Information Retrieval Find the right evidence before generating anything.
Information retrieval is the process of locating items from a searchable collection that are useful for an information need. In AI search, those items can include documents, passages, entities, structured records, images, product data or other evidence units.
Retrieval is not generation. A system can retrieve without generating, and a generative system can fail because retrieval returned the wrong candidates, stale evidence, redundant sources or passages that are relevant to the topic but insufficient for the claim.
Retrieval answers one question first: what evidence should enter the candidate set?
Before synthesis, an information system must decide what to search, what representation to compare, which constraints to apply, how many candidates to retain and which evidence units deserve deeper scoring.
Locate relevant candidate information.
Retrieval maps an information need to candidate items in a corpus. The result can be a ranked list, a filtered set, passages, records or another evidence collection.
- corpus / index defines searchable inventory
- query representation expresses the need
- candidate generation prioritizes recall
- reranking can improve ordering and precision
Use retrieved context to construct output.
Generation happens after or alongside retrieval in RAG-style systems. It can compare, summarize or synthesize retrieved material, but it cannot repair missing evidence merely by producing fluent prose.
- context assembly
- claim construction
- reasoning / synthesis
- citation or source-path generation
Change the retrieval strategy. Change what enters the evidence set.
Select a mode. The same information need produces different candidate behavior under lexical, semantic, hybrid, filtered, passage-level and freshness-sensitive retrieval.
“Relevant” can be computed through different representations.
Sparse, dense and hybrid retrieval are not interchangeable labels for quality. Each representation creates different strengths, blind spots and failure signatures.
Lexical retrieval
Represents queries and documents through terms or weighted token features. Exact language, rare identifiers and technical terms can be highly discriminative.
q: "rel=canonical migration"
strong signal: shared rare terms
STRENGTH / EXACT LANGUAGE
Semantic retrieval
Represents items in a learned vector space so conceptually similar language can be close even without high token overlap. Quality depends heavily on representation and corpus.
q: "pages disappeared after redesign"
candidate: "indexation loss after site move"
STRENGTH / PARAPHRASE
Hybrid retrieval
Combines lexical and semantic candidate evidence. Useful when exact terms matter but conceptual recall is also necessary.
score* = lexical + semantic + filters
*illustrative editorial formula
STRENGTH / BALANCED RECALL
Constraint-aware retrieval
Applies structured conditions such as geography, language, date, product model, permissions, user eligibility or content type.
country=HR AND updated≥2026
model="X120"
STRENGTH / APPLICABILITY
Passage retrieval
Searches at a smaller evidence unit than the whole document. A long page may contain one passage directly capable of answering the subproblem.
document → section → paragraph
retrieve focused evidence unit
STRENGTH / LOCAL SPECIFICITY
Cross-modal retrieval
Matches information needs across text, images, audio, video, product attributes or other representations when the task is not text-only.
image object + text qualifier
→ manual / catalog / compatible part
STRENGTH / NON-TEXT EVIDENCE
First find enough plausible evidence. Then spend precision where it matters.
Many retrieval systems separate broad candidate generation from a more expensive reranking or scoring stage. This is a useful architecture because recall and precision create different optimization pressures.
10M searchable items
Documents, passages, products, records or other units available to the retrieval system.
SEARCH SPACETop 1,000
Fast retrieval aims to avoid missing useful evidence even if the set still contains noise.
RECALL FIRSTTop 50
Deeper scoring can inspect richer query-document relationships, constraints or learned relevance features.
PRECISION UPTop 5–12
Only a limited evidence set may enter downstream synthesis or claim construction.
CONTEXT BUDGETRetrieval quality is not simply “did we find something relevant?”
Recall asks whether useful items were found. Precision asks how much of the retrieved set is actually useful. Real systems balance these goals against latency, cost, context limits and task risk.
Did the candidate set miss critical evidence?
A diagnostic retrieval set that finds 9 broad SEO guides but misses the one migration-specific canonical document may have poor task recall despite looking topically rich.
How much retrieved material is actually useful?
A large candidate set can have high recall but low precision if most items are generic, duplicated, stale or only loosely related to the user’s evidence need.
Every retriever has a blind spot. Design around the failure, not the buzzword.
The matrix treats retrieval mode as a tool selection problem: what information signal is strongest, what can be missed, what filters matter and what should happen before synthesis.
A relevant document can contain one decisive paragraph.
Document-level relevance and passage-level evidence are different. A long page can be broadly relevant while only one section actually supports the answer claim.
Site Migration Documentation
A long technical document contains sections on redirects, canonicals, sitemaps, analytics, DNS and monitoring.
Canonical transfer after URL migration
The selected section specifically explains how canonical targets should move from old URLs to their corresponding new destinations.
Claim-ready context
“Canonical targets still point to the retired URL set” is now a testable diagnosis tied to a specific evidence unit rather than inferred from generic migration advice.
Same machinery. Different candidate contracts.
Retrieval strategy should follow the evidence requirement. Factual lookup, diagnosis, current commercial comparison and multimodal identification do not want the same corpus, filters or ranking emphasis.
Knowledge Graph launch date
One historical fact with a strong preference for a primary or authoritative source.
Pages disappeared after migration
Needs broad recall across several technical failure classes, then aggressive reranking by observed symptoms.
CRM under a fixed budget
Time-sensitive product data must survive seat-count, nonprofit and budget constraints.
