Vector Search And Ranking

Vector Search And Ranking

Vector distance measures similarity in the configured embedding space. Ranking combines retrieval candidates; it does not decide truth, ownership, or effective support. Canonical hydration and lifecycle policy make those decisions.

Embeddings And Cosine Distance

Document and query vectors must use the same active embedding contract. The contract fixes provider/model, dimensions, normalization, and format compatibility. Search rejects incompatible dimensions or generations.

For nonzero vectors:

cosine_similarity(a, b) = dot(a, b) / (norm(a) * norm(b))
cosine_distance(a, b) = 1 - cosine_similarity(a, b)

Lower cosine distance means a closer vector match. It is not calibrated probability, assessor confidence, or an evidence-support score.

PostgreSQL calculates vector distance. Approximate candidate strategies and full-precision reranking are described in ANN, HNSW, And Halfvec.

Candidate Lists

Full-text retrieval supplies lexical matches. Vector retrieval supplies semantic-similarity candidates. Explicit Entity expansion can contribute directly connected evidence. These lists are authorized and bounded before fusion, then checked again during canonical hydration.

Evidence and Relationship discovery are separate frontiers. Community summaries and Hypothesis statements do not become vector evidence or change the primary evidence result budget.

Reciprocal Rank Fusion

The evidence adapter combines branch rank contributions:

score(id) = sum(weight / (60 + branch_rank))
full-text weight = 1
vector weight = 1
explicit Entity-expansion weight = 0.5

Known evidence is excluded. Equal evidence scores use the immutable evidence ID as the stable tie-breaker.

Relationship discovery uses equal weight for lexical and vector branches and groups equivalent authored Relationships by semantic identity. Its stable ordering and coverage checks prevent equivalent copies from filling the frontier.

For two branches with equal weights:

Candidate	Lexical rank	Vector rank	RRF score
A	1	4	0.032018
B	3	2	0.032002

A ranks slightly ahead of B. This describes query correlation, not stronger evidence.

Cross-Profile Conflict Reranking

Conflict policy governs which support paths remain eligible or preferred for the selected time view. It is separate from rank fusion. Authored disagreement must remain visible through the bounded conflict projection rather than being hidden by a source-count popularity score.

A resolved case can suppress a losing current support path while preserving that evidence through another unrelated eligible path. Current and temporal views apply the appropriate case/event state.

Memory-Space Fusion

Private-bound recalls search shared and permitted private evidence separately. The application fuses authorized branch results with equal-weight RRF constant 60. Equal scores prefer the private branch, then shared, then immutable ID.

This fusion does not permit cross-space Entity resolution or support. Current derived Relationship/community/Hypothesis expansion remains on shared data. See Recall Ranking.

Runtime Control

Ranking weights and fusion logic are implemented server policy. There is no public ranking-mode parameter, control-portal RRF toggle, or general mutable ranking-contract API. Search-index generations separately retain compatible candidate/index parameters.

A ranking change requires code review, regression evidence, and the applicable deterministic retrieval comparison. Observational recall feedback does not automatically train a model or update weights.

Degradation

Read search_states and degradations alongside results. A vector frontier may be unavailable while authorized lexical results remain useful. Required canonical hydration failures must remain errors.

Do not reconstruct a fact from stale search-document text, treat an empty candidate set as proof that no evidence exists, or claim full search quality from a degraded result.

Evaluation And Risks

Compare baseline and candidate using identical seed/suite identity and declared provider/configuration fingerprints. Measure retrieval quality and cost, including exact-vs-ANN candidate recall when the index strategy changes.

Failure mode	Mitigation
Similar wording is interpreted as entailed evidence	Keep ranking separate from support and lifecycle validation.
Approximation omits the right candidate	Compare with an exact scoped oracle; reranking cannot restore an omitted row.
Duplicate authored facts fill the response	Group by semantic identity and maintain cross-frontier coverage.
A new vector contract is compared with old vectors	Enforce contract/dimension/source fences.
Feedback changes live ranking silently	Keep feedback observational and require evaluated policy changes.