startthat.work home

PE-13 Retrieval-augmented analysis (RAG (retrieval-augmented generation); grounded question answering over a corpus)

Retrieve passages relevant to a question and have a model answer from them with citations, giving an answer with quoted passages and a claims register.

Prepare and extractEvidence: TestedPrimary technique, runs in step 5.3Strongest statement: describe

What it does

Retrieves passages relevant to a question and has a model answer from them, with citations. Artefact: an answer with quoted passages and references. Read the passages first and verify each cited claim.

When to use it

Use when you have a question that a few passages in a large corpus can answer and you need cited support.

When it does not fit

Fits questions that a few passages answer. Retrieval gives examples, not prevalence.

Result format

The result is an answer to the question made of cited claims, each with its quoted passage and reference. It lists the retrieved passages first and marks each cited claim as verified or not against its passage.

Requirements on the material

PropertyRequirement typeRequirement
StructureApplies in some casesrequires=unstructured; requires=semi-structured
Content typeAlways appliesrequires=text
Unit of analysisAlways appliesrequires=document; requires=topic
Appraises source credibilityApplies in some casesno

Not restricted on: 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 materialnone
Additional sources needednone
Not for (rules)none
Not for (text)The question asks for prevalence or counts across the corpus, since retrieval returns examples and not frequencies.

Choosing and ordering

Kind of techniquePrimary technique, runs in step 5.3
Phase taganalyse
Other phase tagsverify and prepare
Kind of findingpattern
Question answeredwhat
Strongest statement it supportsdescribe
Needs as input
Makes as output
Must exist firstNone
Loop targetproceed 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.

StakesCheckPerson involvedCounts as a different method
exploratoryspot checkoptionalno
decision-informingspot checkoptionalno
decision-informingsource-span citationnoneno
decision-informinghuman review sampleneeded; left out below high stakesno
decision-informingprompt variationnoneno
high-stakesspot checkoptionalno
high-stakessource-span citationnoneno
high-stakeshuman review sampleneededno
high-stakesprompt variationnoneno
high-stakesindependent reviewneededno
high-stakesprovenance lognoneno

Failure modes and limits

Model suitabilitymodel plus code
Typical model failureAttaches citations that do not support the sentence they follow, and treats retrieved examples as typical of the corpus.
Error sourcesinput data; language model
Certainty limitsCertainty drops because sentences are often not fully supported by their citations (risk of bias) and retrieval may miss relevant passages (reporting bias).

Evidence and tag basis

Tested. In generative search engines only 51.5 per cent of sentences were fully supported by their citations and 74.5 per cent of citations supported their sentence (2023). The original retrieval-augmented generator paper calls provenance for outputs an open problem. (Liu et al. 2023, verifiability in generative search engines; Lewis et al. 2020, arXiv 2005.11401)

Tag basis: input profile inference; selection inference; sequencing inference; robustness inference. All tag values are inference unless a source is named.