Skip to main content

Grounding

Grounding ties every extraction back to the exact span of source text it came from. It is what lets you trust (and verify) what the model produced, rather than taking a JSON blob on faith. This page explains the concept; to act on it (filtering, slicing, highlighting), see Work with results.

char_interval: where an extraction lives​

Each Extraction carries a char_interval that records where LangExtract found its text in the source. The interval is half-open: start_pos is inclusive and end_pos is exclusive, so source_text[start_pos:end_pos] returns the matched substring.

A set interval means grounded; None means ungrounded​

The most important thing char_interval tells you is whether an extraction is real:

  • Set: LangExtract located the text in the source, and the interval is its exact position.
  • None: the model produced a value that does not appear in the source. This is the signal for a likely hallucination, or for content the model copied from your examples rather than the input.

Treating a None interval as "unverified" is the core safety habit when working with extractions.

Exact and fuzzy alignment​

When an extraction is grounded, alignment_status records how LangExtract matched its text to the source:

  • MATCH_EXACT: the extraction text matched the source character for character.
  • MATCH_FUZZY: it matched approximately, within a configurable similarity threshold, which absorbs small differences such as whitespace or punctuation.
  • MATCH_GREATER / MATCH_LESSER: it matched a span larger or smaller than the extraction text.

LangExtract prefers exact matching; fuzzy alignment lets grounding survive the minor reformatting a model sometimes applies. The thresholds and algorithm that govern fuzzy matching are tuning parameters, documented in the API reference.

See also​