2.26.0

New features

• Custom score rerankers in global score modifiers for hybrid search (#1370). You can now rerank hybrid search results using BM25 or vector closeness scores (per-field or aggregate via sum/max/avg) by specifying marqo__score_* fields in scoreModifiers. Scores/aggregates are divide-by-max normalized. This is supported only for semi-structured indexes created with Marqo 2.26.0 or later.
• Opt-in lexicalOperand parameter for hybrid search (#1405). You can now explicitly control the lexical query operator (or, and, weakAnd) in hybrid search via the new lexicalOperand parameter in hybridParameters. The lexicalOperand can also be set in relevanceCutoffMethod to override it for the probe query. This is supported only for semi-structured indexes.
• matchAllTokens parameter for typeahead suggestions (#1385). A new matchAllTokens boolean parameter switches typeahead retrieval from OR to AND logic across tokens, requiring all query tokens to match. This improves relevance for multi-word queries.
• CONTAINS filter for semi-structured indexes (#1394). A new CONTAINS filter operator enables token-level matching on lexical fields. Supports bare values and grouped multi-word values, and is case-insensitive.
• _id IN filter support for semi-structured indexes (#1387). You can now filter by _id IN (id1, id2, ...) for semi-structured indexes, supporting up to 10,000 IDs by default (configurable via MARQO_MAX_IN_FILTER_IDS).
• Apply relevance cutoff to facets and trackTotalHits queries (#1388). When relevance cutoff is enabled with affectFacets: true, facets and totalHits now only count documents that pass the cutoff threshold, instead of all matching documents.

Bug fixes and minor changes

• Fix a crash when updating documents whose IDs contain special characters such as # or non-printable ASCII characters (#1407).
• Fix min_sort_candidates not being applied when trimming sort results (#1406).
• Fix bfloat16 indexes generating incorrect Vespa schema definitions by parameterizing the tensor numeric type (#1390).
• Fix documentation URLs for the restructured docs site (#1400).

Other changes

• Remove LanguageBind model support (#1382). LanguageBind models have been removed as they were never registered in the model registry.
• Remove NLTK from the API container and upgrade NLTK in the inference container (#1395).
• Move version to components/common as a single source of truth (#1391).