Reference architecture

Production RAG Reference Architecture & Retrieval Pipeline Guide

Production architecture blueprint for multi-stage RAG pipelines—covering chunking strategies, hybrid search fusion (RRF math), cross-encoder re-ranking, vector database indexing, and evaluation guardrails.

18 minVerified 2026-08-073 primary sources
A governed production AI reference architecture with observable, secured service boundaries.

Architecture

Building an enterprise-grade Retrieval-Augmented Generation (RAG) pipeline requires coordinating text ingestion, chunking, dual-embedding indexing, hybrid retrieval fusion, and cross-encoder re-ranking into a deterministic execution pipeline.

Document ingestion & semantic chunking
Dual embedding (Dense HNSW + Sparse BM25)
Hybrid retrieval execution
Reciprocal Rank Fusion (RRF) score normalization
Cross-encoder re-ranking top-K filter
Context injection into LLM generation prompt
Conceptual teaching model synthesized from:LangGraph Agentic StateGraph Execution Engine RepositoryBuilding Effective AI Agents

1. Mathematical Score Normalization: Reciprocal Rank Fusion (RRF)

Standard vector similarity scores (e.g., Cosine Distance) and lexical search scores (e.g., BM25 TF-IDF) operate on different scale distributions. Combining them directly leads to biased ranking.

Reciprocal Rank Fusion (RRF) normalizes search ranks without requiring arbitrary score calibration:

Reciprocal Rank Fusion (RRF) Score Equation
Mathematical Formulation
RRF\_Score(d \in D) = \sum_{m \in M} \frac{1}{k + r_m(d)}

Combines rank positions across dense vector search and sparse lexical search into a unified relevance score.

2. Architecture Comparison: Vector Search vs. Lexical vs. Hybrid RAG

Retrieval Indexing Architecture Trade-Off Matrix
Architecture OptionPrimary Best-For Case

Decisions

| Decision | Required evidence | Review trigger | |---|---|---| | Use Reciprocal Rank Fusion (RRF) for hybrid search score normalization. | Benchmark evaluation showing improved top-5 recall over single vector search | Keyword lookup accuracy dropping below 90% | | Enforce cross-encoder re-ranking top-K filter on candidate pools. | Latency budget audit confirming re-ranking latency $\le 30\text$ | P99 query response latency exceeding SLA | | Combine dense HNSW vector index with BM25 sparse keyword index. | Query retrieval evaluation confirming exact SKU and technical term precision | Unhandled technical jargon search failures |

Alternatives and trade-offs

Pure dense vector search provides strong semantic discovery but fails on specific alphanumeric identifiers. Hybrid search with RRF fusion resolves exact keyword lookup failures at a modest latency cost (+6ms). Adding a cross-encoder re-ranker increases precision significantly but adds GPU inference overhead to the query path.

Failure modes

  • Fixed-size chunking splitting mathematical equations or code blocks across chunk boundaries.
  • Single-vector search missing exact product SKUs or serial number queries.
  • Re-ranker timeouts under high concurrent request spikes degrading retrieval to raw vector ordering.

Operational checklist

  • [ ] Chunking parser respects markdown headers and sentence boundaries.
  • [ ] Dual-embedding pipeline generates HNSW dense vectors and BM25 sparse tokens concurrently.
  • [ ] RRF smoothing constant is configured to $k = 60$.
  • [ ] Cross-encoder re-ranking top-K is capped at 10 items to preserve low-latency response times.

Connected practice

Sources

  • langgraph-repo
  • anthropic-effective-agents
  • anthropic-trustworthy-agents