Crate khive_retrieval 

Hybrid search and ranking with deterministic scoring for khive.

Combines HNSW vector search, BM25 keyword search, and RRF fusion into a unified retrieval layer. All scores use DeterministicScore for cross-platform consistency. See docs/architecture.md for module layout, design principles, ID bridging strategies, and trait composition guide.

Re-exports

pub use error::ErrorKind;

pub use error::Result;

pub use error::RetrievalError;

pub use hit::HybridSearchOutcome;

pub use hit::RankScoreKind;

pub use hit::SearchHit;

pub use hit::SearchSignals;

pub use hit::SearchSource;

pub use hybrid::fuse_search_results;

pub use hybrid::fuse_search_results_checked;

pub use hybrid::DualIndexConfig;

pub use hybrid::DualIndexRouter;

pub use hybrid::DualIndexStrategy;

pub use hybrid::HybridConfig;

pub use hybrid::HybridSearcher;

pub use hybrid::IdentityReranker;

pub use hybrid::KeywordSearch;

pub use hybrid::Query;

pub use hybrid::Reranker;

pub use hybrid::VectorSearch;

pub use materialization::materialize_ranked_prefix;

pub use materialization::DropCounts;

pub use materialization::DropDiagnostic;

pub use materialization::DropReason;

pub use materialization::MaterializationDecision;

pub use materialization::MaterializationError;

pub use materialization::MaterializationLimitError;

pub use materialization::MaterializationLimits;

pub use materialization::MaterializedItem;

pub use materialization::MaterializedPrefix;

pub use materialization::RankedCandidate;

pub use materialization::MAX_MATERIALIZATION_CANDIDATES;

pub use materialization::MAX_MATERIALIZATION_DIAGNOSTICS;

pub use materialization::MAX_MATERIALIZATION_DROP_REASONS;

pub use materialization::MAX_MATERIALIZATION_LOADER_BATCH;

pub use materialization::MAX_MATERIALIZATION_OUTPUTS;

pub use metrics::MetricEvent;

pub use metrics::MetricValue;

pub use metrics::MetricsSink;

pub use metrics::NoopSink;

pub use metrics::RecordingSink;

pub use policy::filter_by_policy;

pub use policy::filter_by_predicate;

pub use policy::ClearanceLevel;

pub use policy::SearchPolicy;

pub use query_ir::FilterPredicate;

pub use query_ir::FuseStrategy;

pub use query_ir::QueryNode;

pub use query_ir::RerankMethod;

pub use search_config::SearchConfig;

pub use timeout::search_with_cancellation;

pub use timeout::search_with_deadline;

pub use timeout::search_with_optional_timeout;

pub use timeout::search_with_timeout;

Modules

embed

Re-exports from lattice-embed. Use these instead of depending on lattice-embed directly.

error

Error types for retrieval operations.

eval

Retrieval evaluation types and metrics.

hit

Search result types returned by hybrid retrieval pipelines.

hybrid

Unified hybrid search interface.

materialization

Policy-free materialization of a bounded ranked candidate prefix.

metrics

Observability hooks for retrieval indices.

policy

Policy integration for access-controlled retrieval.

query_ir

Query Intermediate Representation for the retrieval pipeline.

search_config

Per-call hybrid search configuration (fusion strategy + top_k). Default: RRF k=60, top_k=10.

timeout

Timeout and cancellation support for search operations.

Enums

FusionStrategy

Validated selection for combining ranked result lists.

Constants

DEFAULT_RRF_K

Default RRF constant k=60, standard in literature (Cormack et al., 2009).

Functions

fuse

Fuse ranked sources and retain at most top_k results.

normalize_weights

Lossily normalize finite positive weights; all-zero input becomes equal weights.

reciprocal_rank_fusion

Fuse score-descending sources by sum 1/(max(k, 1) + one_based_rank).

weighted_fusion

Fuse per-source min-max-normalized scores with lossy normalized weights.

weights_are_normalized

Returns true if positive weights sum to within tolerance of 1.0.