RAG in Detail — 5-Step Pipeline and 12 Optimization Strategies

Retrieval-Augmented Generation (RAG) has become the dominant architecture for grounding LLM outputs in external knowledge. This article dissects the five-stage pipeline and presents twelve production-tested optimization strategies.

The 5-Step RAG Pipeline

Step 1: Document Ingestion

Raw documents pass through a parsing pipeline handling PDFs, HTML, markdown, and plain text. Each document is cleaned, normalized, and split into chunks. Chunk size directly impacts retrieval quality: smaller chunks improve precision but miss contextual relationships.

Step 2: Embedding Generation

Each chunk is vectorized using an embedding model such as OpenAI text-embedding-3-small, Cohere Embed v3, or open-source alternatives like BGE and E5. The embedding dimension determines storage requirements and search latency.

Step 3: Vector Index Construction

Embeddings are stored in a vector database with an Approximate Nearest Neighbor (ANN) index. Options include Pinecone (managed serverless), Weaviate (open-source), Milvus (distributed), and pgvector (PostgreSQL). Index type selection (IVF, HNSW, DiskANN) trades recall against query speed.

Step 4: Query Processing

User queries undergo transformation before retrieval. Query rewriting, hypothetical document embeddings, and multi-query expansion improve recall by generating multiple search variations.

Step 5: Generation

Retrieved chunks are inserted into a structured prompt alongside the original query. The LLM synthesizes an answer from provided context.

12 Optimization Strategies

1. Chunk Overlap: 10-20% overlap prevents context fragmentation
2. Hierarchical Indexing: Summary-level then detail-level retrieval
3. Hybrid Search: Dense vector search combined with BM25 keyword matching
4. Reranking: Cross-encoder rerankers improve top-k relevance by 15-25%
5. Context Window Management: Trim retrieved chunks to fit LLM limits
6. Metadata Filtering: Pre-filter by date, source, or category before vector search
7. Query Routing: Classify queries to select domain-specific retrievers
8. Embedding Fine-tuning: Domain-adapted embeddings improve accuracy by 10-30%
9. Caching: Cache frequent query results at the retrieval layer
10. Streaming: Stream output for better UX on long responses
11. Evaluation: Automated hit-rate and MRR evaluation on held-out sets
12. Asynchronous Ingestion: Parallel document processing with backpressure

Production Architecture

A production RAG system separates ingestion (batch, async) from serving (real-time, synchronous). The serving path targets sub-500ms query-to-response latency.

Conclusion

RAG remains the most practical approach for knowledge-intensive LLM applications. The difference between a demo and a production system lies in these optimization details.