmbedError

ExecutionProvider

Where inference runs.

ParityCacheOutcome

What the load-time parity check did with the opt-in verdict cache (EmbedderConfig::with_parity_cache_dir).

VectorError

Constants

ACCELERATOR_PARITY_MIN_COSINE

An accelerator is admitted only if every parity prompt’s native-dimension vector has at least this cosine similarity to the CPU provider’s vector for the same prompt. Two correct runs of the same Q4 graph agree at about 0.999; a DirectML run of EmbeddingGemma 2 Q4 measured 0.58-0.72.

DEFAULT_INGEST_BATCH

Default static batch of the E40 ingest graphs on DirectML (4 or 8 are accepted).

DEFAULT_MAX_BATCH

Default rows per ONNX call for embed_batch / embed_query_batch.

DEFAULT_VECTOR_DIM

EMBEDDINGGEMMA2_COREML_MODEL_ID

EmbeddingGemma 2 Core ML build (macOS 15+): EmbeddingGemma2Text.mlpackage (one fixed-shape function per length 32, 64, 128, 256, 512) and a host-side bf16 token table, run through the Core ML framework directly (an .mlpackage cannot be loaded through ONNX Runtime). Selected when the model directory holds that layout. Same weights, different numerics from the Q4 and E40 models, so a different pipeline id.

EMBEDDINGGEMMA2_E40_MODEL_ID

EmbeddingGemma 2 E40 export: fp16, five fixed-length graphs and a host-side token table (embed_tokens.f16.npy), run on DirectML (Windows) or CPU. Selected when the model directory holds that layout. Same weights as the Q4 model, different numerics, so a different pipeline id.

EMBEDDINGGEMMA2_MODEL_ID

EmbeddingGemma 2 text-only Q4 export (b64, accuracy level 4). Selected when the model directory’s top-level config.json model_type is embedding_gemma2.

EMBEDDINGGEMMA_MODEL_ID

NEURAL_ENGINE_MAX_TOKENS

Longest graph (in tokens) the default map sends to the Neural Engine; longer graphs use the GPU.

RRF_K

The damping constant of the original RRF paper (Cormack, Clarke and Büttcher, 2009).

Functions

bundle_files

The files a model bundle in dir must hold, relative to dir, for the layout load would pick: the Core ML package (macOS), the E40 bundle (its manifest and the files it names, or the legacy graphs) or the single-file Q4 model (model_q4.onnx). None when dir holds none of these. Every listed file must exist: a missing one is MissingArtifact, as in load. A Core ML package on a non-macOS host is an InvalidConfig error.

cosine_similarity

installed_model_dir

Model directory for subdir of an installed application, derived from its executable only: the resources directory rightkit_onnx::installed_resource_dir names, joined with subdir. It returns the candidate without checking that it exists. None for a relative executable path, or when subdir is not exactly one plain name (empty, an absolute path, .., . or a nested path such as a/b).

reciprocal_rank_fusion

Fuses ranked lists, best first, by reciprocal rank. An item scores sum(1 / (k + rank)) over every list that holds it, with rank starting at 1, so an item near the top of several lists beats one at the top of only one. Returns each item once, best score first. Items with equal scores keep the order in which they first appear across the lists.

truncate_and_normalize