Module embeddings

Enums§

CosineComparison: v0.7.0 H7 — dimension-aware outcome of a recall-time cosine comparison between a live query embedding and a stored embedding whose producing model may have changed since the row was written.
EmbedRole: Retrieval role of a text handed to the embedder. Drives the asymmetric task-instruction prefix for backends that require one (Ollama nomic-embed-text-v1.5); symmetric backends (the in-process candle MiniLM-L6-v2) ignore it. See #1520.
EmbedStatus: v0.7.0 F6 — outcome of a single embedding call. Returned by Embedder::embed_with_status alongside the (possibly absent) embedding vector.
Embedder: Semantic embedding engine supporting multiple backends.
EmbeddingFormatError: Errors produced by the embedding BLOB codec. Distinguishes the three failure modes operators want to triage independently:

EMBEDDING_DIM: Constant for backward compatibility — dimension of the default (MiniLM) embedding.
EMBEDDING_HEADER_BE_F32: Magic byte declaring “big-endian f32” payload follows. Reserved — the reader rejects this until v0.7 adds endianness conversion.
EMBEDDING_HEADER_LE_F32: Magic byte declaring “little-endian f32” payload follows.
EMBED_MAX_BYTES: v0.7.0 F6 — soft cap on the input size handed to the embedder. 64 KiB matches the F10 store-path threshold so a single content blob that the embedder can’t realistically process is reported as Skipped("content > 65536 bytes") rather than blowing up the chat/embed RPC. Operators who want larger embeddings can grow this constant alongside the F10 HTTP threshold.

Embed: v0.7.0 L0.7 — minimal dyn-compatible trait that abstracts “produces embedding vectors” away from the concrete Embedder enum.

decode_embedding_blob: Decode an embedding BLOB back into Vec<f32>.
decoded_dim: Number of f32 elements encoded in bytes, regardless of header presence. Used by the dim_violations stats path to compute per-row dim without allocating a Vec<f32>.
embedding_document: #1558 batch 5 wave 2 — the canonical embedding/rerank document template: "{title} {content}".
encode_embedding_blob: Encode a [f32] slice as a length-prefixed BLOB suitable for the memories.embedding column.
oversize_embed_reason: #1595 — single source of the EMBED_MAX_BYTES oversize check + its human-readable skip reason. Some(reason) when byte_len exceeds the cap, None otherwise. Shared by Embedder::embed_with_status (store path) and the backfill / reembed sweeps so the client-side guard and its WARN text can never drift between the write-time and batch paths.