sqlitevec

type Index struct {
	// contains filtered or unexported fields
}

type OptionFunc func(opts *Options)

type Options struct {
	// EmbeddingsModel identifies the embeddings model. It is recorded in the
	// index and opening an existing index with a different model is rejected
	// (the stored vectors would no longer be comparable).
	EmbeddingsModel string
	// VectorSize is the dimension of the vec0 embedding column (default
	// DefaultVectorSize). It is recorded in the index; opening an existing
	// index with a different size is rejected.
	VectorSize int
	// MaxWords bounds the size of the chunks sent to the embeddings model.
	MaxWords int
	// EmbeddingsConcurrency bounds how many embedding batches are computed in
	// parallel for a single document. Large documents split into many batches;
	// computing them concurrently cuts indexing latency at the cost of more
	// simultaneous requests to the embeddings endpoint. Defaults to
	// DefaultEmbeddingsConcurrency when <= 0.
	EmbeddingsConcurrency int
	// ReadPoolSize is the number of dedicated read connections opened by
	// NewIndexAtPath (defaults to DefaultReadPoolSize when <= 0). Ignored by
	// NewIndex, which uses the single caller-owned connection.
	ReadPoolSize int
	// CoarseQuantization enables the two-stage search: a fast KNN on the
	// binary-quantized vectors (Hamming distance) preselects k×8 candidates,
	// re-scored with the full float vectors. ~30× faster scans at a marginal
	// quality cost, worthwhile on large corpora. Requires a vector size
	// divisible by 8.
	CoarseQuantization bool
}

type SnapshottedMetadata struct {
	Model string
}

type SnapshottedRecord struct {
	Source      string
	SectionID   string
	ChunkIndex  int
	Embeddings  []byte
	Collections []string
	// Metadata is the document's metadata JSON (schema v3), repeated on each of
	// its chunks. Snapshots taken before v3 decode it as "", which restores a
	// document without metadata — the same state an index had back then.
	Metadata string
}