The first prize is 60,000, and the Higress AI Gateway Development Challenge is officially launched
|
Sep 23, 2025
|
Registration Address:https://competition.atomgit.com/competitionInfo?id=c16f796963021f21711ad25819c664b4#heading-0-0
1. Event Overview
1. Event Background
With the rapid development of Generative AI technology, large language models (LLM) are gradually becoming the core of modern software architecture. This transformation presents new requirements for the underlying infrastructure, especially for API gateways, which serve as the traffic entry point and the core for policy execution. Traditional API gateways mainly handle request routing, authentication, and traffic control, but in the AI era, their role must evolve into an "AI gateway" that natively supports and accelerates AI applications.
Higress is a cloud-native API gateway built on Envoy and Istio, born within Alibaba Group, and validated in large-scale production environments with hundreds of thousands of requests per second. It features dynamic configuration, millisecond-level effectiveness, stateless architecture, and support for various plugin extensions (Wasm), providing crucial support for core AI services such as Tongyi Qianwen from Alibaba. Higress strives to be a bridge connecting AI technology with real-world services, creating the next-generation AI native gateway by offering capabilities such as AI proxies, observability, security protection, and advanced plugins.
Currently, the Higress community has open-sourced a series of AI plugins, providing developers with foundational capabilities. However, to truly unleash the potential of the AI gateway, we need to address the cutting-edge, most challenging technological issues in the AI field. This challenge competition is based on this goal, soliciting solutions from developers worldwide, with a focus on three core directions: Accelerating AI Agent construction, RAG (Retrieval-Augmented Generation) enhancement, and Intelligent Routing. We hope to combine cutting-edge theories from academia with best practices from industry through this competition, enriching Higress's AI plugin ecosystem and driving innovation and evolution in AI infrastructure.
2. Event Information
Event Name: Higress AI Gateway Development Challenge
Event Theme: Building the next-generation intelligent AI gateway
Event Purpose:
Greatly enrich the Higress AI plugin ecosystem by introducing solutions supported by cutting-edge theories and excellent engineering outcomes.
Explore and practice advanced architectures to accelerate AI Agent construction at the gateway layer, optimize the full link of RAG, and achieve intelligent LLM model routing.
Build a vibrant global developer community around the open-source Higress project, attracting top talents to jointly create key infrastructure for the AI era.
Accelerate the transformation of the latest academic research results in artificial intelligence systems into practical, verifiable industrial applications.
Event Time: September 2025 – December 2025
3. Organizing Unit
Organizer: OpenAtom Foundation
Undertaking Unit: Alibaba Cloud | Higress Community
2. Competition Content
This competition has three independent sub-directions, and participants may choose any one of them. Each direction aims to solve the key technological bottlenecks in the current AI gateway field, requiring entries not only to be functionally complete but also to achieve leading levels in technical depth, theoretical basis, and engineering outcomes.
1. Direction One: Accelerating AI Agent Construction
1. Basic Description of the Competition Topic
Current Status and Challenges: The construction of AI Agents is moving from academic exploration to industrial implementation, but developers generally face a significant gap of "from idea to production." Currently, the development process of Agents is highly complex and faces several challenges: the first is integration hell, where securing reliable connections between Agents and the complex legacy APIs and data sources (i.e., "tools") within enterprises requires a large amount of custom development work; next is disruptive development experience, where the logic orchestration of Agents, tool development, and management of the underlying infrastructure (such as gateways and service meshes) are disconnected, slowing down iteration speed; finally, the transition from prototype to production-level Agents imposes extremely high requirements for security, observability, and governance abilities, which is exactly the shortfall of most Agent frameworks.
Competition Objective: This competition topic aims to leverage Higress's core capabilities as an AI native gateway to empower and accelerate the process of AI Agent construction. Participants need to create a solution based on Higress's plugin architecture, LLM API access, and MCP (Model Context Protocol) that allows developers to rapidly build, test, deploy, and manage production-level AI Agents in a zero-code or low-code manner. The final goal is to achieve a rapid transition from an idea to a powerful functional Agent.
