12 Advanced RAG Techniques Beyond Naive Retrieval [2026]

Emily Winks

Data Governance Expert

Updated:05/18/2026

Published:05/18/2026

22 min read

Watch Context Layer Live Get the Context Layer Ebook

Key takeaways

SOTA RAG scores 63% factual accuracy; straightforward RAG without advanced techniques scores just 44%.
Hybrid retrieval + sentence window chunking is the highest-ROI starting point for most production systems.
ARAGOG found Cohere Rerank showed no notable advantage over naive RAG; benchmark on your own corpus.
Data quality bounds retrieval quality: governed metadata improves AI agent SQL accuracy by 38% (Atlan research).

Quick Answer: What Are Advanced RAG Techniques?

A 2024 comprehensive RAG benchmark found that state-of-the-art RAG systems answer only 63% of factual questions correctly, while straightforward retrieval without advanced techniques scores just 44%. Advanced RAG techniques close this gap through smarter retrieval, better chunking, and self-correcting generation. This guide covers 12 proven techniques, from hybrid retrieval and reranking to GraphRAG and RAPTOR, with real benchmarks, complexity ratings, and a decision framework for choosing the right approach for your stack.

Core components

State-of-the-art RAG answers only 63% of factual questions correctly; straightforward RAG without advanced techniques scores just 44% (CRAG Benchmark, 2024).
Contextual Retrieval reduces retrieval failure rates by up to 67% -- not hallucination in general, specifically retrieval failures (Anthropic, 2024).
Data quality is the upstream lever: Atlan research shows governed metadata improves AI agent SQL accuracy by 38%.

Is your data estate AI-agent ready?

Assess Your Readiness

Advanced RAG techniques (including hybrid retrieval, cross-encoder reranking, Self-RAG, RAPTOR, and Contextual Retrieval) address a measurable accuracy ceiling. A 2024 comprehensive RAG benchmark found that state-of-the-art RAG systems answer only 63% of factual questions correctly, while straightforward RAG without advanced optimizations scores just 44%. The 12 techniques in this guide span beginner-friendly additions (hybrid search, sentence window chunking) to architectural upgrades (CRAG, GraphRAG, Adaptive RAG), each with benchmarks, complexity ratings, and framework support.

What you’ll find in this guide:

Naive RAG vs. advanced RAG: what the accuracy gap looks like in practice and why it matters
A comparison table for all 12 techniques: what each does, best use case, accuracy gain, and complexity rating
Per-technique breakdowns with real benchmark numbers and arXiv citations
A decision framework for choosing the right technique for your pipeline
The upstream factor most teams miss: data governance as a prerequisite for retrieval quality

Inside Atlan AI Labs and The 5x Accuracy Factor

Learn how context engineering drove 5x AI accuracy in real customer systems. Explore real experiments, quantifiable results, and a repeatable playbook for closing the gap between AI demos and production-ready systems.

Download E-Book

What makes RAG “advanced”?

Naive RAG follows a fixed four-step pipeline: chunk documents, embed chunks, retrieve the top-K by cosine similarity, and pass results to the LLM. It’s straightforward to implement, and it hits a hard ceiling on accuracy. A 2024 comprehensive RAG benchmark found that state-of-the-art RAG systems answer only 63% of factual questions correctly. Straightforward RAG without advanced techniques scores 44%; LLMs with no retrieval at all score around 34%.

Advanced RAG adds quality-control layers at one or more stages of that pipeline. These fall into four categories: pre-retrieval (query transformation, HyDE), retrieval-time (hybrid search, contextual chunking), post-retrieval (reranking, compression, self-reflection), and architecture-level changes (Self-RAG, Adaptive RAG, Modular RAG). Each category addresses a different failure mode. The right choice depends on where your pipeline is breaking down.

