Nova Search Ranking (Design Spec)¶

This document defines how Nova meta is indexed and ranked to improve discoverability without increasing metadata maintenance overhead.

Goals¶

• Make canonical datasets and metrics surface first.
• Allow analysts to search by business terms (synonyms, use cases, domains).
• Support KPI discovery without creating duplicate per‑slice metric models.

Indexed Nova Fields¶

Nova meta is indexed into dedicated search fields:

Column‑level meta.nova.synonyms, semantic_type, and example_values are appended to the columns search field to improve column discovery (e.g., “UK” → country_code).

Default Boosts (Config)¶

These values are tuned for high signal with minimal noise:

DBT_NOVA_ALIAS_BOOST=18.0
DBT_NOVA_NAME_BOOST=12.0
DBT_NOVA_DESCRIPTION_BOOST=6.0
DBT_NOVA_COLUMN_BOOST=4.0
DBT_NOVA_TAG_BOOST=3.0
DBT_NOVA_PATH_BOOST=2.0
DBT_NOVA_CODE_BOOST=1.5

DBT_NOVA_META_SYNONYMS_BOOST=7.0
DBT_NOVA_META_MEASURES_BOOST=8.0
DBT_NOVA_META_METRIC_BOOST=10.0
DBT_NOVA_META_SENSITIVITY_BOOST=6.0
DBT_NOVA_META_PII_BOOST=8.0
DBT_NOVA_META_COMPLIANCE_BOOST=6.0
DBT_NOVA_META_DOMAINS_BOOST=4.0
DBT_NOVA_META_USE_CASES_BOOST=4.0
DBT_NOVA_SEARCH_STAGING_DEBOOST_FACTOR=0.6
DBT_NOVA_SEARCH_MEASURE_MATCH_MULTIPLIER=1.15
DBT_NOVA_SEARCH_METRIC_MATCH_MULTIPLIER=1.20
DBT_NOVA_SEARCH_SYNONYM_MATCH_MULTIPLIER=1.20
DBT_NOVA_SEARCH_CANONICAL_MATCH_MULTIPLIER=1.08
DBT_NOVA_SEARCH_CANONICAL_META_MATCH_MULTIPLIER=1.35
DBT_NOVA_SEARCH_CANONICAL_META_MATCH_BONUS=2.5
DBT_NOVA_SEARCH_ENGINEER_EXACT_MATCH_MULTIPLIER=2.0

Ranking Behavior¶

1) Exact business terms (synonyms/metric names) score close to name matches.
2) Measures are next‑highest (sessions, revenue) so KPI‑driven queries surface the canonical model even without a metric model. 3) Domains/use_cases provide intent routing but stay below synonyms/measures. 4) description, columns, tags, path, and code remain as backstops. 5) Staging-layer models are de‑boosted so curated models rank higher. This deboost is layer-rule driven (DBT_NOVA_LAYER_RULES_JSON) and applies when the resolved layer is staging, stage, or stg. 6) Nova meta re‑ranking boosts canonical models when query tokens match meta.nova.measures, meta.nova.metric(s), or meta.nova.synonyms. 7) Governance terms (sensitivity/pii/compliance) get elevated when present. 8) Analyst semantic readiness adds a bounded multiplier for entities with metric/measure definitions, grain/time-field metadata, and query-aligned dimensions. Non-semantic entities receive a slight de-boost. 9) Analyst near-tie hinting emits analysis_hints when top candidates are within a small score gap, prompting get_entity/get_context validation before SQL generation.

Scoring Pipeline (Lexical + Vector + RRF + Reranker)¶

Nova fuses multiple retrieval channels into a single ranked list:

1) Lexical search uses Tantivy BM25 with field boosts. 2) Vector (dense) and sparse retrievers produce their own ranked lists. 3) RRF fusion combines all retrievers into a single score. 4) Reranker (if enabled) rescales the top‑N results using a cross‑encoder.

Weighted RRF formula¶

For each retriever list, Nova assigns a score based on rank position:

rrf_score = weight / (k + rank + 1)

• weight is the persona‑specific weight for that retriever (e.g., vector, bm25, sparse).
• k is the RRF smoothing constant (DBT_NOVA_SEARCH_RRF_K).
• rank is 0‑based within that retriever list.

Final RRF scores are summed across retrievers to produce the fused ranking.

Reranker adjustments¶

If DBT_NOVA_SEARCH_ENABLE_RERANKER=true, Nova reranks the top‑N fused hits (DBT_NOVA_SEARCH_RERANK_TOP_N) using a cross‑encoder model. The reranker reorders the top‑N results while keeping the remainder of the list in fused order.

Persona Weights¶

Persona-specific weights (analyst/engineer/governance) are applied after base scoring. They amplify different signals without changing the underlying Nova metadata fields.

Current defaults (configurable via the persona weight env vars below):

If you need different weightings, override via env:

DBT_NOVA_SEARCH_PERSONA_ANALYST_WEIGHTS="vector=1.6,docs=1.3"
DBT_NOVA_SEARCH_PERSONA_ENGINEER_WEIGHTS="bm25=1.6,path=1.4"
DBT_NOVA_SEARCH_PERSONA_GOVERNANCE_WEIGHTS="tags=1.6,tests=1.5"
DBT_NOVA_SEARCH_PERSONA_DEFAULT_WEIGHTS="bm25=1.0"

Any weight not listed remains at its default.

Persona Resource-Type Priorities¶

After lexical/vector/sparse fusion, Nova applies persona-aware resource-type multipliers to reduce cross-persona noise:

• Analyst: boosts metric, semantic_model, saved_query, and curated model results; de-boosts test and macro.
• Engineer: boosts model, source, test, and macro; slightly de-boosts pure business artifacts (metric, semantic_model, saved_query).
• Governance: strongly boosts test, and also boosts model, source, and exposure; de-boosts macro.

These multipliers are deterministic defaults in code and complement persona retriever weights. The goal is to keep top-K results high-signal for each agent persona without requiring strict resource-type filters on every call.

Why This Avoids Metric Model Explosion¶

With nova_metric and nova_measures indexed:

• Query: “conversion rate” → matches metric(s).synonyms in canonical base model.
• Query: “sessions” → matches measure name/synonyms.

Analysts can then apply filters (web/app, country, device) at query time instead of having one metric model per slice.

When to Add Metric Models Anyway¶

Metric models are still recommended for:

• Cross‑model joins or complex logic.
• Executive or compliance KPIs that must be enforced centrally.

See Also¶

• Tools Reference - Search tool documentation
• Configuration Reference - Environment variables
• Search Defaults - Default search behavior
• Personas - Persona-specific search behavior