Technical Requirements and Architecture Reference:
The format of the entries can be diverse, including innovative AI Agent plugins or comprehensive solutions combining various capabilities of Higress. The design of the solution can refer to the following content (but is not limited to):
Core Framework: The solution must be based on Higress's core capabilities. Utilize its Wasm plugin mechanism as a carrier for Agent "skills" or "tools" to decouple from the core logic of the Agent. The Higress gateway layer is responsible for handling fundamental infrastructure tasks such as security authentication, traffic governance, and API protocol conversion.
Low-code/Zero-code Agent Construction Experience: The solution should provide a low-code/zero-code experience covering the entire process of Agent construction, rather than being limited to tool integration. This includes visual orchestration of Agent logic, definition of workflows, rapid connection and configuration of tools, and one-click deployment and debugging capabilities, allowing developers to focus their energy on the business value of the Agent itself.
Agent Capability Opening and Dynamic Management: The solution should achieve seamless integration with Higress's existing AI open platform Himarket. Developers can not only dynamically manage the various capabilities (skills) of the Agent through the platform (such as hot-swapping, version control, and dynamic configuration) but, more importantly, can encapsulate the overall capabilities of the constructed Agent into standard API services, making them reusable assets that can be easily integrated into other applications through Himarket.
Production-Level Value and Evolution: The solution must demonstrate its value for production use, rather than being a merely demonstration tool for rapid prototype verification (POC). The evaluation will focus on the robustness, observability, and scalability of the solution itself, ensuring that the Agent built on that solution has the potential for rapid evolution into a production environment, capable of bearing real business loads.
2. Evaluation Criteria for the Competition Topic
The evaluation of this competition topic will focus on the innovation of the solution in accelerating Agent construction, the smoothness of the development experience, the comprehensive capabilities of generated Agents, and the production usability of the solution.
Evaluation Criteria | Sub-Evaluation Criteria | Evaluation Standards |
|---|---|---|
Solution Architecture Design (30 points) | Development Experience (15) | A key focus is on whether the solution significantly simplifies the entire process of Agent construction, and whether it truly achieves the goals of "zero-code" or "low-code." The usability and fluidity of the user interface, toolchain, or documentation are core evaluation points. |
Innovation and Value (10) | Does the solution propose novel and effective solutions to the current pain points in Agent construction? How significant is the value reflected in the entire lifecycle of accelerating Agents from idea to realization? | |
Technical Architecture Rationality (5) | Does the solution fully and reasonably utilize core technologies such as Higress plugins, MCP, LLM API, etc.? Is the architecture design clear, modular, and scalable? | |
Solution Code Implementation (50 points) | Himarket Integration and Capability Opening (20) | The focus is on the closeness and effectiveness of the solution's integration with Himarket. Has the Agent capability been successfully encapsulated as a standard API and opened through Himarket? Evaluate the completeness and convenience of it as a reusable asset. |
Comprehensive Agent Capability (GAIA Test Set) (15) | How does the Agent built using the participant's solution perform on the GAIA benchmark test set? The evaluation focuses on the Agent's success rate and accuracy in tasks requiring complex tool calls, web browsing, and reasoning. | |
Functional Completeness (15) | Does the entry fully implement the core functions promised in its design solution, such as one-click API transformation, dynamic management of Agent skills, effective execution of security policies, etc.? | |
Non-Functional Indicators (20 points) | Production-Level Value (10) | A key focus is on whether the solution transcends the prototype (POC) stage and has production usability. Evaluation dimensions include the robustness, observability, scalability of the solution, and whether there is a clear path to evolve into a production environment. |
Performance and Stability (5) | Operating efficiency of the solution, resource consumption, and stability in continuous operation and anomaly handling. | |
Code Quality and Documentation (5) | Code style is clear, comments are adequate, and comprehensive design documentation, architecture diagrams, and user guides (README) are provided. |
2. Direction Two: RAG Enhancement
1. Basic Description of the Competition Topic
Current Status and Challenges: Higress has provided RAG MCP Server capability (implemented based on the golang-filter plugin), supporting connection with Agents to achieve simple RAG retrieval recall capabilities. However, leading industry RAG systems have already surpassed the simple "retrieve-generate" model, evolving into a complex pipeline comprising multiple stages including query optimization, multi-path recall, result reordering, content compression, and self-correction. Existing plugin functions are singular and lack the overall optimization capability for RAG processes.