What it covers	12 advanced RAG techniques with real benchmarks
Why it matters	Naive RAG answers only 63% of factual questions correctly
Techniques covered	Hybrid retrieval, Reranking, Self-RAG, RAPTOR, GraphRAG
Difficulty range	Low (sentence window) to High (Self-RAG, GraphRAG)
Key research	ARAGOG, RAPTOR (arXiv:2401.18059), Self-RAG (arXiv:2310.11511), Contextual Retrieval (Anthropic, 2024)

Five criteria separate techniques worth implementing from those worth skipping:

Accuracy lift: measured improvement on standard evaluation benchmarks (ARAGOG, QuALITY, open-domain QA datasets)
Implementation complexity: time, tooling, and whether fine-tuning is required
Latency impact: extra LLM calls or compute per query
Framework support: native support in LangChain, LlamaIndex, or Haystack
Production adoption: practitioner consensus from GitHub, Reddit, and HN discussions

When you find yourself asking whether RAG is better than fine-tuning for your use case, these criteria give you a consistent basis for comparison.

Comparison table: all 12 advanced RAG techniques at a glance

Technique	What it does	Best for	Accuracy gain	Complexity
Hybrid Retrieval	Dense (vector) + sparse (BM25) + RRF merge	Production default; exact-match + semantic queries	Significant (de facto standard)	Medium
Cross-Encoder Reranking	Second-pass scoring of (query, doc) pairs	Precision-critical pipelines	Consistent NDCG/MRR lift	Low-Medium
Contextual Retrieval	LLM-prepended chunk context before embedding + BM25	Isolated chunks losing document context	67% fewer retrieval failures (Anthropic)	Medium
HyDE	Generate hypothetical answer doc, embed it for search	Short/vague queries vs. technical corpus	nDCG@10: 61.3 vs. 44.5 baseline	Low-Medium
Self-RAG	LLM decides when to retrieve; reflection tokens grade output	Factuality; over-retrieval avoidance	ICLR 2024 Oral; beats standard RAG on open QA	High
CRAG	Evaluator grades retrieved docs; fallback to web search	High-stakes domains (legal, medical)	Significant over RAG on 4 datasets	Medium
Adaptive RAG	Classifier routes query to no/single/multi-step retrieval	Mixed-complexity production workloads	Efficiency + accuracy on open-domain QA	Medium
GraphRAG	Knowledge graph + community summaries + graph traversal	Multi-hop, relationship-heavy queries	“Substantial” improvement (Microsoft Research)	High
RAPTOR	Recursive clustering + abstractive tree indexing	Long documents; cross-section reasoning	+20% absolute on QuALITY benchmark	High
RAG Fusion	Multi-query generation + RRF merge	Ambiguous queries; recall-priority tasks	Broader coverage vs. single-query	Low-Medium
Sentence Window / Parent-Child	Index small chunks, retrieve surrounding window/parent	Precision matching + rich generation context	#1 retrieval precision in ARAGOG	Low-Medium
Modular RAG	Swappable pipeline modules (retrieval, rerank, memory)	Evolving production systems	Architecture-level improvement	Medium-High

The 12 advanced RAG techniques: quick overview

Here are the 12 techniques covered in this guide, with jump links to the full breakdown for each:

Hybrid Retrieval: the production-standard combination of dense and sparse search
Cross-Encoder Reranking: second-pass precision scoring after initial retrieval
Contextual Retrieval: LLM-generated chunk context that cuts retrieval failure rates by 67%
HyDE: hypothetical document embeddings that bridge vocabulary gaps
Self-RAG: on-demand retrieval with built-in output critique
CRAG: retrieval evaluation with web search fallback
Adaptive RAG: query routing to right-sized retrieval pipelines
GraphRAG: knowledge graphs for multi-hop relationship queries
RAPTOR: recursive tree indexing for long-document reasoning
RAG Fusion: multi-query generation for broader coverage
Sentence Window / Parent-Child Chunking: decoupled retrieval and generation granularity
Modular RAG: swappable pipeline architecture for evolving systems

The 12 advanced RAG techniques in depth

Technique 1: Hybrid retrieval

Hybrid retrieval combines dense vector search (semantic similarity via embeddings) with sparse keyword search (BM25 or SPLADE), then merges ranked lists using Reciprocal Rank Fusion. Dense handles paraphrasing and synonyms; sparse handles exact terms, product codes, and rare names that embeddings miss. It is the de facto production standard as of 2025-2026.

