UltraRAG

1. Project/Product Overview

! UltraRAG Architecture Diagram

• Figure: UltraRAG overall architecture-Pipeline (process definition), Client (scheduling hub), Server (function execution), UI (interactive demonstration) four-layer collaboration *

2. What can it do mainly (low-code MCP RAG Pipeline architecture detailed explanation, componentized design)

Core Ideas of 2.1 MCP Architecture

The biggest difference in UltraRAG is to completely "MCP" the RAG system ". Based on the Model Context Protocol standard protocol proposed by the Anthropic, it is UltraRAG to standardize and package the functional modules of RAG into independent MCP Servers, each of which exposes the function-level Tool interface to the outside world. By scheduling these servers through the MCP Client, developers can complete complex process orchestration by simply writing YAML configurations.

Four core roles:

-📑Pipeline (process definition) :YAML compiles task logic, defines the execution sequence and business logic of each component, and configures the inference process.

-**🕹Client (Scheduling Center) : Analyze Pipeline configuration, coordinate tool calling and data transfer among servers, and ensure accurate process execution.

-⚙Server (function execution) : Standardize the core functions as independent services, and support the rapid expansion of new modules through simple interfaces.

-🖥️ UI (Interactive Presentation) : Transform the logical key defined by YAML into an intuitive conversational Web interface, significantly improving debugging efficiency and demonstration effect.

2.2 Built-in MCP Server Components

2.3 UltraRAG 3.0 Three Core Breakthroughs

(1) One-click leap from logic to prototype

Provide "WYSIWYG" Pipeline Builder to automatically handle cumbersome interface packaging. Developers focus on logical arrangement, and static code instantly becomes an interactive Demo system. YAML is written and applied without additional UI code.

(2) Full-link white-box transparency-inference tracking visualization

In the traditional RAG framework, the reasoning chain of multiple rounds of dynamic decision-making is hundreds of steps long, and only the back-end log can be checked after errors are made. The UltraRAG 3.0 presents the intermediate state of each loop, branch, and tool call in real time through the "Show Thinking" panel, structured streaming display. When Bad Case appears, directly compare the retrieval slice with the final answer on the interface to quickly judge whether it is "data layer noise" or "model layer illusion".

(3) Built-in intelligent development assistant

A AI assistant embedded in the understanding framework, which assists in generating Pipeline configurations, optimizing Prompt, and interpreting parameter meanings through natural language interaction. For example: "help me modify the current Pipeline, add a Citation module for fact checking", "cut the back end of the generated model to OpenAI, and replace the model with qwen3-32b"-the assistant automatically generates the corresponding YAML configuration.

2.4 Visual RAG IDE

UltraRAG UI is not just a chat interface, but a RAG integrated development environment that integrates orchestration, debugging, and demonstration:

-Canvas Mode: Intuitively drag and drop UI components to assemble complex logic such as Loop and Branch, like building blocks

-Code mode: Directly edit the YAML configuration file and render the Canvas canvas synchronously in real time.

-One-click Build & Verify: Automatic logic self-check and syntax verification during build, and dynamic generation of parameter configuration panel

-Knowledge Base Management: Built-in document upload, parsing, and index management components. You can build custom knowledge bases.

2.5 Multimodal Capability (introduced in v2.1)

-VisRAG Pipeline: Based on the VisRAG paper (ICLR 2025), it realizes a complete closed loop generated from local PDF index to multi-modal retrieval. Combine modeling document image information (charts, formulas, layout structure) with text content to significantly improve the QA capabilities of complex scientific documents.

-Unified Multimodal Interface: The Retriever, Generation, and Evaluation servers all support multimodal tasks and can flexibly connect various visual, text, or cross-modal models.