Competition Objective: This competition topic requires participants to implement a unified and powerful Enhanced RAG capability. Integrate cutting-edge RAG optimization technologies into a modular, configurable Higress MCP Server, significantly improving the accuracy, relevance, and timeliness of content generation in complex query scenarios.
Technical Requirements and Architecture Reference:
The implementation of the plugin must reflect a profound understanding of full-link optimization for RAG, and the design solution can refer to the following content (but is not limited to):
Pre-Retrieval Optimization: Before executing any retrieval operation, the original user query must be processed. Participants should implement relevant technologies, such as:
Query Rewriting and Decomposition: Drawing on research results like
PreQRAG, the plugin should automatically analyze query intentions, rewriting them to enhance retrieval recall rates, or decompose complex questions into multiple independently retrievable sub-questions.Multi-Path & Hybrid Retrieval: The plugin needs to unify the access interface for various data sources (e.g., vector databases, full-text search engines, structured databases, etc.) and support multi-path recall strategies. In particular, it must achieve Hybrid Search, maximizing the advantages of both sparse retrieval (e.g., BM25) and dense retrieval (vector search) to recall relevant documents.
Post-Retrieval Processing: After obtaining an initial list of documents from the data source, fine-tuning must be carried out, including:
Re-ranking: Use stronger models (e.g., Cross-Encoder) or algorithms to reorder the top-K results initially retrieved, enhancing the precision of the rankings.
Context Compression: Based on the latest document compression research, intelligently compress or summarize the retrieved document content to inject more critical information into a limited LLM context.
Corrective Retrieval: The plugin can consider implementing a closed-loop mechanism for Corrective-Action RAG (CRAG) enhancement. This mechanism may include:
A lightweight Retrieval Evaluator to assess the relevance of retrieved documents to the query.
Trigger different actions based on the confidence levels given by the evaluator: if
Correct, refine the document for use; ifIncorrect, discard the document and acquire new knowledge through web search; ifAmbiguous, combine the refined document and web search results.
2. Evaluation Criteria for the Competition Topic
The evaluation of this competition topic will focus on the completeness of the RAG pipeline constructed by participants, the theoretical depth of various optimization technologies, and the final quantifiable improvement in effect.
Evaluation Criteria | Sub-Evaluation Criteria | Evaluation Standards |
|---|---|---|
Solution Architecture Design (30 points) | Completeness and Theoretical Support (15) | Does the solution design completely cover key stages such as pre-retrieval, multi-path retrieval, post-retrieval, and corrective retrieval? Do the technical choices for each stage have sufficient theoretical basis (e.g., citing papers related to CRAG, PreQRAG, etc.)? |
Modularity and Unification (10) | Is the architecture clear and modular, allowing for the independent activation/deactivation or replacement of various optimization technologies? Has a unified abstraction and access to multiple back-end data sources been successfully achieved? | |
Advancement and Innovation (5) | Does the solution propose novel combinations of RAG technologies, or does it have unique implementations in certain aspects (such as the retrieval evaluator)? | |
Solution Code Implementation (50 points) | Pre/Post-Retrieval Implementation (15) | Fully implement at least one pre-retrieval optimization technology (e.g., query rewriting) and one post-retrieval processing technology (e.g., re-ranking). |
Core Logic Implementation of RAG Enhancement (20) | For example, implementing the CRAG mechanism requires a fully functional retrieval evaluator and associated corrective actions. | |
Verifiable Effectiveness (15) | Provide a complete and reproducible evaluation scheme for testing on public datasets (e.g., HotpotQA, Natural Questions, etc.) and provide quantitative indicators (such as retrieval Precision/Recall, generated Factuality Score, etc.) to demonstrate significant improvements compared to baseline RAG. | |
