Hyphae documentation

Exact vectors, lexical text, and hybrid fusion.

Hyphae 0.2.0 makes durable multimodal retrieval part of the autonomous engine while leaving vector generation and query-text generation under caller control.

Retrieval is part of durable format 2

Named vector spaces, signed-Q15 vectors, and lexical-index definitions survive restart, snapshot, compaction, backup, restore, migration, and derived-index rebuild.

  • Vector-space dimension is immutable.
  • Lexical field definitions and weights are canonical.
  • Mutation UUIDs keep vector retries durably idempotent.

Exact vector ranking is deterministic

Every durable vector in one named space is scored with canonical integer cosine nanos. Results sort by score descending, then binary object key ascending.

  • Semantics identifier: hyphae-exact-reference-semantics-v2.
  • Signed Q15 avoids platform-dependent floating-point ranking.
  • Minimum score and top-margin policies can return typed abstention.
  • Candidate, decoded-byte, result, and time budgets return no partial ranking.

Lexical retrieval needs no provider

Pinned NFKC normalization, Unicode default case folding, Unicode-alphanumeric tokenization, and BM25F-compatible integer scoring make the lexical path reproducible without an embedding service.

  • The query and configured document fields use the same tokenizer.
  • Field weights and scoring constants are versioned.
  • Missing, non-string, and empty fields have explicit behavior.

Hybrid fusion is explainable

Hyphae executes lexical semantics v1 and durable exact semantics v2 at the same snapshot checkpoint, then fuses one-based ranks with integer reciprocal-rank fusion.

  • The RRF constant is fixed at 60.
  • Each result records modality ranks, scores, contributions, fusion score, and final rank.
  • Branch errors fail the complete request instead of silently degrading.

Every retrieval mode can prove its outcome

Separate exact, lexical, and hybrid hyphae-retrieval-proof-v1 .hyrproof payloads bind canonical request, outcome, semantics, and format-2 snapshot witness for complete offline reexecution.

  • The caller obtains the expected anchor from a trusted channel.
  • Proof verification detects request, result, explanation, semantics, witness, and anchor tampering.
  • A self-consistent proof is evidence, not identity or authorization.
consolecopy-safe example
hyphae verify-retrieval --kind exact \
  --proof result.hyrproof \
  --snapshot witness.hysnap \
  --anchor <expected-anchor>

Know the deliberate boundary

Hyphae 0.2.0 does not ship approximate nearest-neighbor indexing or an embedding model. It provides exact durable vectors, deterministic lexical retrieval, and deterministic hybrid fusion.

  • Applications choose whether and how to create vectors.
  • Hyphae never calls an embedding provider.
  • GET /v1/capabilities publishes effective retrieval limits.
Next guideResult proofs