Identify a replacement part
Visual evidence must resolve to an entity before text retrieval can reliably find compatibility data.
Generation often gets blamed for retrieval failures upstream.
When an answer is wrong, the root cause can be missing corpus coverage, bad indexing, poor query representation, weak candidate generation, incorrect filtering, stale evidence or reranking failure.
The correct source is not in the searchable collection, so no scoring method can retrieve it.
EXPAND / FIX CORPUSThe information exists but is absent, stale or incorrectly represented in the searchable index.
REINDEXLexical retrieval misses a useful document because query and source use different terminology.
SEMANTIC / EXPANSIONDense retrieval finds conceptually similar content that is wrong for the user’s specific entity, constraint or task.
FILTER / RERANKLocation, model, language, date, permissions or user constraints are dropped during retrieval.
PRESERVE STATEAn old but highly relevant document outranks current evidence for a state-sensitive claim.
TIME-AWARE RETRIEVALGood candidates are retrieved, but deeper scoring promotes generic or weak evidence above specific support.
RECALIBRATEA sentence appears relevant when isolated but means something different in the surrounding section or document.
RESTORE CONTEXTToo many redundant or weak candidates consume the synthesis context and reduce signal density.
PRUNE EVIDENCEUse IR vocabulary precisely. Do not invent a search-engine internals story.
These controls separate general information-retrieval concepts from proprietary production systems and public Google documentation.
Illustrative similarity values or candidate order on this page are not public Google Search metrics.
MODEL ≠ PRIVATE SCOREExact identifiers, names, rare terminology and structured filters can make sparse signals extremely valuable.
MODE DEPENDS ON TASKA candidate can be conceptually close while being stale, inapplicable, contradictory or wrong for the specific entity.
SIMILAR ≠ TRUERetrieval provides context; generation uses that context. The quality of either stage can independently fail.
RETRIEVE ≠ GENERATESmaller units improve focus but can remove qualifiers, definitions or exceptions located outside the selected passage.
LOCAL FIT / GLOBAL CONTEXTReturning more candidates may increase recall while reducing precision and increasing downstream context noise.
K IS A TRADE-OFFAny claimed private vector space, score cutoff or AI retrieval weight should be treated as unsupported unless publicly documented.
NO INTERNAL VISIBILITYA page can be indexed and eligible for Search yet never be selected as supporting evidence for a given AI response.
ELIGIBLE ≠ SELECTEDGoogle’s current guidance keeps AI-feature eligibility inside normal Search technical requirements and visible content practices.
FOUNDATIONS REMAINGround Google-specific claims in what Google actually publishes.
Google’s public guidance confirms that generative AI Search features retrieve relevant, up-to-date pages from the Search index and can use query fan-out. It does not disclose the private scoring architecture behind candidate selection.
RAG depends on retrieval from the Search index
Google describes retrieval-augmented generation in its Search AI features as relying on core Search ranking systems to retrieve relevant, up-to-date web pages, then using specific information from those pages to generate responses with links.
Supporting links still depend on Search eligibility
Google states that pages shown as supporting links in AI Overviews or AI Mode must be indexed and eligible to appear in Search with a snippet. There are no additional published technical requirements specifically for those AI features.
Audit the candidate set. Then audit what deserved to survive.
This framework separates corpus coverage, indexing, query representation, retrieval mode, filters, recall, precision, reranking, freshness, passage context and downstream evidence quality.
AI Search research. From retrieval foundations to agentic systems.
Continue through RAG, grounding, source selection, passage retrieval, optimization, entity visibility and multimodal interfaces while keeping retrieval mechanics distinct from generation.
What Is AI Search?
Define the retrieval, evidence and synthesis architecture behind AI-oriented search interfaces.
OPEN NODE → AIS / 02 · QUERIESQuery Fan-Out
Decompose complex requests into distinct retrieval branches and evidence responsibilities.
OPEN NODE →Information Retrieval
Locate candidate documents, passages, records and evidence before downstream synthesis.
CURRENT NODERetrieval-Augmented Generation
Use retrieved information as context for model-generated responses.
OPEN NODE → AIS / 05 · EVIDENCEGrounding
Connect generated claims to information capable of supporting or constraining them.
OPEN NODE → AIS / 06 · SOURCESCitations & Source Selection
Evaluate directness, freshness, specificity, provenance and claim support.
OPEN NODE → AIS / 07 · PASSAGESPassage Retrieval
Design focused information units that can resolve precise evidence needs.
OPEN NODE → AIS / 08 · OPTIMIZATIONAI Search Optimization
Build technically accessible, differentiated and evidence-rich information systems.
OPEN NODE → AIS / 09 · ENTITIESEntity Visibility in AI Search
How explicit identity, attributes and relationships support machine interpretation.
OPEN NODE → AIS / 10 · FRONTIERMultimodal & Agentic Search
Search systems that combine language, images, context, tools and task execution.
OPEN NODE →CORPUS / CANDIDATES / RERANK / EVIDENCE
Generation cannot rescue evidence that retrieval never found.
Information retrieval determines which parts of the corpus become available to downstream reasoning. Strong systems therefore treat corpus coverage, candidate recall, ranking precision, filters, freshness, passage context and provenance as separate engineering problems rather than collapsing everything into one vague idea of “relevance.”