Non-Functional Indicators (20 points) | Code Quality and Documentation (10) | The code quality is high, documentation is clear, and references to academic literature and technical information are clearly marked. |
Performance and Efficiency (10) | The end-to-end processing latency and resource consumption of the entire RAG pipeline. |
3. Direction Three: Intelligent Routing
1. Basic Description of the Competition Topic
Current Status and Challenges: As the application of large models deepens, enterprises and service providers typically deploy a "model matrix" composed of various models rather than relying on a single general-purpose large model. This matrix may include one or more flagship models with top capabilities but high costs (e.g., qwen3-max), as well as multiple fine-tuned vertical small models that perform exceptionally well in specific fields (e.g., code generation, financial report analysis) and are more cost-effective. However, current API gateways predominantly adopt simple, content-agnostic routing strategies (e.g., based on URL paths) that are unable to understand the intrinsic semantics of AI requests. This "semantic-blind" routing approach causes all requests, regardless of simplicity or complexity, to be routed to the same default model, resulting in severe resource misallocation: simple tasks use expensive models leading to cost waste, while complex tasks are sent to under-capable models, causing subpar results.
Competition Objective: This competition topic requires participants to design and implement an intelligent routing plugin for Higress that has semantic awareness. The core goal of this plugin is to be an intelligent scheduler for LLM traffic, capable of real-time parsing of API requests compliant with the OpenAI protocol, deeply understanding their intrinsic task intentions, complexities, and required tools, and then dynamically and accurately forwarding requests to the most suitable models in a predefined model pool. Additionally, the plugin should rely on the gateway's traffic interception capability to implement a "data flywheel" mechanism: automatically generating labeled data from real API interactions for continuous iteration and optimization of the routing model itself.
Technical Requirements and Architecture Reference:
Participants need to design and train a lightweight "smart routing model" and integrate it into the Higress plugin. The design solution can refer to the following content (but is not limited to):
Core of the Routing Model: The core of the plugin is a predictive model capable of quickly classifying or assessing input requests. This model receives the user prompt as input and outputs the identifier of the target model. Participants can explore various implementation paths:
Classifier Method: Train an efficient text classifier (e.g., based on DistilBERT or lighter models) to map different user intents (e.g., "chitchat," "code generation," "summary extraction") to different back-end model services.
Semantic Matching Method: Define a set of embedding vectors of "typical queries" for each back-end model. The plugin calculates the embedding of the new request at runtime and determines the best routing target through vector similarity calculation (e.g., cosine similarity).
Self-Optimizing Data Pipeline: The plugin needs to implement a configurable "data generation mode". In this mode, the plugin can simultaneously send requests to both the "economical model selected by the routing model" and a "benchmark flagship model", recording information such as
{request, model_A_output, model_B_output, latency, cost}to form a high-quality training/evaluation dataset, providing data support for the continuous iteration of the routing model.Performance and Controllability: The routing decision process must be completed at the millisecond level to avoid significant impacts on end-to-end latency. At the same time, the entire system must provide a clear and operable trade-off mechanism that allows operational personnel to flexibly switch between "cost-prioritized" and "quality-prioritized" strategies through simple configuration (e.g., adjusting a confidence threshold).
2. Evaluation Criteria for the Competition Topic
The evaluation of this competition topic will comprehensively assess the degree of intelligence in routing strategies, the actual optimization effects on system performance (cost, latency, accuracy), and the innovation and engineering quality of the overall architecture.
