Answer Faithfulness

Available

rag.faithfulness

Measures how well the answer sticks to the retrieved sources, penalising claims the context does not support.

Contract

Field	Value
version	1.0.0
metric_type	pointwise
scorer_contract	per_row
direction	higher_is_better
entrypoint	aip_metrics_rag.unanswered.faithfulness
target_kind	None

Required columns

• input_id
• prompt
• sut_response
• retrieved_context

Accepted schemas

[
  {
    "name": "gdi_text_v1",
    "task_types": [
      "multi_turn_rag",
      "single_turn_rag"
    ]
  }
]

Methodology

1. An LLM judge (custom Ragas-based faithfulness metric) decomposes the answer into atomic claims.
2. A relevance filter drops claims that are not relevant to the question.
3. Each surviving claim is checked against the retrieved context with a natural-language-inference verdict of supported vs not-supported (a failing claim also carries a finer reason of unsupported, contradicted, wrong_entity, or overclaim).
4. The score is the mean of the per-claim verdicts, normalised to 0-1 (0.0 if no claims survive).

Score semantics

Scores range 0-1, the fraction of relevant answer claims that are supported by the retrieved context. A high score means every claim is grounded; a low score means the answer hallucinated or contradicted the sources. Higher is better.

Worked example

An answer yields 5 claims; 1 is filtered as irrelevant, and of the 4 judged, 3 are supported and 1 contradicts the context -> 3/4 = 0.75.

Configuration schema

{
  "properties": {
    "concurrency_limit": {
      "default": 10,
      "title": "Concurrency Limit",
      "type": "integer",
      "x-aip-param-role": "operational"
    },
    "inverted": {
      "default": false,
      "title": "Inverted",
      "type": "boolean",
      "x-aip-param-role": "scoring_metadata"
    },
    "lang": {
      "default": "en",
      "enum": [
        "en",
        "de",
        "fr"
      ],
      "title": "Lang",
      "type": "string"
    },
    "max": {
      "default": 1.0,
      "title": "Max",
      "type": "number",
      "x-aip-param-role": "scoring_metadata"
    },
    "metric_name": {
      "default": "faithfulness",
      "title": "Metric Name",
      "type": "string",
      "x-aip-param-role": "scoring_metadata"
    },
    "min": {
      "default": 0.0,
      "title": "Min",
      "type": "number",
      "x-aip-param-role": "scoring_metadata"
    },
    "model": {
      "anyOf": [
        {
          "type": "string"
        },
        {
          "type": "null"
        }
      ],
      "default": "gpt-4o",
      "title": "Model"
    },
    "system_type": {
      "const": "rag",
      "default": "rag",
      "title": "System Type",
      "type": "string",
      "x-aip-param-role": "judge_wiring"
    },
    "weight": {
      "default": 1.0,
      "title": "Weight",
      "type": "number",
      "x-aip-param-role": "scoring_metadata"
    }
  },
  "title": "InitializationConfig",
  "type": "object"
}

Execution

{
  "emits_metric_family": false,
  "function_name": null,
  "max_concurrency": 8,
  "processing_kind": "network",
  "stream_batch_size": 5000,
  "timeout_seconds": 21600
}

Complete manifest

accepts:
- name: gdi_text_v1
  task_types:
  - multi_turn_rag
  - single_turn_rag
config_schema:
  properties:
    concurrency_limit:
      default: 10
      title: Concurrency Limit
      type: integer
      x-aip-param-role: operational
    inverted:
      default: false
      title: Inverted
      type: boolean
      x-aip-param-role: scoring_metadata
    lang:
      default: en
      enum:
      - en
      - de
      - fr
      title: Lang
      type: string
    max:
      default: 1.0
      title: Max
      type: number
      x-aip-param-role: scoring_metadata
    metric_name:
      default: faithfulness
      title: Metric Name
      type: string
      x-aip-param-role: scoring_metadata
    min:
      default: 0.0
      title: Min
      type: number
      x-aip-param-role: scoring_metadata
    model:
      anyOf:
      - type: string
      - type: 'null'
      default: gpt-4o
      title: Model
    system_type:
      const: rag
      default: rag
      title: System Type
      type: string
      x-aip-param-role: judge_wiring
    weight:
      default: 1.0
      title: Weight
      type: number
      x-aip-param-role: scoring_metadata
  title: InitializationConfig
  type: object
dependencies: []
description: How well does the answer stick to the retrieved sources without unsupported
  claims?
direction: higher_is_better
display_name: Answer Faithfulness
entrypoint: aip_metrics_rag.unanswered.faithfulness
execution:
  emits_metric_family: false
  function_name: null
  max_concurrency: 8
  processing_kind: network
  stream_batch_size: 5000
  timeout_seconds: 21600
kind: metric
manifest_version: '1'
max_prompt_slots: 0
metric_metadata:
  methodology:
  - An LLM judge (custom Ragas-based faithfulness metric) decomposes the answer into
    atomic claims.
  - A relevance filter drops claims that are not relevant to the question.
  - Each surviving claim is checked against the retrieved context with a natural-language-inference
    verdict of supported vs not-supported (a failing claim also carries a finer reason
    of unsupported, contradicted, wrong_entity, or overclaim).
  - The score is the mean of the per-claim verdicts, normalised to 0-1 (0.0 if no
    claims survive).
  score_semantics: Scores range 0-1, the fraction of relevant answer claims that are
    supported by the retrieved context. A high score means every claim is grounded;
    a low score means the answer hallucinated or contradicted the sources. Higher
    is better.
  summary: Measures how well the answer sticks to the retrieved sources, penalising
    claims the context does not support.
  worked_example: An answer yields 5 claims; 1 is filtered as irrelevant, and of the
    4 judged, 3 are supported and 1 contradicts the context -> 3/4 = 0.75.
metric_type: pointwise
name: rag.faithfulness
partition_types: []
required_columns:
- input_id
- prompt
- sut_response
- retrieved_context
required_kinds: []
scorer_contract: per_row
target_kind: null
unsupported_trace_shapes: []
version: 1.0.0