Expand description
Lightweight task-aware embedding interfaces, independent of search and runtimes.
Modules§
- diagnostics
- Opt-in, content-free phase markers for external memory samplers. Set SEYK_MEMORY_DIAGNOSTICS=1 before process startup. No source content, token IDs, paths, errors or vectors are emitted.
- quant
- Offline Photo Anchor 4-bit codec for already normalized photo embeddings.
Structs§
- Dimension
- A nonzero embedding width, independent of any model’s supported dimensions.
- Embedding
- A vector with two explicit construction policies.
- Embedding
Request - One borrowed input and its embedding purpose.
- Model
Identity - Shared embedding-space configuration used by every modality.
Modality-specific weights/preprocessing additionally use
Embedder::input_identity. - Video
Frame - One decoded video frame (RGB8, row-major).
Enums§
- Embedding
Task - The purpose of an embedding. Each backend documents its supported mapping. These are embedding tasks, never instructions to generate answers or summaries.
- Error
- Embedding errors, with no search or inference-runtime types.
- Modality
- The kind of content a resource holds.
- Resource
Input - A borrowed input to embed. The request selects its task independently.
Traits§
- Embedder
- Synchronous, task-aware local embedding interface. Successful batches return one result per request, in input order. Async callers can use their executor’s blocking pool.
Functions§
- cosine_
similarity - Cosine similarity, including for vectors deserialized without normalization. Zero vectors have no cosine and return an error.
- dot_
product - Dot product. Unequal dimensions are errors, including in release builds.
- normalize
- L2-normalize in place using f64 accumulation to avoid f32 overflow/underflow. Zero (including empty) and non-finite inputs fail without modifying the input.