Evaluation Criteria | Sub-Evaluation Criteria | Evaluation Standards |
|---|---|---|
Solution Architecture Design (30 points) | Routing Strategy and Theoretical Support (15) | Is the design of the routing model reasonable, efficient, and based on clear theoretical underpinnings? Is there sufficient justification for the choices of routing strategies (such as classification, semantic matching, etc.)? Are the designs of the data flywheel and self-optimizing mechanisms feasible and innovative? |
Solution Completeness (10) | Does the design comprehensively consider practical issues such as multi-model management and dynamic configuration? Is the provided cost/quality trade-off mechanism well-designed and easy to operate? | |
Advancement and Innovation (5) | Does it propose novel and effective implementation methods for semantic routing, data generation, or model iteration? | |
Solution Code Implementation (50 points) | Core Routing Logic Implementation (20) | Comprehensively and efficiently implement the logic for semantic-aware routing forwarding. The plugin should be able to correctly parse requests and route traffic to various back-end LLMs based on the routing model's decision. |
Verifiable Effectiveness (15) | Provide a complete benchmark testing scheme and scripts. On the given test set, quantitatively assess the plugin's performance in terms of answer accuracy, access latency, and comprehensive cost, demonstrating significant holistic benefits compared to simple polling strategies, | |
Non-Functional Indicators (20 points) | Code Quality and Documentation (10) | Code style is clear, comments are sufficient, and documentation is detailed, including architectural diagrams, design decisions, and user guides. |
Performance and Stability (10) | The latency introduced by the routing plugin itself is minimal. Under concurrent requests, the system runs stably, and resource consumption is reasonable. |
3. Competition Mechanism
1. Competition Format
Participants can compete as individuals or teams, with no more than 5 members per team.
2. Participants
Open to enterprise developers, students, researchers, and individual developers globally, regardless of nationality or age.
3. Registration
Participating teams must register uniformly through the official website of the Third Open Atom Competition; specific registration channels will be announced later.
4. Preliminary Submission of Work
Submission Content: A complete project design and implementation principle introduction document, project source code hosted on the AtomGit platform, and a video clearly demonstrating the core functions and performance testing process of the work must be submitted.
Submission Specifications: The competition entries must be original and not infringe on any third party's intellectual property rights. If using open source code, the applicable open source licensing agreements must be followed, and the source, agreements, and dependencies must be clearly stated in the documentation.
5. Final Submission of Work
Teams advancing to the finals must prepare and submit the final version of the defense PPT and the fully reproducible source code. The finals will be conducted in an offline roadshow format, where participating teams need to run the code on-site and showcase the results of their work.
4. Review Mechanism and Prize Setting
1. Review Rules
The competition will form an expert review committee composed of technical experts from Alibaba Cloud, key contributors from the Higress community, and renowned scholars from academia. The preliminary stage will use online material reviews, and the finals will use offline roadshow defenses.
2. Supplementary Explanation for Prize Assessment
To ensure the authority and technical content of the awards, there are strict basic requirements for each award's evaluation:
First Prize: A score of 90 or above is required
Second Prize: A score of 85 or above is required
Third Prize: A score of 80 or above is required
Open Source Contribution Award: A score of 60 or above is required.
Note: One first prize and two second prizes will total three winners, with one winner selected from each of the three sub-directions based on comprehensive evaluation.
3. Prize Setting
The total prize pool for this competition is 200,000 RMB, and the specific prize settings are as follows:
Award | Prize (RMB) | Quantity |
|---|---|---|
First Prize | 60,000 RMB | 1 |
Second Prize | 50,000 RMB | 2 |
Third Prize | 10,000 RMB | 3 |
Open Source Contribution Award | 2,000 RMB | 5 |
Total | 200,000 RMB | 11 |
5. Competition Support
1. Official Materials
Higress Project Codebase: https://github.com/alibaba/higress
Higress AI Gateway Official Website: https://higress.ai/
2. Contact Us
To help participants better understand the competition topics and choose directions, the competition organizing committee offers pre-competition consultations. Those with questions about registration or technical choices for each direction can contact the organizing committee via email: zty98751@alibaba-inc.com.
