Search across one or more Machine Library indexes.

Benchmarks: merge a binary-only retrieval pool into the reranker pool of a documents-corpus search with L2 (default false).

Multi-query score combiner.

Rank decay applied by the combiner.

Temperature for the weighted combiner.

Number of sub-query ranks considered by the combiner.

Final-stage hosted cross-encoder reranker. On by default; pass false to skip it. No-op unless the reranker is configured (reranker.enabled). Runs after Summa candidate selection.

Override the number of candidates retrieved and reranked (0 = config default). Capped at MAX_CROSS_RERANK_POOL.

Collapse duplicate documents.

Benchmarks: force (true) or disable (false) merging a whole-document L2 pool into a passages search's reranker pool. Omit for the default, which merges for automatic passage searches.

Earliest issue time as a Unix timestamp.

Latest issue time as a Unix timestamp.

Restrict results to a publisher.

Extra weight for exact phrase matches inside the BM25 branch (default 1.0). Zero disables the bonus while keeping quoted phrase constraints.

Weight of the stemmed BM25 (full-text) channel in hybrid fusion; 0 disables it. Omit for the service default (1.0). Takes effect only on indexes whose schema carries the text fields.

Summa candidate heap multiplier.

Include raw document/passage features for formula-ranked returned hits. Training collection uses the separate complete-union export path.

Include exact organic RRF votes alongside fusion results.

Maximum number of hits to return.

Maximum number of reformulated query dimensions; zero uses the service default.

Result offset. offset + limit cannot exceed 500.

Optional sparse-vector pruning threshold.

Natural-language query (at most 16 KiB UTF-8). Wrap a span in double quotes ("float-zero determinants") to require ordered indexed tokens. Straight, curly and guillemet pairs work; all quoted spans are required and receive a lexical score bonus. Tokenizer normalization still applies. Unquoted words guide relevance; use structured parameters for filters. Quoted identifiers are text constraints, not direct lookup requests. An empty query can be combined with filters.

Relevance objective, independent of corpus and lexical/vector retrieval. Results remain document groups for both objectives.

Include enriched document metadata and query-matched snippets. Set to false for bulk ranking jobs that only need document IDs and scores.

RRF smoothing constant for hybrid fusion. Omit for the service default (60, the original-paper value); smaller values weight top-ranked results more aggressively.

Whether the sparse branches nominate candidates (false keeps them as ranking features only). Omit for the default: on for passage ranking, off for whole-document ranking.

Return bounded per-shard and per-vertical execution traces.

Minimum generated sub-query weight.

Relevance objective, independent of corpus and lexical/vector retrieval. Results remain document groups for both objectives.