RAG Patterns — Cheat Sheet

Tags: rag, vector-search, embeddings, reranking, hybrid-search, llm

Level: intermediate

TL;DR: Architecture patterns for Retrieval-Augmented Generation, from baseline naive retrieval to advanced hybrid routing and reranking pipelines.

Key concepts

Naive RAG uses direct semantic search over chunked documents without post-processing.

Advanced RAG adds pre-retrieval query rewriting and post-retrieval reranking stages.

Modular RAG decouples indexing, retrieval, routing, synthesis, and evaluation components.

Chunking strategies include fixed-size sliding window, semantic boundary, and recursive splitting.

Hybrid search merges dense vector similarity with sparse lexical matching via BM25.

Reciprocal Rank Fusion combines ranked lists from distinct sparse and dense retrievers.

Cross-encoder rerankers rescore candidate chunks by scoring query-document pairs jointly.

Contextual compression trims irrelevant sentences before assembling the final prompt payload.

Hypothetical Document Embeddings generate synthetic candidate passages to improve dense retrieval.

Multi-query expansion decomposes ambiguous questions into parallel domain-specific search vectors.

Typical questions and ready answers

Q: What is the difference between dense retrieval and sparse retrieval in RAG?R: Dense retrieval maps text to semantic embedding spaces capturing conceptual similarity, while sparse retrieval uses exact keyword frequencies via algorithms like BM25; combining both via hybrid search mitigates vocabulary mismatch while preserving context.

Q: Why do we use cross-encoders for reranking instead of using them for initial search?R: Cross-encoders process the query and passage together through all attention layers, which is highly accurate but computationally intractable across millions of records; we use bi-encoders to fetch top-k fast, then cross-encoders to rerank only those candidates.

Q: How do you handle multi-hop reasoning in complex RAG workflows?R: I use iterative or agentic retrieval patterns, decomposing the user prompt into sub-queries, retrieving intermediate evidence, and feeding findings back to generate follow-up lookups before synthesizing the final output.

Pitfalls and common mistakes

Embedding entire large documents directly instead of splitting into context-preserving chunks.

Failing to include document metadata filters, causing semantic drift across temporal domains.

Setting retrieval top-k too high, exceeding context windows and introducing irrelevant distraction.

Ignoring out-of-vocabulary terms and specialized acronyms by relying exclusively on dense embeddings.

Neglecting chunk overlap, splitting sentences across chunk boundaries and breaking logical context.

Snippets and commandsTypeScript

```import { Document, VectorStoreIndex, CohereRerank } from "llamaindex";

// Hybrid retrieval with cross-encoder reranker

const index = await VectorStoreIndex.fromDocuments(documents);

const retriever = index.asRetriever({ similarityTopK: 20 });

const reranker = new CohereRerank({ topN: 5 });

const initialNodes = await retriever.retrieve({ query: userQuery });

const finalNodes = await reranker.postprocessNodes(initialNodes, userQuery);

Reciprocal Rank Fusion (RRF) calculation

def reciprocal_rank_fusion(dense_ranks, sparse_ranks, k=60):

rrf_scores = {}

for doc_id, rank in dense_ranks.items():

rrf_scores[doc_id] = rrf_scores.get(doc_id, 0.0) + (1.0 / (k + rank))

for doc_id, rank in sparse_ranks.items():

rrf_scores[doc_id] = rrf_scores.get(doc_id, 0.0) + (1.0 / (k + rank))

return sorted(rrf_scores.items(), key=lambda item: item[1], reverse=True)

```

To review

[ ] Multi-vector retrieval using summary-to-chunk lookups

[ ] Parent document retrieval via recursive chunk hierarchical trees

[ ] Self-RAG evaluation loops for hallucination detection

[x] Dense vector similarity metrics: Cosine, Dot Product, Euclidean

[x] Basic BM25 tokenization and inverse document frequency scoring

References

Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (Lewis et al.)

Precise Zero-Shot Dense Retrieval without Relevance Labels (HyDE paper)

Reciprocal Rank Fusion outperforms Condorcet and individual approaches (Cormack et al.)

Jarvis