When to use it: Default for any production RAG system; critical when queries contain proper nouns, SKUs, or acronyms. Complexity: Medium. Framework support: LangChain (EnsembleRetriever), LlamaIndex (QueryFusionRetriever), Haystack, Weaviate, and Qdrant all support it natively.

The hybrid RAG pattern addresses both the semantic gap and the exact-match gap in one step, making it the highest-ROI starting point for most teams upgrading from naive retrieval.

Technique 2: Cross-encoder reranking

Reranking applies a second-pass cross-encoder model that scores each (query, document) pair directly. Cross-encoders are more accurate than the bi-encoder used for initial retrieval. Top-K reranked results feed the LLM, and the technique is considered the easiest high-ROI upgrade after hybrid search.

What it is: Cross-encoders read query and document together (not independently), producing more accurate relevance scores. MiniLM cross-encoders trained on MS MARCO consistently improve NDCG/MRR. Complexity: Low-Medium. Key tools: Cohere Rerank, ColBERT, MiniLM, Flashrank.

Important nuance: The ARAGOG benchmark found that Cohere Rerank showed no notable advantage over the naive RAG baseline on its evaluation corpus, while LLM-based reranking did show improvement. This is corpus-dependent: commercial rerankers are not universally superior to open-source cross-encoders. Always benchmark reranking on your own data before assuming gains. The relationship between retrieval precision and hallucination depends heavily on what the reranker was trained on.

Technique 3: Contextual retrieval

Contextual Retrieval prepends chunk-specific context (generated by an LLM from the full document) to each chunk before embedding and BM25 indexing. This solves the core problem of isolated chunks losing their document context. Combined with reranking, it reduces retrieval failure rates by 67%.

Benchmarks (Anthropic internal testing):

Contextual embeddings alone: 35% reduction in retrieval failures (5.7% to 3.7%)
Plus contextual BM25: 49% reduction (5.7% to 2.9%)
Plus reranking: 67% reduction (5.7% to 1.9%)

Complexity: Medium (one LLM call per chunk at index time, not per query). A one-time cost with permanent accuracy gains. Best for: Any enterprise document corpus where chunk quality and retrieval accuracy are misaligned because chunks lose meaning in isolation. See also: context caching for managing the token costs of prepending context at index time.

Technique 4: HyDE (hypothetical document embeddings)

HyDE generates a hypothetical document that would answer the query using an LLM, then embeds that document for vector search rather than embedding the raw query. The generated document captures the semantic “shape” of a good answer, closing the query-document vocabulary mismatch between short queries and long technical documents.

Benchmark: nDCG@10 = 61.3 on DL-19 versus 44.5 for the Contriever baseline, approaching fine-tuned retriever performance in a zero-shot setting. ARAGOG also confirms HyDE significantly enhances retrieval precision.

Complexity: Low-Medium (one extra LLM call per query). Framework support: LlamaIndex (HyDEQueryTransform), Haystack (native). Best for: Short or vague queries against technical or specialized corpora; zero-shot retrieval without labeled data.

Technique 5: Self-RAG

Self-RAG trains the LLM to decide when to retrieve, then critique both retrieved passages (IsREL) and its own generated output (IsSUP, IsUSE) using special reflection tokens. This on-demand LLM reasoning approach avoids unnecessary retrieval for simple queries and catches unsupported claims before they reach the user.

Benchmark: ICLR 2024 Oral (top 1%). Self-RAG 7B and 13B significantly outperform Llama2 and standard RAG on open-domain QA, reasoning, and fact verification. Only 2% of correct predictions came from outside retrieved passages, versus 15-20% in Alpaca/Llama2 baselines.

Complexity: High (requires LLM fine-tuning with reflection tokens). When to use: Factuality-critical applications where citation accuracy matters and self-reflective retrieval is worth the training investment. Framework support: LangGraph (Self-RAG workflows). For teams deploying Self-RAG at scale, see scaling in production considerations.

