Retrieval-Augmented Generation is no longer one pattern. Production systems combine lexical and semantic retrieval, metadata, rerankers, context-preserving indexing, graphs, corrective loops, and agents. The right design depends on what your baseline gets wrong.
A strong 2026 RAG design starts by proving that retrieval is needed, then separating retrieval quality from generation quality. Public recipes are useful starting points; your own query set should decide which pattern survives.
If the relevant corpus fits comfortably in context and cost/latency are acceptable, full-context prompting can be simpler than building retrieval infrastructure.
If the correct evidence never enters the candidate set, a larger generator or more elaborate prompt will not reliably repair the system.
Preserve document IDs, permissions, chunk lineage, scores, reranker outputs, and citations so failures can be diagnosed and reproduced.
Graphs, corrective loops, and agents add indexing or runtime cost. Introduce them for a measured failure mode, not because they sound more advanced.
The original page correctly separated ingestion from inference. The updated version makes the control points explicit: permissions, context enrichment, multiple retrieval channels, reranking, citations, and evaluation.
This replaces the old numeric bubble chart. The map is intentionally qualitative: it shows what kind of problem each family addresses without pretending that “complexity 8/10” is an objective measurement.
Prove whether retrieval is needed and whether simple dense retrieval is adequate.
Long-context baselineFlat vector RAGMetadata filtersUse when correct evidence exists but phrasing, exact terms, or ranking create misses.
Hybrid searchQuery transformationRerankingMulti-vector / late interactionUse when chunks lose document context or answers depend on hierarchy and relationships.
Contextual retrievalParent-childRAPTOR / hierarchicalGraphRAGMultimodal RAGUse when the system must decide whether/how to retrieve, recover from weak evidence, or select tools dynamically.
Corrective RAGSelf-reflective RAGAgentic retrievalFilter by where the pattern acts. Each card explains the failure it addresses, the main trade-off, and the evidence you should collect before adopting it.
Provide the relevant corpus directly when it fits the model context and latency/cost remain acceptable.
Task success, citation quality, latency and cost versus a retrieval baseline. Do not assume a bigger context window makes retrieval unnecessary.
Chunk documents, embed them, retrieve top-k nearest neighbors, then generate from the selected evidence.
Recall@k of answer-bearing chunks, context precision, answer groundedness, latency and cost.
Constrain search by tenant, source, product, date, document type, geography or permissions before/alongside semantic retrieval.
Filter accuracy, empty-result rate, recall within the filtered corpus and permission violations.
Fuse lexical retrieval such as BM25 with dense semantic retrieval so exact terms and conceptual similarity can both contribute.
Hybrid lexical + semantic retrieval is now a mainstream production pattern. Anthropic’s contextual retrieval recipe explicitly combines BM25 and embeddings before reranking.
Rewrite, expand, decompose or generate multiple search queries before retrieval.
Multi-query, decomposition, step-back questions and hypothetical-document-style query expansion are useful variants, but should be evaluated per query class.
Retrieve a broader candidate set quickly, then use a stronger relevance model to reorder and trim evidence before generation.
Modern rerankers can score lexical, semantic and semi-structured candidates; keep first-stage retrieval broad enough that the reranker has the right document to promote.
Prepend chunk-specific context: such as document identity, section, time period or surrounding meaning: before building dense and lexical indexes.
Anthropic reported substantially fewer retrieval failures from contextual embeddings + contextual BM25, with further gains from reranking. Treat those results as vendor experiments and reproduce on your corpus.
Index small child chunks for precise matching, then return a larger parent section or document window for generation.
Retrieval recall using children, answer quality using parent context, and context utilization rather than just token volume.
Build multi-level representations: chunks, clusters and summaries: so retrieval can operate at different levels of abstraction.
RAPTOR recursively clusters and summarizes chunks into a tree. More recent work continues to explore lower-cost and more update-friendly hierarchical variants.
Represent a document with multiple token- or region-level vectors rather than compressing all meaning into one vector.
ColBERT-style late interaction separates efficient document encoding from fine-grained query-document matching; multimodal descendants apply similar ideas to visually rich documents.
Extract entities, relationships and communities, then query graph structure plus source text for local or corpus-level reasoning.
The current project distinguishes Basic search, Local search for entity-centric questions, Global search over community reports, and DRIFT search that combines local exploration with broader community context.
Evaluate retrieved evidence and trigger corrective actions: such as alternate search, evidence refinement or fallback: when retrieval is weak.
The CRAG paper uses a retrieval evaluator to determine whether corrective retrieval actions should be triggered instead of blindly trusting top-k results.
Make retrieval conditional and critique evidence or generated output rather than retrieving a fixed number of passages for every question.
Self-RAG formalized adaptive retrieval and self-reflection rather than indiscriminately adding a fixed top-k context to every generation.
Let a policy or agent choose among vector search, SQL, graph, web, document/image retrieval or tools, then iterate when evidence is insufficient.
Agentic retrieval should be bounded by approved tools, budgets, maximum steps, evidence requirements and fallback behavior. “Agentic” is orchestration, not a substitute for good indexes.
These are qualitative design heuristics: not benchmark scores. Actual cost and latency depend on corpus size, models, indexes, hardware and query mix.
| Pattern | Primary failure addressed | Index-time impact | Query-time impact | Best evidence to collect |
|---|---|---|---|---|
| Flat vector RAG | Need external/current knowledge | Low | Low | Recall@k, groundedness, latency |
| Hybrid retrieval | Exact-term + semantic misses | Moderate | Moderate | Recall by query type, fusion sensitivity |
| Reranking | Noisy candidate ordering | Low | Moderate | nDCG/MRR uplift vs latency |
| Contextual retrieval | Chunks lose document context | High | Low | Retrieval failure rate, index cost |
| Parent-child | Precision vs context trade-off | Moderate | Low | Context precision + answer completeness |
| Hierarchical / RAPTOR | Multi-level / long-document reasoning | High | Moderate | Multi-hop QA, update/rebuild cost |
| GraphRAG | Entity, relationship and global-theme queries | High | High | Query-class quality vs indexing/query spend |
| Corrective / self-reflective | Weak evidence and recovery | Low | High | Recovery rate, false correction, added calls |
| Agentic retrieval | Heterogeneous tools and dynamic strategy | Variable | High | Task success, tool errors, steps, cost, safety |
A single “answer accuracy” score hides why RAG fails. Keep a versioned golden query set with known evidence and segment results by query type.
A RAG system is full of choices that should be measurable instead of buried in code. DataKnobs can make those choices explicit across creation, control and experimentation.
Build ingestion, parsing, chunking, embedding, indexing, retrieval, reranking and answer-generation workflows as reusable capabilities.
Govern source permissions, PII handling, lineage, freshness, citations, approved tools, budgets, human review and production evidence.
Experiment with chunk size, embedding model, hybrid weights, filters, top-k, reranker, thresholds, graph mode, routing and agent autonomy.
These references support the newer patterns and terminology added to the original guide.
There is no universal best architecture. Start with the simplest design that meets your target quality, latency, cost and governance requirements. Add patterns only for a measured failure mode.
Hybrid search is especially useful when exact identifiers and semantic meaning both matter. Reranking is useful when first-stage retrieval finds the right evidence but ranks too many irrelevant candidates above it.
Use it for query classes that depend on relationships, multi-hop paths, entity neighborhoods or corpus-wide themes. Compare its quality improvement against indexing and runtime cost on those specific query classes.
Not automatically. Long context can be a strong baseline for bounded corpora, but retrieval remains valuable for larger or changing corpora, permissions, freshness, citations and cost control.