RAG Retrieval

Definition

RAG retrieval selects source material at run time and places the relevant evidence into a model's working context. The architectural boundary includes query formation, retrieval, filtering, and context assembly; it does not by itself guarantee ranking quality, citations, or faithful generation. Harvey decomposes those stages, while GoDaddy shows that retrieval timing can be as important as retrieval itself. Harvey GoDaddy

Corpus Support

The analysis counts 304 deduplicated agentic systems (30.7%) across 15 industries. Evidence comprises 190 article, 193 transcript, and 56 summary quotes; 436 are stated and 3 implied. Corpus analysis

Problem It Solves

Models cannot reliably answer from private, changing, or domain-specific knowledge that is absent from their parameters. Retrieval supplies task-relevant records at decision time: Stripe grounds data questions in a catalog before generating SQL, and Deutsche Telekom routes knowledge questions through search and FAQ services. Stripe Deutsche Telekom

Applicability Signals

• Answers depend on private or frequently changing documents. Harvey
• A user query must be mapped to enterprise schemas or catalogs before execution. Stripe
• Personalization depends on retrieving user-specific records. Manus
• The source corpus is too large or noisy to place wholesale in the prompt. GoDaddy

Architecture and Variants

• Intent-gated retrieval: classify the request before fetching context, avoiding premature retrieval. GoDaddy
• Decomposed legal retrieval: rewrite queries, retrieve passages, generate an answer, and attach citations as distinct stages. Harvey
• Catalog-grounded execution: retrieve schema and metadata before text-to-SQL generation. Stripe
• Personalized vector search: scrape and normalize records, embed them, and search a user-specific collection. Manus

Why Teams Choose It

Retrieval can update knowledge without retraining, constrain answers to enterprise material, and expose the evidence used by downstream generation or evaluation. Sword additionally logs retrieval behavior so it can be evaluated as a component rather than hidden inside an end-to-end score. Sword Health

Tradeoffs

• Retrieving too early can lower accuracy by flooding the model with irrelevant context; GoDaddy delayed retrieval until intent was known. GoDaddy
• Pure document RAG may be insufficient for structured questions; Stripe pairs catalog grounding with executable data systems. Stripe
• More retrieval stages improve control but create additional latency and evaluation surfaces. Harvey evaluates rewrite, retrieval, generation, and citations separately. Harvey

Gotchas and Resolutions

• Raw documents consume tokens without guaranteeing relevance; use scoped chunks and task-aware filtering. GoDaddy
• A plausible answer can hide retrieval failure; Sword records retrieval traces and evaluates components. Sword Health
• Retrieval over the wrong schema produces confident but invalid actions; Stripe grounds generation in its data catalog. Stripe

When Not to Use It

Do not add retrieval when the answer is fully captured by a small stable prompt, or when the required operation is deterministic database access rather than document search. GoDaddy found that indiscriminate retrieval harmed performance, and Stripe uses retrieval to prepare execution rather than pretending retrieval alone answers structured-data questions. GoDaddy Stripe

Boundaries and Related Patterns

retrieval-ranking orders candidate evidence; hybrid-search combines lexical and semantic retrieval; RAG retrieval is the broader context-acquisition loop. source-citation-grounding is an unresolved vocabulary name: citations may follow retrieval, but a citation contract is not identical to finding context. Tool orchestration may invoke retrieval, while agent memory stores or recalls state across interactions. Vocabulary normalization

Operational Checklist

• Define when retrieval runs and what can bypass it. GoDaddy
• Trace query rewriting, candidates, ranks, selected context, and generation separately. Harvey
• Evaluate retrieval against domain-labeled examples, not only final-answer fluency. Sword Health
• Enforce access controls before retrieved content enters model context. Stripe

Representative Implementations

Methodology and Limitations

This page synthesizes a representative subset of 304 deduplicated systems. The corpus establishes recurrence, not that RAG improves every task; published cases also underreport failed retrieval designs. Corpus methodology