Retrieval contracts¶

search_vectors in exact mode selects and ranks with the same sequential f64 cosine used by document search. Stored embeddings remain normalized f32 values. Graph mode selects approximate candidates and then computes their canonical scores. Equal scores use document ID and chunk index order.

Sparse query results report Jaccard similarity for every route. Hopfield attention is available through query_hopfield; its weights are a separate retrieval contract. Index training can affect approximate candidate membership, but it does not change the meaning of similarity.

After train_document_index, broad metadata-filtered hybrid queries reuse the exact bounding-box tree and retain at most candidateLimit lexical owners before semantic fusion. The tree checks document filters before scoring a row. Small filtered sets use a direct scan. Both paths retain the filtered corpus's BM25 statistics and canonical scores. This changes the execution plan, not the search or filtering contract.

Filtered lexical queries and trained hybrid queries use filtered postings when their work is smaller than scanning every eligible chunk for every term. The captured eligibility is also used by dense traversal. Posting acceleration is capped at 524,288 store slots; the dense eligibility set is separately capped at that many matching documents. Beyond these limits the existing scan or predicate path applies. BM25 normalization belongs to the captured filter and generation, including legacy zero-frequency terms.

The trained exact tree remains usable after mutations through a bounded changed-document delta. Changed base documents are excluded before top-k selection, and current changed chunks participate in canonical scoring and fusion. The delta holds at most 128 document IDs, 1,024 live chunks and 64 KiB of ID bytes. Exceeding a bound drops the acceleration; queries retain their exact fallback and maintenance can rebuild. Replacing the same ID updates its delta entry. Explicit training publishes only at its captured generation and resets the delta. The trained status still means the base itself is current.

Dense training shares the existing vector-build admission gate and reports a retryable maintenance error when occupied. Queries add no background workers or queues. Server admission retains its bounded queue and busy response; cancellation of a waiting request releases its queue entry, while a running blocking query owns its permit until completion. These bounds describe additional search state, not a process RSS guarantee.

Phrase queries use persisted positions where available. For a candidate chunk without positions, the engine reconstructs them from retained text using that chunk's analyzer. An unrelated chunk without positions does not disable phrase search. A candidate whose phrase cannot be resolved because both positions and text are absent returns an error; it is never silently excluded. Persisting positions avoids reconstruction work.

Whole-document lexical ingest still requires positions, but it can also accept one dense embedding in the configured embedding space. This supports phrase matching across chunk boundaries and vector/hybrid retrieval of the same article. It does not create a sparse text hypervector in lexical ingest mode.

The standalone FilteredIndex accepts dimensions from 1 through 66,051 so every squared u8 Euclidean distance fits its u32 result. It accepts at most 1,048,576 labels in its dense label tables, validates training and CSR shapes, and bounds result capacity by the live candidate count. Use a compact label mapping for sparse external label identifiers.

Reasoning budgets¶

Use multi_hop_with_limits and synthesize_concepts_with_limits with writerslogic_warm::core::query_limits::QueryLimits to handle resource limits explicitly. Their defaults are 10,000,000 work units, 16,384 frontier entries and 64 MiB of accounted bytes. Work includes scanned triples, candidate visits, sorting and comparisons. Byte accounting conservatively covers retained inputs, intermediate frontiers and working buffers; it is not a process RSS measurement.

Budget exhaustion returns an error identifying the resource and limit. Results are complete or absent because of the error; there are no silently truncated answers. Multi-hop retrieval preserves distinct paths even when they reach the same endpoint. Concept synthesis generates one candidate frontier per seed instead of retaining every pair of neighbors. Common dimensions can still require quadratic work, so the work limit remains necessary.

The legacy Rust multi_hop and synthesize_concepts signatures retain their Vec return types and apply the default limits. They panic on budget exhaustion. Call the checked companions in applications that must report or recover from an exhausted budget.

Durable meaning governance¶

WarmCore::govern_memory returns Result<GovernanceReport> and uses the default QueryLimits; govern_memory_with_limits accepts explicit limits. It captures a coherent revision and bounded live snapshots, plans on isolated memories outside the mutation gate, and publishes one journal transaction only if the revision is still current. GovernanceConflict means another writer invalidated the plan; no governance change was committed, and callers may retry. Work, frontier, byte, operation-count and serialized-size exhaustion also fail before publication. The atomic transaction is capped at 4,096 operations and 32 MiB; byte accounting is a conservative allocation budget, not an RSS measurement.

Live triple owners survive consolidation even when their vectors are identical. Scoped atom/composite deletions preserve unrelated raw vectors with the same ID. Refinement recomposes affected owned composites and their matching raw-vector mirrors; ambiguous or missing ownership fails explicitly. Reopen and compaction retain these changes. The standalone MemoryGovernor primitive operates only on the supplied in-memory stores; applications persisting an WarmCore use its durable governance methods.

New stores use schema 5. Meaning-enabled stores with schemas 2–4 keep their writer format and must first be explicitly copied with warm-format upgrade <source> <destination> before durable governance. The source remains unchanged, and current arenas/checkpoints are copied without re-encoding their existing journal frames.