Seven RAG Patterns Beyond the Basic Chatbot and When Each One Fits
A field guide to naive, hybrid, corrective, self-reflective, multi-hop, graph and contextual RAG: what gap each one closes, its components, and how to choose.
This article is published in English.
Sources: Beyond the Chatbot: The Sophisticated Maturation of RAG Architectures
Retrieval-augmented generation has outgrown its experimental phase. Early on it meant dropping a PDF into a vector database and asking a chatbot for a summary. Now the industry is moving toward compound AI systems, in which retrieval is a layered cognitive architecture and not a single step. At enterprise scale the naive approach keeps failing on accuracy, latency and multi-step reasoning, and a set of specialized patterns has appeared in response. Here is a map of those patterns and a way to pick among them.
Why the designs keep multiplying
Basic RAG cuts documents into fixed chunks, embeds them, grabs the top k by cosine similarity and hands them to an LLM. That is adequate for simple FAQs. It breaks down on exact identifiers, long reports full of ambiguous pronouns, questions spanning several documents, and high-stakes domains where a hallucination is unacceptable.
Each newer variant targets one or more of these gaps:
- Poor recall on literal terms like policy IDs, part numbers or section references.
- Noisy or off-topic chunks getting through to the model.
- No fallback when the internal corpus has no answer.
- No self-check on whether retrieval or generation is grounded.
- No way to reason across several documents in multiple hops.
- Loss of document-level context inside small chunks.
The baseline: naive RAG
This pattern is the starting point, and it already reduces hallucination relative to a closed-book model by tying answers to retrieved text.
Documents are split into fixed-size chunks, say 512 tokens. A dense model such as bge-large or text-embedding-3-large embeds them, and the vectors go into a store like Qdrant, Weaviate or Pinecone. At query time the query embedding is compared with chunk embeddings by cosine similarity, and the top three to five chunks accompany the question to the LLM.
It suits single-topic internal knowledge bases, simple FAQ lookup, and prototypes where speed matters more than precision. The usual parts are fixed-size chunking with little or no overlap, one dense vector index, no reranker or a trivial one, and any instruction-tuned generator.
The weaknesses are predictable. Exact keyword queries and acronyms trip it up, chunk boundaries cost it context, and nothing detects bad retrieval before generation starts.
Hybrid search with a cross-encoder reranker
A single dense index misses too much once queries mix fuzzy meaning with literal identifiers like policy numbers, section IDs, SKUs or citations of regulations. Benchmarks show hybrid retrieval combined with neural reranking beating single-stage methods on mixed text and table documents.
Two retrievers run side by side. A sparse lexical one, BM25 or SPLADE, matches exact terms. A dense semantic one captures conceptual similarity. Their ranked lists are merged with Reciprocal Rank Fusion or a calibrated weighted scheme. The top fifty to one hundred fused candidates then go to a cross-encoder reranker, which scores each query-and-chunk pair jointly and returns a sharper top five to ten. Those chunks and the query reach the LLM along with citation metadata.
It fits regulated compliance documents where exact section references matter, technical manuals full of part numbers and error codes, financial documents with tables and precise numeric questions, and in general any production system where retrieval quality drives answer quality. Typical components include fixed or semantic chunking with metadata, BM25 or SPLADE alongside a dense bi-encoder, RRF or learned fusion, a cross-encoder such as bge-reranker, Cohere Rerank or Jina Rerank, and an instruction LLM that supports citations.
Corrective RAG
The aim here is to stop a system from answering confidently out of weak or irrelevant chunks. A quality gate follows retrieval, and a fallback route exists for when the internal corpus is not enough.
After retrieval, a lightweight evaluator scores the chunks against the query and emits a confidence label: correct, ambiguous or incorrect. A router then branches. Correct results go on to generation. Ambiguous ones trigger a query rewrite, a decomposition, or another retrieval strategy. Incorrect ones set off a web search or another external source, so the bad chunks are never used.
Customer support is a natural home, where missing information must not turn into invention. Frequently updated technical documentation with gaps is another. So is any domain where an honest "I don't know" with web-sourced context beats a confident wrong answer. The parts are a hybrid or dense retriever, a small classifier or prompted model that grades relevance, a rule-based or LLM-based router, a fallback such as a web search API or secondary store, and a generator told to cite sources and admit uncertainty.
