surface↓ranking↓outliers

Rankingaggregate · returns varchar

OUTLIERS

Find unusual or atypical items via embeddings (+ optional criteria)

Per-group — reads the whole group in one call.

rankingrerankerspecialist-zoojson

Syntax

OUTLIERS({{ texts }})

OUTLIERS({{ texts }}, {{ num_outliers }})

OUTLIERS({{ texts }}, {{ num_outliers }}, '{{ criteria }}')

Arguments

name	type	description
texts	JSON	—
num_outliers	INTEGER	—
criteria(optional)	VARCHAR	—

About

Find outliers — items that don't fit the pattern — in a collection. Returns a JSON array of the N most-unusual items. Two modes depending on whether a `criteria` argument is provided: 1. **Criteria mode (recommended).** When you know what the collection is "supposed to be about", pass a criterion phrase and the cascade uses the cross-encoder reranker (bge-reranker-v2-m3) to score each item against the criterion. The N items with the LOWEST relevance scores are returned as outliers. This matches LLM-style categorical outlier detection cleanly: OUTLIERS(item, 1, 'a type of fruit') → finds 'chicken' in a list of fruits 2. **Unsupervised mode (no criteria).** When no criterion is given, the cascade falls back to nearest-neighbor distance on the embedding space: items whose closest neighbor is far away are "isolated" and ranked as outliers. This works well when items are naturally clustered but can be misled when the collection has no clear topical grouping — the embedding model may see lexically-distinctive items (e.g., "strawberry" among common fruits) as more outlier-ish than taxonomically-distinct items (e.g., "chicken"), because taxonomy isn't a geometric property. If your semantics depend on category membership, pass criteria. For LLM-style categorical outlier detection WITHOUT needing to spell out criteria (the old "just use world knowledge"), use OUTLIERS_LLM — see outliers_llm.cascade.yaml.

Examples

Criteria-guided — correctly identifies chicken as non-fruit outlier

WITH
  test_data AS (
    SELECT
      *
    FROM
      (
        VALUES
          ('apple'),
          ('banana'),
          ('orange'),
          ('grape'),
          ('strawberry'),
          ('watermelon'),
          ('chicken'),
          ('mango'),
          ('pineapple')
      ) AS t (item)
  )
SELECT
  OUTLIERS (item, 1, 'a type of fruit')
FROM
  test_data

Nearby rabbit holes

same domain

ABOUT

Returns 0.0-1.0 relevance score for text vs criterion (cross-encoder)

rankingrerankerspecialist-zootext

ABOUT_LLM

LLM-backed 0.0-1.0 relevance score (escape hatch for ABOUT/RELEVANCE TO)

rankingllmllm-escape-hatchtext

BEST

Pick the single best value from a group by a plain-English quality criterion

OUTLIERS_LLM

LLM-backed outlier detection (escape hatch for OUTLIERS)

rankingllmllm-escape-hatchjson

PAGERANK

PageRank centrality on an ad-hoc edge list (NetworkX)

rankingdeterministicjson

RANK_SEMANTIC

Rank a group of items by a subjective multi-factor criterion

Climb back to The Looking Glass

surface↓ranking↓outliers

Rankingaggregate · returns varchar

OUTLIERS

Find unusual or atypical items via embeddings (+ optional criteria)

Per-group — reads the whole group in one call.

rankingrerankerspecialist-zoojson

Syntax

OUTLIERS({{ texts }})

OUTLIERS({{ texts }}, {{ num_outliers }})

OUTLIERS({{ texts }}, {{ num_outliers }}, '{{ criteria }}')

Arguments

name	type	description
texts	JSON	—
num_outliers	INTEGER	—
criteria(optional)	VARCHAR	—

About

Find outliers — items that don't fit the pattern — in a collection. Returns a JSON array of the N most-unusual items. Two modes depending on whether a `criteria` argument is provided: 1. **Criteria mode (recommended).** When you know what the collection is "supposed to be about", pass a criterion phrase and the cascade uses the cross-encoder reranker (bge-reranker-v2-m3) to score each item against the criterion. The N items with the LOWEST relevance scores are returned as outliers. This matches LLM-style categorical outlier detection cleanly: OUTLIERS(item, 1, 'a type of fruit') → finds 'chicken' in a list of fruits 2. **Unsupervised mode (no criteria).** When no criterion is given, the cascade falls back to nearest-neighbor distance on the embedding space: items whose closest neighbor is far away are "isolated" and ranked as outliers. This works well when items are naturally clustered but can be misled when the collection has no clear topical grouping — the embedding model may see lexically-distinctive items (e.g., "strawberry" among common fruits) as more outlier-ish than taxonomically-distinct items (e.g., "chicken"), because taxonomy isn't a geometric property. If your semantics depend on category membership, pass criteria. For LLM-style categorical outlier detection WITHOUT needing to spell out criteria (the old "just use world knowledge"), use OUTLIERS_LLM — see outliers_llm.cascade.yaml.

Examples

Criteria-guided — correctly identifies chicken as non-fruit outlier

WITH
  test_data AS (
    SELECT
      *
    FROM
      (
        VALUES
          ('apple'),
          ('banana'),
          ('orange'),
          ('grape'),
          ('strawberry'),
          ('watermelon'),
          ('chicken'),
          ('mango'),
          ('pineapple')
      ) AS t (item)
  )
SELECT
  OUTLIERS (item, 1, 'a type of fruit')
FROM
  test_data

Nearby rabbit holes

same domain

ABOUT

Returns 0.0-1.0 relevance score for text vs criterion (cross-encoder)

rankingrerankerspecialist-zootext

ABOUT_LLM

LLM-backed 0.0-1.0 relevance score (escape hatch for ABOUT/RELEVANCE TO)

rankingllmllm-escape-hatchtext

BEST

Pick the single best value from a group by a plain-English quality criterion

OUTLIERS_LLM

LLM-backed outlier detection (escape hatch for OUTLIERS)

rankingllmllm-escape-hatchjson

PAGERANK

PageRank centrality on an ad-hoc edge list (NetworkX)

rankingdeterministicjson

RANK_SEMANTIC

Rank a group of items by a subjective multi-factor criterion

Climb back to The Looking Glass