How Tencent Cloud ES Powers RAG with Hybrid Search and Massive Vector Optimizations

Introduction

Amid the LLM‑driven revolution, tightly coupling search with large models has become essential for knowledge advancement. Elasticsearch (ES), the most popular open‑source search engine, leverages its mature text‑search capabilities and powerful vector retrieval to enable more accurate, comprehensive, and intelligent search.

Why Hybrid Search?

Traditional keyword search excels at precision but lacks semantic understanding, while pure vector search offers strong semantic matching but suffers from lower precision and inability to handle exact keywords. A hybrid approach that combines both retrieves the best of each world.

Core Retrieval Models

Keyword‑based inverted index (Lucene, BM25)

Probabilistic models (BM25, TF‑IDF)

Vector space models (HNSW, ANN)

These models are layered to support multi‑route recall, where text relevance, vector similarity, and category relevance are each scored and then fused.

Tencent Cloud ES One‑Stop RAG Solution

The platform provides a complete pipeline: data ingestion → tokenization & segmentation → embedding generation (custom or built‑in models) → both text and vector indexing → hybrid retrieval → prompt assembly → LLM generation. It also offers built‑in machine‑learning nodes for model deployment, Kibana for debugging, and security‑aware high‑availability features.

Performance Optimizations

To handle billion‑scale vector datasets with sub‑100 ms latency, ES integrates a dynamic memory‑MMAP strategy: hot indexes stay in RAM, while the majority are stored on disk and accessed via a pre‑loaded MMapFS layer. Additional techniques include vector quantization, efficient file encoding, and a custom HNSW implementation that reduces memory usage to 1/10‑1/20 of the original while preserving recall.

These optimizations cut memory consumption by ~80 % and boost query throughput 5‑10× compared with the open‑source baseline.

Real‑World Case Study

A digital‑book platform with over a billion vectors adopted the ES RAG stack. By uploading a fine‑tuned embedding model to ES, the system performed simultaneous text and vector search, assembled prompts, and generated answers via a large model, dramatically lowering cost, simplifying operations, and meeting 100 ms latency at massive scale.

Result Re‑ranking

ES uses a rank‑by‑inverse‑position fusion algorithm and optional LTR/Reranker models (including LLM‑based rerankers) to produce a final ordered list of high‑value results.

Conclusion

Tencent Cloud ES delivers a production‑grade, hybrid search‑RAG platform that bridges traditional information retrieval and generative AI, offering scalability, performance, and ease of integration for enterprise knowledge‑base applications.