3. Applicable Scenario

4. Not quite the scene

5. Core Competence List

6. Architecture/deployment/integration approach

6.1 deployment method

Method 1: Source code installation (recommended)

# 安装 uv（快速 Python 包管理器）
pip install uv

# 克隆仓库
git clone https://github.com/OpenBMB/UltraRAG.git --depth 1
cd UltraRAG

# 核心依赖（仅 UI 等基础功能）
uv sync

# 全功能安装（检索+生成+语料+评测）
uv sync --all-extras

# 按需安装
uv sync --extra retriever    # 仅检索模块
uv sync --extra generation   # 仅生成模块

# 激活虚拟环境
source .venv/bin/activate

Method 2: Docker deployment

# 拉取镜像（可选择 CPU / GPU / 全功能版本）
docker pull hdxin2002/ultrarag:v0.3.0-base-cpu
docker pull hdxin2002/ultrarag:v0.3.0-base-gpu
docker pull hdxin2002/ultrarag:v0.3.0

# 启动容器（默认映射 5050 端口）
docker run -it --gpus all -p 5050:5050 <镜像名>
# 浏览器访问 http://localhost:5050 即可使用 UI

6.2 Integrated Ecosystem

6.3 Architecture Diagram

! UltraRAG Architecture

• Figure: UltraRAG four-tier architecture-Pipeline defines business processes, Client parses configuration and schedules, Server layer includes independent services such as Retrieval/Generation/Corpus/Prompt/Benchmark, UI layer provides conversational Web interaction *

How to use #7.

7.1 the simplest Hello World

# examples/sayhello.yaml
name: sayhello
pipeline:
  - step:
      name: greet
      server: sayhello
      tool: greet

# 运行 Pipeline
ultrarag run examples/sayhello.yaml

# 输出：Hello, UltraRAG v3!

7.2 Custom MCP Server (Take SayHello as an Example)

# servers/sayhello/src/sayhello.py
from typing import Dict
from ultrarag.server import UltraRAG_MCP_Server

app = UltraRAG_MCP_Server("sayhello")

@app.tool(output="name->msg")
def greet(name: str) -> Dict[str, str]:
    ret = f"Hello, {name}!"
    app.logger.info(ret)
    return {"msg": ret}

if __name__ == "__main__":
    app.run(transport="stdio")

# servers/sayhello/parameter.yaml
name: UltraRAG v3

7.3 Complex Pipeline example: multi-way recall reordering generation

name: advanced_rag
pipeline:
  - step:
      name: dense_retrieve
      server: retriever
      tool: search_dense
  - step:
      name: sparse_retrieve
      server: retriever
      tool: search_sparse
  - step:
      name: merge_rerank
      server: reranker
      tool: rerank
      inputs:
        - dense_retrieve.results
        - sparse_retrieve.results
  - step:
      name: condition_check
      server: router
      tool: check_confidence
  - step:
      name: generate_with_context
      server: generation
      tool: generate
      inputs:
        - merge_rerank.top_results
      condition: "condition_check.confidence > 0.7"
  - step:
      name: iterative_search
      server: retriever
      tool: iterative_search
      condition: "condition_check.confidence <= 0.7"
      loop:
        max_iterations: 3
        condition: "not_enough_context"

7.4 start visual UI

# 启动 UltraRAG UI（管理员模式）
ultrarag show ui --admin

# 浏览器访问 http://localhost:5050
# 可在 Canvas 和 Code 模式间切换，可视化管理 Pipeline

7.5 evaluation process

# 下载评测数据集
# 配置 Benchmark Pipeline
ultrarag run examples/experiments/eval_benchmark.yaml

# 查看 Case Study 可视化分析
# 在 UI 中深度追踪每个中间输出，辅助分析归因

8. What can I say before sales

8.1 a sentence positioning

UltraRAG = Tsinghua, the world's first MCP architecture low-code RAG framework, uses YAML choreography instead of hard coding, allowing RAG development to return from "writing engineering code" to "designing algorithm logic".

