KickSide Search Index

kickside.search_index:index runs hybrid retrieval over tables that the calling module owns. The library owns no schema and no connection: the caller passes an open connection and a table spec.

Table spec

• vectors: Vector table: a vec0 virtual table on SQLite, a table with a vector(512) column on PostgreSQL.
• vector_key: Column in vectors that references the row id.
• vector_column: Vector column, default embedding.
• rows: Row table that holds the searchable content.
• row_key: Row id column.
• partition: Column present on both tables that scopes every search (a vec0 partition key on SQLite).
• columns: Row columns returned with each hit.
• fts / fts_key: SQLite fts5 table and its row-id column.
• tsvector: PostgreSQL tsvector column on rows.
• json_columns: Row columns decoded from JSON text.

Functions

• search(db, spec, vector, {partition, limit, filters, query_text}): dense search without query text; with query text, reciprocal-rank fusion of lexical and dense candidates.
• dense_search, lexical_search, lexical_query, fuse: the parts of search.
• filters is a list of {column, values, table = "rows" | "vectors"} equality or IN constraints.
• vector_literal, placeholders, insert_rows_tx, VECTOR_DIMENSIONS.

db is a pooled connection, not a transaction: on PostgreSQL a dense search runs in its own read transaction with hnsw.iterative_scan = strict_order, so an HNSW index keeps scanning until the partition and filters admit limit rows on pgvector 0.8 or later. Earlier versions retain their existing index scan behavior without receiving an unsupported setting. The setting is local to the read transaction and does not change the caller's session defaults.

Each hit carries id and similarity; fused hits also carry vector_similarity, vector_rank, lexical_rank and hybrid_score.