Documentation
¶
Index ¶
Constants ¶
const ( ScaleBM25 = "bm25" ScaleRRF = "rrf" )
Score scales. BM25 magnitudes are raw and corpus-dependent; RRF scores are bounded near 1/rrf_k. They share a JSON key, so each result names its own.
const ( HighlightOpen = "\x02" HighlightClose = "\x03" )
Highlight markers wrap matched terms in FTS5 snippets. They are sentinels, not markup: the output layer turns them into ANSI or strips them. Using <b> here leaked HTML into JSON and into the terminal.
Variables ¶
This section is empty.
Functions ¶
This section is empty.
Types ¶
type BM25 ¶
type BM25 struct {
// contains filtered or unexported fields
}
BM25 runs a full-text search using SQLite FTS5's built-in BM25 ranking.
type Hybrid ¶
type Hybrid struct {
// contains filtered or unexported fields
}
Hybrid orchestrates BM25 + vector search + RRF fusion.
func NewHybrid ¶
func NewHybrid(bm25 *BM25, vector *VectorSearch, embedding providers.EmbeddingProvider, cfg config.SearchConfig) *Hybrid
type Result ¶
type Result struct {
DocID int64 `json:"doc_id"`
ChunkID int64 `json:"chunk_id"`
Collection string `json:"collection"`
Path string `json:"path"`
Title string `json:"title"`
HeadingPath string `json:"heading_path"`
Snippet string `json:"snippet"`
Timestamp string `json:"timestamp"`
Score float64 `json:"score"`
// Scale names the unit of Score. BM25 magnitudes and RRF scores differ by
// two orders of magnitude and are not comparable across commands.
Scale string `json:"scale"`
Explain *ScoreExplain `json:"explain,omitempty"`
}
Result is a single search hit.
func Finalize ¶
func Finalize(results []Result, opts SearchOpts) []Result
Finalize applies the ordering and trimming every command needs after retrieval: date sort over the whole candidate pool (never after truncation, or "newest" would only mean "newest of the most relevant"), duplicate and per-document capping, then the caller's limit.
type ScoreExplain ¶
type ScoreExplain struct {
BM25Score float64 `json:"bm25_score,omitempty"`
BM25Rank int `json:"bm25_rank,omitempty"`
VectorDist float64 `json:"vector_distance,omitempty"`
VectorRank int `json:"vector_rank,omitempty"`
RRFScore float64 `json:"rrf_score,omitempty"`
RerankScore float64 `json:"rerank_score,omitempty"`
}
ScoreExplain breaks down how a score was computed.
type SearchOpts ¶
type SearchOpts struct {
Query string
Collection string // empty = all collections
TopK int
Pool int // candidates to retrieve before dedupe/cap; defaults to TopK
Mode string // lexical | hybrid | deep
Explain bool
Since string // YYYY-MM-DD, inclusive lower bound on document timestamp
Until string // YYYY-MM-DD, inclusive upper bound
Sort string // "" = relevance, "date" = newest first
}
SearchOpts configures a search operation.
type VectorSearch ¶
type VectorSearch struct {
// contains filtered or unexported fields
}
VectorSearch performs KNN search using pure Go cosine similarity. Embeddings are loaded from the DB and compared in memory. For large corpora, a dedicated vector index (sqlite-vec, etc.) is preferred.
func NewVectorSearch ¶
func NewVectorSearch(database *db.DB, fingerprint string) *VectorSearch
func (*VectorSearch) Search ¶
func (v *VectorSearch) Search(ctx context.Context, queryEmbedding []float32, topK int, opts SearchOpts) ([]Result, error)
Search returns up to topK results nearest to the query embedding.
Source Files