Self-reflective RAG
This design moves retrieval and grounding checks into generation itself. The model learns when to retrieve, whether what it retrieved is relevant, and whether its own answer is backed by the evidence.
It works as a cyclic loop, driven by special tokens or structured prompts. The model decides whether the query needs retrieval at all, judges the relevance of retrieved documents, writes an answer conditioned on them, critiques that answer for grounding and completeness, and then either revises or declines to answer if support is thin. Training or prompting nudges it to favor grounded statements and flag unsupported claims.
Legal contract analysis and clause comparison fit. So do medical report summaries, clinical guidelines, audits and compliance reviews that must be traceable, and every setting where an unsupported claim is costly. Components include a hybrid or domain-specific retriever, a generator trained or prompted for self-critique, optionally an external verifier that checks citations against the retrieved chunks, and evaluation metrics for groundedness, faithfulness and abstention behavior.
Agentic multi-hop and sub-query RAG
Some questions cannot be answered in one retrieval pass: comparing metrics across companies and years, tracing dependencies across documents, or pulling together input from several departments.
An agent receives the complex query and breaks it into several sub-queries. Each runs against the retriever, in sequence or in parallel. The agent then reconciles contradictions, merges the evidence and writes a master answer, and it may loop again if intermediate results reveal gaps.
Typical uses include comparing 10-K filings with earnings calls, analyzing policy across documents, investigations that cut across departments of a large company, and research assistants that aggregate findings from many sources. The system needs a planner agent that decomposes queries, one or more indices with possibly different scopes, short-term memory for intermediate results, a synthesizer that merges sub-answers, and optionally a step that checks consistency across them.
GraphRAG
Flat chunk retrieval misses relationships that span several links, such as connecting suspicious accounts in anti-money-laundering work, following a clinical pathway, or understanding a complex supply chain.
An LLM or extraction pipeline finds entities and relationships in the documents. These go into a knowledge graph, in Neo4j or an in-memory structure. Queries then walk the graph to find multi-hop connections, for instance that a given entity treats a certain condition at a particular clinic. The retrieved subgraphs are combined with text chunks and passed to the LLM.
It belongs in fraud detection and financial crime compliance, clinical decision support, organizational or supply chain mapping, and any domain where the links between entities matter as much as the text. The building blocks are an entity-and-relation extractor, a graph store such as Neo4j, networkx or a managed graph database, a retriever combining traversal with optional vector search over node descriptions, and a generator that sees both text and graph context.
Contextual or late chunking
A chunk stripped of its surroundings can lose its meaning. A sentence reading "revenue increased by 15 percent" does not say which company, year or segment it describes.
The fix is to have an LLM prepend fifty to a hundred tokens of document-wide context to every chunk before embedding. That header can carry the document title, section, fiscal year, entity name and other key metadata. Each enriched chunk is embedded and indexed, so retrieved chunks arrive with their context and leave the generator less to guess.
It pays off with long PDF reports containing tables, figures and ambiguous pronouns, multi-year financial statements and regulatory filings, and technical standards with repeating section patterns. In short, any corpus where phrases like "the company," "this condition" or "they" show up without a clear local referent. The parts are a chunker, LLM-driven or rule-based, that adds the headers, a hybrid or dense retriever, an optional cross-encoder reranker, and a generator able to use the extra context tokens.
Choosing a pattern
Treat the following as a starting point, then adjust based on evaluation results.
- Simple internal search and prototypes: stay with naive RAG.
- Production systems, particularly with exact terms and tables: step up to hybrid search with cross-encoder reranking.
- Corpus with gaps, and wrong answers that cost real money: add the corrective layer.
- Legal, medical or audit work needing strict grounding: choose self-reflective RAG.
- Questions that span many documents and need several steps: use agentic multi-hop retrieval.
- Tasks that hinge on how entities relate: bring in GraphRAG.
- Long documents whose ambiguous references hurt quality: add contextual chunking.
Mature systems usually combine several of these. A strong, common stack in 2026 pairs hybrid retrieval with RRF fusion, cross-encoder reranking and contextual chunking, with a corrective fallback for low-confidence queries.