Retrieval, evidence and AI search
Trace the answer back to the context it used.
Retrieval augmented generation, or RAG, combines a search for relevant material with generated text. The retrieved context, the answer and the evidence supporting each claim are different things. Keep those distinctions visible so a fluent response is never mistaken for a verified source.
Free learning from Semantec SEO.
Separate retrieval, generation and provenance.
Retrieval selects candidate material for a question from an available collection. Generation produces a response using the supplied context and the model’s learned behavior. Provenance records where a passage came from, which version was used and how the answer relates to it. A system can retrieve a relevant passage and still produce an unsupported sentence.
Begin with the question and its scope: subject, date, audience and decision. Then identify the collection, access permissions and source versions. A retriever cannot find a record that is absent from the collection or inaccessible to its process. Ranking a passage highly does not certify its truth, currentness or fitness for the question.
A citation is an attribution link, not an automatic proof. It must lead to the right source and support the nearby claim under the same conditions. A correct source URL attached to an exaggerated sentence is still a failed evidence trace. Keep quoted source text distinct from an authored synthesis.
Understand the research without assuming a live product design.
Lewis and colleagues’ 2020 RAG research combined a learned retriever with a sequence generator and an external Wikipedia index. The paper compared approaches that conditioned on the same retrieved passages across the generated sequence with an approach that could use different documents as tokens were generated. Those are descriptions of the reported research models, not a specification for every current AI search product.
The related Dense Passage Retrieval research by Karpukhin and colleagues studied question and passage encoders for finding passages in open domain question answering. Retrieval and the downstream reader were distinct components. Their test results depend on the datasets, corpus, training and tests in that paper; they do not establish a ranking rule for a website today.
The RAG experiments used a particular dated Wikipedia collection and passage preparation. Do not turn that preparation into a universal recommended passage length. The work supports a useful distinction between stored model knowledge and externally retrieved material, while its discussion also recognizes that external sources can be biased or inaccurate.
Current providers may combine search, retrieval, generation, tools and citation presentation differently. Use current provider documentation for a named feature. Research architecture alone cannot reveal its hidden queries, selected corpus, ranking weights or exact citation mechanism.
Prepare context that preserves the decision.
Retain the source title, stable reference, version date, passage location, relevant text and scope. Include the conditions around a sentence: an exception in the next paragraph can reverse an apparent answer. Keep units, time periods, permissions and negation with the claim they qualify.
Compare conflicting records before synthesis. Check if one supersedes another, describes another product version or applies to another audience. If the conflict cannot be resolved from the available material, expose it and narrow the answer. Combining two incompatible passages into a confident compromise creates a new unsupported claim.
For each planned answer sentence, identify supporting passage references. Distinguish direct support, a clearly labelled inference and missing evidence. Material absent from the supplied context should remain unknown unless a separate, identified source establishes it. A request for a complete answer does not authorize invention.
Worked example: question to passages to answer.
The following example is entirely fictional and illustrates an evidence trace rather than a real system run. The question is: “Can a course participant move a booked session, and when must they ask?” The collection contains course policy CP-4, dated 1 October 2026, and course schedule CS-2 for the same fictional course.
Passage P1 from CP-4 states that participants may request one move to another listed session if they email the course team at least seven calendar days before their booked session. Passage P2 from the same policy states that a move needs team approval and an available place; a request does not itself change the booking. Passage P3 from CS-2 lists two sessions but gives no place availability.
The authored answer is: “You may request one move by emailing the course team at least seven calendar days before your booked session [P1]. The move requires team approval and an available place, so the booking is unchanged until approved [P2]. Two sessions are listed, but the supplied schedule does not establish available places [P3].” These bracket references identify the fictional records in the adjacent table.
The answer preserves the actor, permitted request, quantity of one, seven calendar days, approval and availability conditions. It does not promise a move, invent an email address or claim that a listed session has space. If P2 were omitted during retrieval, the draft would be incomplete even if its first sentence were accurate.
| Reference | Retained source passage | Answer use | Boundary |
|---|---|---|---|
| P1 · CP-4 | One move request; email at least seven calendar days before booked session | Permission, quantity and deadline | No promise of approval |
| P2 · CP-4 | Team approval and an available place required; booking unchanged until approval | Conditions and current booking state | No automatic move |
| P3 · CS-2 | Two sessions listed; no place availability supplied | Listed options with availability unknown | Do not infer spare places |
Check citations and factual limits before relying on the answer.
Open each cited record and compare the sentence with its passage. Check identity, version, date, scope, qualifications and link destination. Look for unsupported additions such as “free”, “guaranteed” or “any session”. Confirm that a cited page has not changed since the retained passage was collected.
Check the answer against the question as well as the sources. An accurate description of refund rules would not answer this rescheduling question. A source can be relevant without containing enough evidence for the requested decision. Record unresolved facts, the action needed to obtain them and the person responsible for review.
RAG does not guarantee truth or eliminate fabricated statements. Failures can arise from missing documents, poor retrieval, stale context, hostile instructions inside source material, misunderstood conditions or unsupported synthesis. Treat retrieved text as evidence to assess, not instructions that override the task or access policy.
Retain the question, retrieved passage identifiers, source versions, generated draft, citation mapping and review decision. This makes a later correction possible without pretending that an illustration, a research test or a citation count establishes factual correctness.
Sources and further reading.
The linked papers describe particular retrieval and generation research models. The examples and decision records here are editorial teaching material, not observations from a live audit.
Continue learning
Choose the next related task.
Use the passage retrieval and answer systems hub. Return to the semantic SEO learning overview to choose another subject.