UltraRAG 3.0 is here: no more blind-box development, every line of reasoning logic in plain sight
“Validating an algorithm prototype takes a week; building a usable system takes months.” It sounds like a joke, but it is the reality every algorithm engineer has to face.
Today, THUNLP at Tsinghua University, NEUIR at Northeastern University, OpenBMB, ModelBest, and AI9Stars jointly release UltraRAG 3.0. Built for these pain points, it is a framework designed to understand researchers and developers, with three core advantages:
From logic to prototype in one click, letting algorithm engineers get back to algorithms: a what-you-see-is-what-you-get pipeline builder takes over tedious interface wrapping. Focus on orchestrating logic, and static code instantly becomes an interactive demo system.
Full-pipeline white-boxing with pixel-level visualization of reasoning traces: a transparent reasoning verification window shows every loop, branch, and decision the model makes in complex long-chain tasks, in real time.
A built-in intelligent development assistant, your interactive development guide: an embedded AI assistant that understands the framework helps generate pipeline configurations and optimize prompts through natural language, greatly lowering the learning curve.
This article includes video content. Please view it in the original WeChat post.
Get started in 30 seconds — launch the UltraRAG UI in one click
We welcome your suggestions in the Issues section!
➤ Repository
🔗https://github.com/OpenBMB/UltraRAG
➤ Tutorials
🔗 https://ultrarag.openbmb.cn/
Logic as application: zero distance from orchestration to interaction: the endpoint of an algorithm is no longer a cold console log. UltraRAG 3.0 automates tedious interface wrapping and parameter wiring, so the moment logic orchestration is done, an interactive demo interface is generated alongside it:
Configuration as application: just define the pipeline’s YAML config file, and the framework parses it and turns it into a standard interactive demo.
A dual-mode builder: to balance usability and flexibility, we built a construction engine that keeps the visual canvas and code in sync in real time:
Canvas mode: assemble complex logic such as Loop and Branch intuitively from UI components, like building blocks.
Code mode: edit the YAML config file directly and watch the canvas update in real time, giving developers precise control over parameter tweaks.
One-click build and validation: when construction is done, click the “Build” button at the top. The system runs logic self-checks and syntax validation automatically and dynamically generates a parameter configuration panel. The moment parameters are ready, static algorithm logic becomes an interactive system — what you write is what you get, and what you get is what you use.
UltraRAG Pipeline Builder — seamless flow from pipeline construction to parameter configuration to interactive validation
Rejecting the black box: making the reasoning path of complex RAG visible
As RAG evolves from simple single-pass retrieval to multi-turn dynamic decision-making, reasoning chains often run to hundreds of steps. Without intermediate-state monitoring, debugging is like starting over in thick fog, and locating errors comes down to guesswork. UltraRAG 3.0 redefines the Chat interface — it is not only the entry point for user interaction but also a window for logic validation. We know that for developers, knowing “what the result is” is far from enough; seeing “how the result came about” is the key to optimization.
Through the “Show Thinking” panel, we visualize the system’s entire “thinking” process at pixel level in real time — from complex loops and branches to individual tool calls, all intermediate states are streamed in structured form. Even for complex long-flow tasks like DeepResearch, developers can follow execution progress in real time, so the process is no longer a pitch-black wait. When a bad case appears, developers no longer need to dig through backend logs; they can compare retrieval chunks with the final answer directly in the interface and quickly judge whether the problem lies in “noise at the data layer” or “hallucination at the model layer,” greatly shortening optimization cycles.
Here we use two typical scenarios from the AgentCPM-Report workflow to show white-box debugging in practice:
Case 1: Medical consultation
Case 2: Analysis of historical facts from the Three Kingdoms
Freeing custom development from the framework trap
Trying a new algorithm logic often means digging into the framework’s internals and rewriting a large number of inherited classes. To achieve 10% core algorithmic innovation, you end up paying 90% of the cost learning the framework.
UltraRAG 3.0 embeds the entire development documentation and best practices into the framework’s built-in intelligent assistant. It may not write an entire project for you the way Cursor does, but it is absolutely the most UltraRAG-savvy assistant you can get. Through natural language interaction, it removes the cognitive friction between “reading docs” and “writing configs”:
Configuration generation: just describe what you need (for example, “I want a pipeline with multi-path recall and reranking”), and the assistant generates a standard pipeline structure draft that you can use after minor tweaks.
Prompt tuning: the assistant offers targeted prompt optimization suggestions based on the current task context, quickly adapting to specific business scenarios.
Understanding support: don’t understand a parameter or a piece of logic? No need to jump to a browser and dig through the official docs — just ask and get development advice and code examples, so coding is never interrupted.
Hands-on demo: what can it do for you? Here we show four real interaction scenarios to see how it turns natural language into executable logic:
1. Structure adjustment: modify a pipeline with one sentence
User: “Please modify the current pipeline to add a Citation module so that generated content can be fact-checked.”
2. Scenario adaptation: targeted prompt optimization
User: “I need to optimize the current prompt for the legal domain. Please adjust it so the generated answers use more professional and accurate terminology and reasoning in this field.”
3. Configuration changes: tweak low-level parameters with ease
User: “I want to switch the generation backend configuration. Please change the generation model backend to OpenAI, the model name to qwen3-32b, and deploy the API service on port 65503.”
4. Free-form tuning: a shortcut from concept to implementation
User: “I want to redesign my RAG pipeline based on this paper: https://arxiv.org/pdf/2410.08821 (DeepNote). Please analyze the core ideas in the paper and help me build a similar pipeline architecture.”
Full workflow demo of assistant-aided development
About the team and an open invitation
THUNLP’s Knowledge Augmentation Research Group brings together researchers, engineers, and students from the Tsinghua University NLP lab, NEUIR at Northeastern University, OpenBMB, ModelBest, and AI9Stars — a group of explorers passionate about RAG technology. More students are welcome to join us!
https://nlp.csai.tsinghua.edu.cn/job/29
Open source thrives on community. Whether you want to complain about a bug or share a bold new feature idea, join our developer community and help make RAG development cooler and simpler!