Technique 6: CRAG (corrective retrieval augmented generation)

CRAG adds a lightweight retrieval evaluator that grades retrieved documents before generation. Each document is decomposed into knowledge strips, scored for relevance, and if quality is too low, CRAG falls back to web search. This prevents bad retrieval from propagating into bad answers.

Three paths CRAG takes:

Retrieve and generate if confidence is high
Distill and augment with web results if partially relevant
Fall back entirely to web search if local corpus misses the query

Note: the web search fallback path in CRAG introduces a prompt injection risk — malicious content in retrieved web pages can be crafted to manipulate the LLM’s output. Sanitize web-retrieved content before passing it to the generator.

Benchmark: Significant improvement over standard RAG across 4 short- and long-form generation datasets (Yan et al., 2024). Note: this is the CRAG technique paper. A separate 2024 comprehensive RAG benchmark study (arXiv:2406.04744) established the 63% accuracy ceiling for SOTA RAG systems generally; these are different papers with related findings. Complexity: Medium (the evaluator module plugs into existing pipelines without full fine-tuning). Best for: Retrieval quality gates in legal, medical, and compliance domains where hallucination is unacceptable. Understanding why AI agents fail in production helps prioritize where CRAG adds the most protection.

Technique 7: Adaptive RAG

Adaptive RAG trains a small query classifier to route each question to the optimal retrieval path: no retrieval for simple factual queries, single-step retrieval for moderate queries, and multi-step iterative retrieval for complex reasoning. It balances cost and accuracy across mixed workloads.

Benchmark: Enhanced efficiency and accuracy on open-domain QA versus iterative and single-step RAG baselines (NAACL 2024). The 2026 practitioner consensus describes Adaptive RAG as the “emerging best practice” for routing queries by complexity in production systems.

Complexity: Medium (train a small classifier once; modular after deployment). Framework support: LangGraph, custom routing logic. Best for: Enterprise systems with heterogeneous query types, which is the common case. For broader comparisons, see agent frameworks compared for implementation options across LangGraph and alternatives.

Technique 8: GraphRAG

GraphRAG constructs a knowledge graph from source documents (entities as nodes, relationships as edges), then uses community detection (the Leiden algorithm) to build hierarchical summaries. At query time, graph traversal enables multi-hop reasoning that pure vector search cannot support.

Benchmark: “Substantial improvements over conventional RAG” for comprehensiveness and diversity on global sensemaking across 1M+ token corpora (Microsoft Research). GraphRAG also uses 26% to 97% fewer tokens than some alternatives on global queries.

Complexity: High (offline knowledge graph construction is compute-intensive). Framework support: Microsoft open-source, Neo4j, LlamaIndex knowledge graph index, RAGFlow. Best for: Multi-entity relationship queries, competitive intelligence, and regulatory analysis. See also: GraphRAG vs. standard vector RAG. GraphRAG also pairs well with a vector database for hybrid graph-plus-embedding retrieval patterns.

Technique 9: RAPTOR

RAPTOR recursively clusters and summarizes text chunks into a tree of increasing abstraction. At inference time, it retrieves from multiple tree levels simultaneously, including both fine-grained chunks and high-level summaries. This enables multi-granularity retrieval for complex, multi-section questions.

Benchmark: +20% absolute accuracy on QuALITY with GPT-4. On the QuALITY dataset: 62.4% versus DPR (60.4%) and BM25 (57.3%). RAPTOR combined with HyDE and reranking achieves approximately 99% retrieval accuracy on SQuAD.

Complexity: High (offline tree construction with indexing and storage overhead). Framework support: LlamaIndex (RAPTOR tree index). Best for: Long documents requiring multi-level document retrieval, including annual reports, legal contracts, and technical manuals where both detail and high-level context matter.

Technique 10: RAG Fusion

Known limitation: Multi-query rewrites can be “nearly identical and lacking in diversity,” limiting the recall benefit. The DMQR-RAG approach (arXiv:2411.13154) addresses this with diversity-maximizing methods.

