How DeepSearch Elevates RAG: From RAG 1.0 to a Multi‑Agent AI Search Engine

Amid the rapid growth of large language model (LLM) technology, Alibaba Cloud OpenSearch launched an LLM‑enabled version in 2023 and has spent two years iterating its Retrieval‑Augmented Generation (RAG) capabilities, progressing from RAG 1.0 to RAG 2.0.

From RAG 1.0 to RAG 2.0

RAG 1.0 follows a single‑step retrieve‑read‑reason workflow, which is fast but struggles with multi‑step reasoning and cross‑document queries. RAG 1.5 introduces a ReAct‑based single‑agent approach that adapts retrieval over multiple turns, yet couples reasoning with generation and suffers from prompt rigidity and context bloat. RAG 2.0, embodied by DeepSearch, adopts a plan‑based multi‑agent system that iteratively plans, searches, reads, and reflects to converge on optimal answers.

DeepSearch Architecture

DeepSearch consists of several specialized agents:

Problem Planning Agent – generates high‑level plans, manages iteration limits, updates global memory, and enforces explicit termination conditions.

Clarification Agent – disambiguates user intent to improve retrieval precision.

Search Agent – performs mixed retrieval (vector, text, and external sources), filters results, and extracts key information.

Summarization Agent – assembles the final answer, supports multiple LLM back‑ends, and outputs structured formats (text, tables, images).

The system also features a modular tool layer (search, web, NL2SQL, graph queries) and supports multiple knowledge sources: an internally built vector store, user‑provided Elasticsearch indices, and live web data.

Implementation Highlights

Multi‑agent registration mechanism for unified management.

Agent collaboration workflow driven by the Planning Agent.

Independent tool services with generic interfaces.

Support for various LLMs (fine‑tuned Qwen, open‑source Qwen, DeepSeek) and prompt optimizations for Chinese and English.

Scalable architecture allowing future tool and modality extensions.

Performance Evaluation

Benchmarks on public datasets show that DeepSearch’s full‑hit recall improves markedly with question complexity, especially beyond three hops, outperforming traditional RAG methods. On the xBench‑DeepSearch leaderboard, the system achieved a 63% full‑hit rate over five iterations.

Conclusion

DeepSearch represents a significant leap from RAG 1.0 to RAG 2.0, turning AI search from simple retrieval into iterative reasoning. It demonstrates strong potential for enterprise knowledge management, intelligent customer service, and technical documentation Q&A, and sets a foundation for future advances in AI‑driven search.