// Retrieval-Augmented Generation patterns combining vector search and LLMs for knowledge-intensive tasks including chunking strategies, hybrid search, and reranking.
[HINT] Download the complete skill directory including SKILL.md and all related files
| name | rag-patterns |
| description | Retrieval-Augmented Generation patterns combining vector search and LLMs for knowledge-intensive tasks including chunking strategies, hybrid search, and reranking. |
Retrieval-Augmented Generation combines parametric (LLM) and non-parametric (vector database) memory for knowledge-intensive tasks. Originally proposed by Lewis et al. (2020), RAG retrieves relevant documents and uses them as context for LLM generation.
Sources: Lewis et al. NeurIPS 2020, LangChain Docs, Pinecone RAG Guide
Fixed-Size: Split by tokens/characters (500-1500 chars) with 10-20% overlap. Simple but ignores semantic boundaries.
Semantic: Group sentences by embedding similarity. Preserves meaning but computationally expensive.
Recursive: Split by structure (headers, paragraphs, sentences). Best for structured docs.
Document-Based: Use Markdown structure, code classes, tables as boundaries.
Recommendation: Start with 1000 chars + 200 overlap. Tune based on query patterns and embedding model context window.
| Model | Dims | Use Case |
|---|---|---|
| text-embedding-3-small | 1536 | Cost-effective, general |
| text-embedding-3-large | 3072 | Higher accuracy |
| Cohere embed-v3 | 1024 | Multilingual |
Best Practice: Use instruction prefixes (query: for searches, passage: for docs). Store original text in metadata to avoid separate DB lookup.
Combine semantic (dense vectors) + keyword (BM25) for robust retrieval.
BM25: TF-IDF refinement for exact term matching (handles acronyms, names, IDs)
Dense Vectors: Semantic similarity (handles paraphrases, synonyms)
Reciprocal Rank Fusion (RRF): Merge results from both methods
score = 1 / (60 + rank) // constant k=60
Retrieve 2x desired results from each method, merge with RRF, return top-k.
After retrieval, rerank with cross-encoder models (e.g., BGE Reranker) that score query-document pairs directly. More expensive but significantly improves precision.
Pattern: Retrieve 20-50 candidates → Rerank → Return top 5-10
Rerankers focus on relevance vs similarity - crucial for multi-hop reasoning.
Metadata Filtering: Pre-filter by date, author, category before semantic search. Vastly more efficient than post-filtering.
Prompt Template:
Context: [Retrieved chunks with sources]
Instructions:
- Answer using only provided context
- Cite sources as [Source N]
- Say "insufficient information" if context lacks answer
Question: {query}
Citation Parsing: Extract [Source N] patterns, map to original document metadata.
Irrelevant Results: Add hybrid search, metadata filters, reranking
Missing Context: Increase overlap (20%), use parent document retrieval (store small chunks but retrieve larger parent)
Hallucination: Enforce citation requirements, lower temperature (0.1-0.3), validate claims against sources
Slow Performance: Cache embeddings (by hash), batch requests, use approximate nearest neighbor (HNSW, IVF)
Retrieval: Precision (relevant/retrieved), Recall (relevant/total), MRR (mean reciprocal rank)
Generation: Faithfulness (uses only context), Answer Relevancy (addresses question), Context Precision (minimal irrelevant chunks)
Use test sets with ground truth Q&A pairs. Measure retrieval quality separately from generation quality.