Complexity: Low-Medium (one extra LLM call for query generation plus standard RRF). Framework support: LangChain (MultiQueryRetriever + RRF). Best for: Ambiguous or underspecified queries and query transformation techniques for information synthesis tasks where recall matters more than precision.

Technique 11: Sentence window / parent-child chunking

Sentence Window Retrieval indexes individual sentences for precise matching but retrieves the surrounding window of sentences for richer generation context. Parent-Child Chunking indexes small chunks but returns the parent block. Both solve the same problem: retrieval granularity and generation context have different optimal sizes.

Benchmark: Sentence Window Retrieval ranked #1 for retrieval precision in the ARAGOG head-to-head evaluation, beating HyDE, Document Summary Index, Multi-query, MMR, and both Cohere Rerank and LLM Rerank. This finding makes it the strongest low-complexity option.

Complexity: Low-Medium. Framework support: LlamaIndex (SentenceWindowNodeParser), LangChain (ParentDocumentRetriever). Best for: Any pipeline where chunking strategy and data quality are the primary bottleneck.

Technique 12: Modular RAG

Modular RAG decomposes the retrieval pipeline into independent, swappable modules (retrieval, reranking, query transformation, memory, generation), each configurable independently. It is an architectural pattern, not a single technique with its own accuracy benchmark. LangChain, LlamaIndex, Haystack, and RAGFlow are all fundamentally modular RAG implementations already. Understanding the broader AI agent stack helps teams position Modular RAG within their full agent architecture.

Why it matters as a design choice: Teams that adopt a modular architecture from the start can swap or upgrade individual components (for example, replacing a bi-encoder retriever with a cross-encoder, or adding a reranking step) without rebuilding the full pipeline. This separates the decisions of “what technique to use” from “how the pipeline is structured,” which matters at enterprise scale. See modular retrieval architecture with MCP for how MCP fits into this pattern.

Complexity: Medium-High (requires upfront interface design decisions). Best for: Production systems that will iterate over time and teams running A/B tests on retrieval components.

How to choose the right advanced RAG technique

Choose by diagnosing your bottleneck first. If retrieval misses exact terms, start with hybrid search. If precision is low after retrieval, add reranking. If chunks lose context, add contextual retrieval. Complex relationship queries need GraphRAG; long documents need RAPTOR; mixed-complexity workloads need Adaptive RAG. Fix chunking before optimizing retrieval algorithms.

If you need…	Try…	Why
Production default, highest ROI first	Hybrid Retrieval + Sentence Window Chunking	Fixes exact-match gaps and chunking granularity with low complexity
Better precision on retrieved results	Cross-Encoder Reranking	Second-pass scoring; test on your corpus first, Cohere didn’t beat naive RAG in ARAGOG
Chunks losing document context	Contextual Retrieval	67% fewer retrieval failures; one-time indexing cost
Short/vague queries on technical docs	HyDE	Closes vocabulary gap; approximately 1 extra LLM call per query
Factuality and citation accuracy	Self-RAG	Highest accuracy ceiling; requires fine-tuning investment
High-stakes, must-not-hallucinate	CRAG	Evaluator + web fallback; plugs into existing pipeline
Mixed query complexity, cost control	Adaptive RAG	Routes queries to right-sized pipeline
Multi-entity relationship questions	GraphRAG	Enables multi-hop reasoning; high offline construction cost
Long documents, cross-section reasoning	RAPTOR	Tree-level retrieval; +20% accuracy on QuALITY benchmark
Ambiguous queries, recall priority	RAG Fusion	Multi-query + RRF; quick to implement
Systems that must evolve over time	Modular RAG	Architectural pattern; iterate components independently

Two practical guidelines the comparison table won’t tell you:

Start with the foundation layer. Hybrid retrieval plus sentence window chunking plus contextual retrieval covers most accuracy gaps for most enterprise use cases. Invest in high-complexity architectures only after these are in place.
Data preparation takes longer than technique selection. Practitioners consistently report spending three or more weeks on data ingestion and cleaning before retrieval technique choice becomes the bottleneck. The best algorithm applied to ungoverned data still underperforms. For the enterprise RAG decision framework, data quality is a prerequisite, not a follow-up.

