Claim Extraction

Extracts atomic claims from a candidate answer using a chat-completion model, writing them to output_field. The question and answer inputs are read from the source columns configured in columns. This is the first step of the claim-based RAG pipeline; the extracted claims are consumed by the Claim Labeling and Link Claim Quotes synthesizers.

Output

One output sample per source sample with the extracted claims written to output_field.

Examples

Example: Claim extraction. Extracts atomic claims from each candidate answer into the claims field.

Claim Extraction Synthesizer
# ...
config_spec:
  - key: "synthesizer_model_key"
    type: "model"
    display_name: "Synthesizer Model"
    description: "Model used to extract claims from question-answer pairs."
definition:
  # ...
  synthesizers:
    - type: claim_extraction
      model_key: "<< config.synthesizer_model_key >>"
      columns:
        question: question
        answer: expected_answer

With the columns bindings above, the question is read from the question column and the candidate answer from the expected_answer column:

Source Sample
{
  "question": "What is the capital of France?",
  "expected_answer": "The capital of France is Paris. Paris is well known for its Eiffel Tower."
}

The synthesizer produces one output sample, which keeps the source columns and adds the extracted claims under claims - one (subject, predicate, object) triplet per fact found in the answer:

Output Sample
{
  "question": "What is the capital of France?",
  "expected_answer": "The capital of France is Paris. Paris is well known for its Eiffel Tower.",
  "claims": [
    {"claim": {"subject": "Paris", "predicate": "is the capital of", "object": "France"}},
    {"claim": {"subject": "Paris", "predicate": "is well known for its", "object": "Eiffel Tower"}}
  ]
}
Note

Set output_field to write the claims to a different column - the Claim Labeling and Link Claim Quotes synthesizers then read that column through their own columns bindings.

The claims are extracted by a model, so the exact triplets vary between runs.

Configuration

Properties

type Literal “claim_extraction” required

The type of the synthesizer.

model_key TemplateValue required

The key of the chat completion model to be used as a synthesizer.

output_field string, TemplateValue

Name of the structured output field written to each generated sample.

Default: claims

instructions string, TemplateValue

Optional domain-specific instructions appended to the built-in system prompt.

Default: None

columns ClaimExtractionColumnBindings

Mapping from built-in inputs to source sample columns.