Skip to main content
Back

Baidu's TURA Beats a 671B AI Model Using a Tiny 4B Agent

Baidu's production AI search combines RAG with agentic tool-use via MCP servers, DAG planning and a distilled 4B executor serving millions.

  • Baidu's TURA merges RAG with agentic tool-use, serving tens of millions of users in production.
  • Three stages: intent-aware MCP server retrieval, DAG-based task planner, and a distilled agent executor.
  • Distilled Qwen3-4B hits 88.3% tool-calling accuracy vs 82.4% for the 671B Deepseek-V3 teacher.
  • DAG parallel planning cuts latency 44% on complex multi-hop queries with no accuracy loss.
  • Live A/B test: Session Success Rate rose from 55.1% to 64.0% over the RAG baseline.
  • MCP-Bench released with 10,683 annotated queries across 61 MCP servers.

Traditional AI search chokes when you ask it something like "is there a bullet train from Beijing to Shanghai tomorrow with seats left?" The retrieval index only knows about static web pages, not the live inventory sitting behind a ticketing API. Baidu's search team has published TURA, a three-stage framework that fuses RAG with agentic tool calling and has been running in production since May 2025, serving tens of millions of users.

Side-by-side comparison of TURA calling a ticket API versus RAG returning stale web results

The static web problem

Current RAG systems, optimized for retrieving from pre-indexed web documents, are fundamentally incapable of accessing dynamic, real-time information that is not present on a static webpage but must be generated via interaction. Flight seats, train schedules, live inventory, weather at a specific hour, none of that lives in a crawled corpus. Search engines historically patched this with hand-curated components like Google's OneBox or Baidu's Aladdin platform, but those approaches rely on brittle, hand-crafted integrations that are difficult to scale.

TURA replaces those bespoke boxes with a unified agentic pipeline where every information source, static web index or dynamic API, is wrapped as a Model Context Protocol (MCP) server. The core objective is stated formally: maximize answer quality subject to a strict latency budget, since this is a live search product where seconds matter.

Three stages, one pipeline

TURA framework overview showing intent-aware retrieval, DAG planner, and distilled executor

The architecture has three modules that hand off to each other:

  1. Intent-Aware MCP Server Retrieval. An LLM decomposes the raw query into atomic sub-intents, then dense-retrieves candidate MCP servers for each.
  2. DAG-based Task Planner. A planner LLM builds a directed acyclic graph of sub-tasks so independent calls run in parallel.