surfaceclassificationstance_llm
Classificationdimension · returns varchar

STANCE_LLM

LLM-backed stance detection on a topic (escape hatch for STANCE)

Per-row classifier — stable across GROUP BY.

classificationllmllm-escape-hatchtext

Arguments

nametypedescription
textVARCHAR
topicVARCHAR
stancesVARCHAR

About

LLM-backed alternative to STANCE for identifying an author's stance on a topic. Reach for this when the base NLI operator can't handle rhetorical nuance — sarcasm, hedging ("I mean, I guess it's fine…"), concessions, or texts that argue both sides before landing. Like STANCE, it requires a topic argument; unlike STANCE, it can read the full argument structure rather than scoring against a hypothesis template. For straightforward stance labeling at scale (editorial corpora, survey responses with a clear question), STANCE is fast and cheap. Use the LLM variant where the author is being clever.

Examples

LLM escape hatch detects stance

SELECT
  stance_llm (
    'I strongly support this initiative',
    'the initiative'
  )

Nearby rabbit holes

same domain
Climb back to The Looking Glass