8.2 customer pain points → solutions

8.3 Differentiated Selling Points

vs LangChain / LlamaIndex

vs RAGFlow / Dify / MaxKB

UltraRAG unique three cards:

1. MCP Native Architecture: The only RAG framework in the industry that fully embraces the MCP protocol, seizing the commanding heights of the technology paradigm shift
2. The Strongest Credibility in Academic Circles : Tsinghua THUNLP OpenBMB(MiniCPM / ChatDev Team) has 5 top meetings, and its academic reputation is unparalleled.
3. Zero Distance from Algorithm to Demo : One-click UI generation by Pipeline Builder AI Assistant to get through the last kilometer of "Paper Algorithm → Demonstrable System"

8.4 Customer Value Story Line

STORY LINE A- For University/Lab Heads:

"Your doctoral student is studying a new RAG algorithm, but every time you want to verify an idea, you need to spend 80% of your time building an engineering framework, docking a search library, writing a front-end Demo, and only 20% of your time doing real algorithm innovation. After using the UltraRAG: YAML orchestration logic + MCP Server registering new components + one-click generation of demonstration UI, returning 80% of the project time to algorithm research. And with built-in evaluation benchmarks, paper experimental reoccurrence and horizontal comparisons are no longer a nightmare. This is the Tsinghua team's own efficiency tool."

Story Line B- Technical Leader for Enterprise AI Team:

"Your team has a lot of RAG scenarios to explore: internal knowledge base Q & A, multimodal document understanding, Deep Research report generation... but each scenario is too expensive to build from scratch. The UltraRAG MCP Server is "plug-in"-the knowledge base is docked once, the retrieval strategy can be switched with YAML, and the generation model can be changed with API Key. More importantly, every step of the Pipeline's reasoning is visible and traceable, and the illusion can quickly locate the problem. This is especially valuable for scenarios where the system behavior needs to be explained to the business side."

9. Frequently Asked Customer Questions

10. PoC Recommendations

10.1 PoC target setting

10.2 Recommended PoC Scenarios

Scenario 1: Research RAG Evaluation PoC

-Run standard evaluation with UltraRAG built-in Benchmark

-Show case study visual analysis interface

-Highlights white-box reasoning tracking capabilities

-Customer value: let researchers intuitively feel the "experiment → evaluation → analysis" one-stop experience

Scenario 2: Multimodal Document QA PoC

-Handle PDF documents with diagrams/formulas with VisRAG Pipeline

-Show combined retrieval and answer generation

-Customer Value: Demonstrate the unique capabilities of multimodal RAG

Scenario 3:Deep Research Demo PoC

-Deploy Deep Research Pipeline AgentCPM-Report

-Enter a research topic and automatically generate a million-word research report

-Customer Value: Demonstrate the ability to fully automate "from problem to report" with strong impact

10.3 PoC Key Indicators

10.4 PoC Success Criteria

-✅UltraRAG successfully installed in customer environment (local or Docker)

-✅Build a knowledge base with customer-provided documents (≥ 50 PDFs) and complete QA

-✅The customer team independently completes a Pipeline modification (e. g. adding a reorder step)

-✅Customers recognize the value of white-box debugging (compared to traditional log troubleshooting methods)

11. Risks and Considerations

12. My Pre-Sales Judgment

12.1 comprehensive recommendation degree

12.2 Customer Portrait

Core Target Customers:

1. University NLP/IR Laboratory : Need for Standardized RAG Experimental Framework Evaluation System Paper Reproduction Tool
2. Enterprise AI Research Institute/Innovation Laboratory : Teams that need to quickly transform algorithmic idea into demonstrable systems
3. Start-up team with RAG as core product : flexible Pipeline arrangement capability is needed to optimize product effect
4. Technical decision makers focused on the MCP ecosystem: as a reference for best practices in the RAG area for MCP protocols

Not suitable for direct referrals:

-Enterprise IT departments that aim purely at "building a knowledge base Q & A system" and have no requirements for algorithms.

-Traditional enterprises with limited technical reserves

12.3 policy recommendations

- Short-term (2026 H2) : cut into universities and enterprise research institutes with "RAG experimental platform for scientific research" to avoid head-on confrontation with RAGFlow/Dify in production scenarios

- Mid-term (2027) : If MCP protocol continues to heat up, UltraRAG may become "MCP RAG standard"; It is suggested to implant the cognition of "MCP native architecture" in customer technology selection.

-Long-term: Pay attention to whether the OpenBMB team will launch a commercial version (Face Wall Intelligent ModelBest already has a commercial service system)

13. REFERENCE

• Date of analysis: 2026-07-02 *