Building a Production Hybrid RAG: Why I Threw Out the LangChain Recipe

2026-05-24 · Source: LLM on Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Cloud Computing & IT Infrastructure · Depth: Advanced, long

Summary

A production-ready hybrid RAG system, detailed in a new open-source project, demonstrates how to move beyond basic tutorials to achieve deployable performance, reducing hallucination rates from 12% to 3%. Key architectural decisions include using Reciprocal Rank Fusion (RRF) for combining dense and sparse retrieval, which is parameter-free and score-agnostic, outperforming individual components with 89% accuracy on mixed queries. It employs cross-encoder reranking with `ms-marco-MiniLM-L-6-v2` for improved relevance (5-8% keyword recall, 7 points NDCG@5) on top-15 candidates after bi-encoder retrieval. The system incorporates explicit abstention logic, first-class citations, and a rigorous evaluation framework using metrics like RAGAS Faithfulness and BERTScore F1. Adaptive routing, semantic caching, and separate guardrail passes for PII/toxicity are also implemented, alongside comprehensive production hardening for observability, scaling, and CI/CD.

Key takeaway

For MLOps Engineers building RAG systems, relying solely on basic tutorials is insufficient for production. You should prioritize robust architectural decisions like Reciprocal Rank Fusion, cross-encoder reranking, and explicit abstention logic to achieve reliable performance and reduce hallucination rates. Implement a comprehensive evaluation framework early to quantify improvements and ensure system stability, avoiding "shipping on vibes."

Key insights

Production RAG requires robust engineering beyond basic tutorials, focusing on hybrid retrieval, rigorous evaluation, and operational resilience.

Principles

• Reciprocal Rank Fusion (RRF) combines retrieval ranks, not scores, for robustness.
• Cross-encoders provide cleaner relevance signals than bi-encoders for reranking.
• Explicit abstention logic prevents hallucinations and enables monitoring.

Method

The proposed RAG pipeline involves bi-encoder retrieval (top-3k), RRF for fusing dense/sparse rankings (top-15), cross-encoder reranking, LLM generation with abstention, and structured citation output. Adaptive routing and semantic caching optimize performance.

In practice

• Implement RRF with `k=60` for combining dense and sparse retrieval.
• Use `ms-marco-MiniLM-L-6-v2` as a cross-encoder for reranking top-15 candidates.
• Add explicit LLM prompt instructions for abstaining on out-of-scope questions.

Topics

• Hybrid RAG
• Reciprocal Rank Fusion
• Cross-Encoder Reranking
• LLM Evaluation Frameworks
• Production MLOps
• Semantic Caching

Code references

• prabhaharanv/production-hybrid-rag
• prabhaharanv/production-hybrid-rag

Best for: Machine Learning Engineer, AI Engineer, MLOps Engineer

Related on AIssential

• Build a Hybrid RAG System with FAISS, BM25, LangGraph and Claude Sonnet Model — Hybrid RAG combines semantic and keyword search via Reciprocal Rank Fusion to enhance retrieval accuracy for diverse query types.
• RAG in Production: The Retrieval Failures Nobody Writes About — Production RAG systems require robust engineering to overcome subtle retrieval and context-use failures.
• Building a Production-Ready RAG Pipeline — Only 5 Seats Left — The workshop focuses on advanced techniques for building production-ready Retrieval Augmented Generation (RAG) pipelines.
• Beyond Semantic Search: The Complete Guide to Advanced RAG with Milvus | Sagar Patil — Advanced RAG requires multi-stage retrieval, combining hybrid search with techniques like query rewriting and reranking for production reliability.
• What No RAG Tutorial Tells You Before You Start Building — Effective RAG implementation hinges on robust retrieval strategies, not just LLM capabilities, to prevent confident but incorrect answers.
• Building HITL Feedback RAG: Embeddings, Retrieval, and Reranking — HITL Feedback RAG improves LLM accuracy by dynamically injecting human-curated corrections via a robust retrieval and reranking pipeline.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by LLM on Medium.