CK-21 Reference-free grounding metrics for retrieval-augmented answers (Ragas; faithfulness, answer relevance and context relevance scoring)
Score whether retrieval finds relevant focused passages, whether the answer uses them faithfully and the quality of the answer, and produce a score per dimension without human reference answers.
What it does
Scores whether retrieval finds relevant focused passages, whether the answer uses them faithfully, and the quality of the answer, without human reference answers. Artefact: scores per dimension. Read each dimension separately.
When to use it
Use when a retrieval-augmented pipeline answers questions from a corpus and you need to separate retrieval problems from answer problems.
When it does not fit
Fits a pipeline with a retrieval step. Misfit: no retrieval, or an evaluator that shares the generator's errors. Reference-free scores are unvalidated against humans unless a human-coded sample is added.
Result format
The result gives a separate score for each dimension, covering retrieval of relevant focused passages, faithful use of those passages and the quality of the answer. It reports each dimension on its own line and does not combine them into one figure.
Requirements on the material
| Property | Requirement type | Requirement |
|---|---|---|
| Content type | Always applies | requires=text |
| Unit of analysis | Always applies | requires=claim |
| Appraises source credibility | Applies in some cases | no |
Not restricted on: Structure, Variable types, Time structure, How the sample was chosen, Level of detail available, Number of sources, Whether sources can be joined, Amount of material, Known data problems, Producer interest, Source dependence, Origin, Document genre, Language, Personal information, Security classification, Licence.
| Minimum material | a retrieval step with logged retrieved passages and a set of test questions |
|---|---|
| Additional sources needed | none |
| Not for (rules) | none |
| Not for (text) | Wrong choice when the pipeline has no retrieval step. Reference-free scores are unvalidated against humans unless a human-coded sample is added. |
Choosing and ordering
| Kind of technique | Verification technique, runs in step 6 |
|---|---|
| Phase tag | validate |
| Other phase tags | report |
| Kind of finding | gap; contradiction |
| Question answered | what |
| Strongest statement it supports | describe |
| Needs as input |
|
| Makes as output |
|
| Must exist first | None |
| Loop target | proceed with limit stated |
Library checks by stakes
Each stakes level includes the checks of the level below it. Checks that need a person apply at high stakes only, so the process leaves them out at lower stakes (decision 38). Repeatability: judgement dependent.
| Stakes | Check | Person involved | Counts as a different method |
|---|---|---|---|
| exploratory | spot check | optional | no |
| decision-informing | spot check | optional | no |
| decision-informing | human review sample | needed; left out below high stakes | no |
| decision-informing | source-span citation | none | no |
| high-stakes | spot check | optional | no |
| high-stakes | human review sample | needed | no |
| high-stakes | source-span citation | none | no |
| high-stakes | gold standard agreement | conditional | no |
| high-stakes | independent review | needed | no |
Failure modes and limits
| Model suitability | model plus code |
|---|---|
| Typical model failure | The judge model rates fluent answers as faithful. |
| Error sources | input data; language model |
| Certainty limits | Certainty drops when the scoring model shares the generator's errors (risk of bias) and when scores are not compared with human grades (indirectness). |
Related techniques
- should follow PE-13 Retrieval-augmented analysis (RAG (retrieval-augmented generation); grounded question answering over a corpus): retrieval-augmented answers are what is scored
- pairs with CK-22 LLM-as-judge with human calibration (model-graded evaluation): judge scores should be calibrated against human grades
- pairs with CK-17 Source-span and quotation check by code (quote-and-verify; extractive grounding; citation verification): span checks confirm citations in the answers
Evidence and tag basis
Documented. The abstract names the three dimensions and not the individual metrics (2023). (Es et al. 2023, arXiv 2309.15217)
Tag basis: input profile inference; selection inference; sequencing inference; robustness inference. All tag values are inference unless a source is named.