How Atlan’s governed context layer improves RAG accuracy

Every advanced RAG technique optimizes retrieval mechanics. But retrieval quality is ultimately bounded by what is in the index. When an agent retrieves “revenue” from an enterprise data warehouse, it may get three definitions from three different systems, none of them certified, none carrying lineage. This is a common pain point for AI agents for data engineering teams, where schema ambiguity compounds retrieval errors. The 2024 comprehensive RAG benchmark documents this ceiling: even state-of-the-art RAG answers only 63% of factual questions correctly. The bottleneck is often data quality and metadata completeness, not the retrieval algorithm.

Atlan’s context layer for AI provides the governed metadata foundation that advanced RAG techniques need to work reliably at enterprise scale. This includes column-level descriptions, data lineage from source to consumption, business glossary definitions, certification status from data owners, and data quality scores. When an AI agent retrieves a data asset through Atlan’s MCP server, the retrieved chunk carries provenance: certified, mapped to a governed glossary term, and tagged with lineage.

The results are measurable. In Atlan’s own research across 522 enterprise queries, AI agents grounded in context-rich metadata achieved 38% higher SQL accuracy versus agents using semantic definitions alone. This is an internal benchmark, not a universal claim, but it illustrates the directional point: advanced retrieval techniques combined with governed context outperform advanced retrieval over ungoverned data. The active metadata that Atlan continuously captures (usage patterns, freshness signals, quality certifications) makes every retrieval call more accurate, not just more relevant.

Build Your AI Context Stack

Get the blueprint for implementing context graphs across your enterprise. This guide walks through the four-layer architecture, from metadata foundation to agent orchestration, with practical implementation steps for 2026.

Get the Stack Guide

Real stories from real customers: context-governed RAG in production

"We're excited to build the future of AI governance with Atlan. All of the work that we did to get to a shared language at Workday can be leveraged by AI via Atlan's MCP server...as part of Atlan's AI Labs, we're co-building the semantic layer that AI needs with new constructs, like context products."

Joe DosSantos, VP of Enterprise Data and Analytics, Workday

Watch Now

"Atlan is much more than a catalog of catalogs. It's more of a context operating system...Atlan enabled us to easily activate metadata for everything from discovery in the marketplace to AI governance to data quality to an MCP server delivering context to AI models."

Sridher Arumugham, Chief Data and Analytics Officer, DigiKey

Watch Now

Why the knowledge foundation matters more than the algorithm

Advanced RAG techniques are necessary, but they are not sufficient. Hybrid retrieval, RAPTOR, and Self-RAG all assume a clean, well-structured, governed knowledge base on the other side of the retrieval call. When that foundation is missing, even the most sophisticated retrieval algorithm returns results that are relevant but untrustworthy: technically correct vectors pointing to uncertified, context-free data. This is precisely why AI agents need an enterprise context layer — the retrieval mechanics alone cannot compensate for ungoverned data.

The practitioners who get the most from advanced RAG techniques are the ones who invest in data preparation first. That means governed metadata, consistent business glossary definitions, certified assets, and lineage that lets the retrieval system explain why a result was retrieved, not just that it was. This is where the context layer for AI agents becomes the deciding factor between agents that demonstrate well and agents that work reliably in production. Effective context management at the enterprise level is what separates retrieval that scales from retrieval that breaks under real workloads.

The 12 techniques in this guide are the algorithmic layer. The governed knowledge base is the foundation layer. Both are required.

Book a Demo

FAQs about advanced RAG techniques

What are the most advanced RAG techniques?

The most benchmark-backed advanced RAG techniques are Contextual Retrieval (67% fewer retrieval failures), RAPTOR (+20% absolute accuracy on QuALITY), Self-RAG (ICLR 2024 Oral, beats standard RAG on open-domain QA), GraphRAG (multi-hop sensemaking), and Hybrid Retrieval (de facto production standard). For most teams, hybrid retrieval plus sentence window chunking delivers the best ROI at the lowest implementation cost.

