TAIJI: MCP-based Multi-Modal Data Analytics on Data Lakes

Chao Zhang, Shaolei Zhang, Quehuan Liu, Sibei Chen, Tong Li, Ju FanarXiv 2025 · DOI · arXiv

Abstract

TAIJI is a novel architecture designed to overcome the "jack-of-all-trades, master-of-none" limitations of monolithic Large Language Models (LLMs) in multi-modal data analytics. By leveraging the Model Context Protocol (MCP), TAIJI offloads specialized tasks to modality-specific servers and implements an iterative "DeepResearch" update loop. Evaluation on multi-modal data lakes shows that TAIJI achieves a 43% reduction in latency and significantly higher accuracy (85% vs. 65%) compared to unified models like GPT-4.

Introduction

Modern enterprise data lakes are heterogeneous environments containing a mix of structured tables, semi-structured JSON, and unstructured media such as images and video. While frontier LLMs like GPT-4 possess cross-modal reasoning capabilities, they face significant hurdles in production: limited expressiveness for complex execution plans, high inference overhead for processing massive unstructured datasets, and data staleness due to fixed training cutoffs.

Comparison of existing monolithic approaches versus the proposed MCP-based architecture.

TAIJI addresses these challenges by shifting from a single, unified LLM to an orchestrated system of specialized models. This approach treats the LLM as a reasoning engine that generates plans, while offloading execution to modular, optimized servers.

The TAIJI Architecture

The core of TAIJI is built on the Model Context Protocol (MCP), an emerging standard for AI agent interaction with external tools and data. The system is composed of several specialized managers and modules designed to handle the lifecycle of a multi-modal query.

Detailed components of the TAIJI system, including the Client Manager, Server Manager, and Augmentor.

Key Components

  • MCP Client Manager: Acts as the central conductor. It utilizes an NL2Operator translator to map natural language intent into a hierarchy of semantic operators. A Query Planner then employs sampling-based cost estimation to determine the most efficient execution path across available resources.
  • Iterative Operator Optimizer: To mitigate query ambiguity, TAIJI enters a feedback loop. It refines queries iteratively until the results meet a predefined confidence threshold, effectively avoiding the "hallucination trap."
  • Multi-Modal MCP Servers: Sub-tasks are dispatched to specialized servers. For instance, vision-specific models (e.g., Qwen2.5-VL) process images, while relational databases like PostgreSQL handle structured data. This modularity ensures that the best tool is used for each specific data type.
  • MCP Augmentor: This module maintains system freshness through a two-fold mechanism: "DeepResearch" for harvesting new data via web crawling and APIs, and Machine Unlearning (via gradient ascent) to remove obsolete or sensitive information without requiring full model retraining.

Experimental Evaluation

TAIJI was evaluated using the Craigslist Furniture Dataset, consisting of 3,000 listings. This dataset requires complex cross-modality reasoning, as it pairs relational metadata (price, category) with unstructured visual data (product images).

Key Results

  • Accuracy & Recall: TAIJI achieved 85% accuracy on complex multi-modal queries, significantly outperforming a GPT-4 baseline which reached only 65%. TAIJI proved far more capable at distinguishing between visually similar items (e.g., a black sofa vs. a black chair) by leveraging specialized vision servers.
  • Efficiency: The system improved end-to-end latency by an average of 43%. Notably, the performance gap widened as data volume increased, demonstrating superior scalability compared to monolithic models.
  • Knowledge Freshness: The DeepResearch module successfully updated the system with real-world information post-dating the base model's training cutoff, ensuring the data lake remains a reliable source of truth.

Discussion and Future Work

TAIJI demonstrates that decoupling reasoning from execution is a powerful paradigm for data lake analytics. By using smaller, specialized models for heavy lifting, the system achieves better performance at lower latency.

However, several areas for future research remain. Security and privacy are paramount when data is distributed across multiple MCP servers. Additionally, optimizing the placement of models and data in hybrid cloud/edge deployments presents a complex optimization challenge. Future work will also focus on developing standardized multi-modal benchmarks to further validate these architectural gains.

Summary updated

Abstract

The variety of data in data lakes presents significant challenges for data analytics, as data scientists must simultaneously analyze multi-modal data, including structured, semi-structured, and unstructured data. While Large Language Models (LLMs) have demonstrated promising capabilities, they still remain inadequate for multi-modal data analytics in terms of accuracy, efficiency, and freshness. First, current natural language (NL) or SQL-like query languages may struggle to precisely and comprehensively capture users' analytical intent. Second, relying on a single unified LLM to process diverse data modalities often leads to substantial inference overhead. Third, data stored in data lakes may be incomplete or outdated, making it essential to integrate external open-domain knowledge to generate timely and relevant analytics results. In this paper, we envision a new multi-modal data analytics system. Specifically, we propose a novel architecture built upon the Model Context Protocol (MCP), an emerging paradigm that enables LLMs to collaborate with knowledgeable agents. First, we define a semantic operator hierarchy tailored for querying multi-modal data in data lakes and develop an AI-agent-powered NL2Operator translator to bridge user intent and analytical execution. Next, we introduce an MCP-based execution framework, in which each MCP server hosts specialized foundation models optimized for specific data modalities. This design enhances both accuracy and efficiency, while supporting high scalability through modular deployment. Finally, we propose a updating mechanism by harnessing the deep research and machine unlearning techniques to refresh the data lakes and LLM knowledges, with the goal of balancing the data freshness and inference efficiency.
BibTeX
@misc{zhang2025taiji,
  author = {Chao Zhang and Shaolei Zhang and Quehuan Liu and Sibei Chen and Tong Li and Ju Fan},
  title = {{TAIJI: MCP-based Multi-Modal Data Analytics on Data Lakes}},
  year = {2025},
  eprint = {2505.11270},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.11270},
}
Figure