What is the difference between naive RAG and advanced RAG?

Naive RAG chunks documents, embeds them, retrieves top-K by cosine similarity, and passes results to the LLM. Advanced RAG adds quality-control layers at multiple stages: query transformation, hybrid retrieval, reranking, and self-reflection. State-of-the-art naive RAG answers only 63% of factual questions correctly; advanced techniques push that ceiling higher at the cost of implementation complexity.

How does Self-RAG work?

Self-RAG fine-tunes an LLM to use special reflection tokens that determine (1) whether retrieval is needed for a given query, (2) whether retrieved passages are relevant (IsREL), and (3) whether the generated output is supported by retrieved evidence (IsSUP) and useful (IsUSE). Unlike standard RAG, retrieval is on-demand rather than always-on. Self-RAG achieves top performance on open-domain QA at the cost of a fine-tuning requirement.

What is Corrective RAG (CRAG) and when should you use it?

CRAG adds a retrieval evaluator that grades retrieved documents before passing them to the LLM. Documents are decomposed into knowledge strips and scored for relevance. If confidence is low, CRAG falls back to web search. Use it in high-stakes domains (legal, medical, compliance) where bad retrieval leading to bad output is unacceptable. It plugs into existing pipelines without requiring full model fine-tuning.

What is HyDE in RAG?

HyDE stands for Hypothetical Document Embeddings. Instead of embedding the user query, an LLM generates a hypothetical document that would answer the query, and that document is embedded for vector search. The hypothesis captures the semantic shape of a good answer even if its specific details are hallucinated. HyDE achieves nDCG@10 of 61.3 versus 44.5 for a standard dense retriever baseline.

What is RAPTOR and how does it improve multi-hop retrieval?

RAPTOR builds a hierarchical tree of text summaries through recursive clustering and abstractive summarization. At retrieval time, it searches across multiple tree levels simultaneously (both granular chunks and high-level summaries), enabling answers that require synthesizing information from across a long document. RAPTOR combined with GPT-4 improved accuracy on the QuALITY reading comprehension benchmark by 20% in absolute terms versus prior state of the art.

How does GraphRAG differ from standard vector RAG?

Standard vector RAG retrieves documents by embedding similarity: it finds chunks that look like the query. GraphRAG builds a knowledge graph from source documents, then uses community detection to create hierarchical summaries of entity clusters. This enables multi-hop reasoning that pure similarity search cannot support. GraphRAG is best for complex relationship queries across large corpora; its main tradeoff is expensive offline graph construction.

What is Adaptive RAG and how does query routing work?

Adaptive RAG trains a small, fast classifier on query examples to predict complexity. Each query is routed to one of three pipelines: direct LLM answer for simple factual questions, single-step retrieval for moderate queries, or multi-step iterative retrieval for complex reasoning. This optimizes for cost and latency on mixed enterprise workloads without sacrificing accuracy on complex queries.

How does hybrid search improve RAG accuracy?

Hybrid search combines dense vector search (finds semantically similar content) with sparse keyword search like BM25 (finds exact term matches). The two ranked result lists are merged using Reciprocal Rank Fusion. Dense retrieval alone misses exact product codes, names, and acronyms; sparse retrieval alone misses paraphrasing and synonyms. Combining both covers the full range of query types and is widely considered the highest-ROI upgrade for any RAG pipeline.

What is RAG Fusion?

RAG Fusion generates multiple alternative phrasings of the original query using an LLM, retrieves documents for each query variant, then merges all ranked lists using Reciprocal Rank Fusion. The result is broader document coverage from multiple query perspectives. It is most useful for ambiguous or underspecified queries and information synthesis tasks. The main limitation: generated query variants can be too similar to each other, limiting diversity gains.

Sources

Share this article

Atlan is the next-generation platform for data and AI governance. It is a control plane that stitches together a business's disparate data infrastructure, cataloging and enriching data with business context and security.

Book a Demo Context Studio Live