Spark Program | Nervos Brain - A Global Developer Onboarding Engine and Cross-Language Hub Powered by Agentic RAG

English
,
1

Project Name

Nervos Brain — A Global Developer Onboarding Engine and Cross-Language Hub Powered by Agentic RAG

Team / Individual Profile & Contact Info

  • Team / Lead: Yunhe Su (Independent Developer)
  • Background: An open-source developer with a profound technical background in AI. Proficient in Python, PyTorch, and cutting-edge AIGC technologies (LLM, Agent, RAG). Previously independently developed automated workflows utilizing Large Language Models and Chain of Thought (CoT), and published highly regarded AI models and ComfyUI plugins on Civitai as an open-source creator. Possesses a deep understanding of CKB’s underlying technologies (UTXO model, Fiber Network).
  • GitHub Profile: iris-Neko (irisNeko) · GitHub
  • Civitai Published Works: ControlNet ASCII Generator
  • Corresponding GitHub Open-Source Project: ComfyUI_ascii_art
  • Contact Info:
    • Telegram: @IrisNeko_QvQ
    • Discord: @iris_neko_

Project Description

The Problem (Market Pain Points & Ecosystem Gaps): The CKB ecosystem is currently at a critical turning point toward mass application but faces two hard bottlenecks hindering growth:

  1. High Onboarding Friction for Developers: CKB’s UTXO/Cell model and RISC-V VM are extremely hardcore. Existing technical support relies heavily on manual troubleshooting by the DevRel team, which is slow and unscalable, leading to the churn of potential developers.
  2. Language and Platform Silos in Technical Discussions: A massive amount of in-depth discussions regarding underlying design trade-offs are buried in historical archives on Nervos Talk. The English and Chinese communities lack real-time synchronization, preventing global developers from efficiently sharing high-value knowledge.

The Solution (Nervos Brain): This project aims to develop a proactive intelligent engine integrated into Discord, Telegram, and technical forums:

  • Dedicated MCP Server: Breaking the limits of traditional bots, the AI directly reads CKB’s latest RFCs, CCC (CKBer’s Codebase) documentation, and historical bug records to assist developers in writing code.
  • Multi-step Reasoning Agentic RAG: Leverages Chain of Thought (CoT) to break down complex queries (e.g., “how to open a channel on Fiber”), synthesizes various documents and community troubleshooting guides, and generates step-by-step practical tutorials.
  • Cross-Language TL;DR Engine: Automatically patrols core communities and GitHub, extracts the core logic of lengthy, hardcore discussions, generates bilingual (English and Chinese) summaries, and pushes them out to proactively foster high-quality community discussions.

Expected Deliverables

MVP Core

  • Infrastructure: A permanently open-source CKB AI knowledge base and Agent system, available for reuse by any subsequent wallets or developer tools within the ecosystem.
    • Vector DB + Graph Planning (Core knowledge base and graph-structured knowledge retrieval framework, including core Nervos documentation and examples).
    • Bot framework based on Agentic RAG.
    • Nervos Talk MCP tool to fulfill basic social functions.
    • Discord Bot / TG Bot (Final delivery depends on specific development progress).
  • Code Repository: Delivery of a complete, open-source codebase.
  • Documentation: Detailed deployment guides and user manuals.

Technical Details

Abandoning the high development and maintenance overhead of traditional heavy graph databases, we adopt an agile architecture of “Lightweight Vector Retrieval + LLM Native Tool Calling” to ensure high-quality delivery within 8 weeks.

  • Vector DB: Utilizing Qdrant (Serverless). It is maintenance-free, lightweight, and extremely fast. Rather than blindly relying on complex graph connections, we attach precise Payload metadata tags and vectorized indexes to CKB documents. When the Agent calls the retrieval tool, it can autonomously pass in filtering conditions to achieve precise hybrid search.
  • Agent Reasoning & Orchestration: Moving away from passive “one-shot retrieve-and-generate,” we adopt a ReAct (Reason + Act) architecture based on LangGraph or LLM native Function Calling. We break down CKB knowledge into discrete tools (e.g., Search_CKB_RFCs, Query_Nervos_Talk_History, Read_CCC_Codebase). The Agent actively decomposes developers’ complex questions, autonomously decides which tools to call and documents to consult, and proactively performs graph-structured searches (multi-hop reasoning) when information is insufficient, thereby maximizing the value extracted from the knowledge base.
  • Embedding Model Selection: Utilizing text-embedding-3-large (OpenAI). It has an extremely low unit cost and excellent bilingual cross-language representation and code comprehension capabilities, making it the optimal solution for handling CKB’s mixed English-Chinese technical discussions.
  • Interaction Layer: An extensible, multi-routing Agentic Bot framework, coupled with an MCP system that interacts with Nervos Talk forum data in real-time.

Funding Required

**Why apply for the maximum grant cap ($2,000)?**A conventional budget of $1,000-$1,500 usually only supports a lightweight Bot with “single model + basic vector retrieval.” As ecosystem infrastructure, Nervos Brain entails higher engineering complexity: it requires building a hybrid retrieval foundation (Knowledge Graph + Vector DB) and developing a dedicated MCP server. Furthermore, parsing hardcore, lengthy CKB documents and performing real-time cross-language distillation will incur massive token consumption. The $2,000 cap is a necessary threshold to ensure in-depth technical execution and stable system deployment.

**LLM API Strategy (Based on X-MAS Theory):**This project does not rely on locally deployed open-source models; instead, it fully utilizes cloud APIs. Based on the X-MAS (Multi-Agent System) theory, the Agent will dynamically route tasks to achieve the optimal “Performance-Cost” ratio:

  • Basic Interaction & Intent Recognition: Routed to highly cost-effective models (e.g., GPT-5-mini / Haiku 4.5) to handle daily, lightweight requests.
  • Long Context & Cross-Language Summarization: Routed to large-context models (e.g., Gemini 3.1 Pro / Claude 4.6 Sonnet), focusing on parsing lengthy RFC/CCC documents and historical Nervos Talk archives.
  • Deep Reasoning & Code Generation: Routed to top-tier models (e.g., GPT-5.3 / Gemini 3.1 Pro), utilizing CoT to tackle hardcore underlying architecture or smart contract coding issues.

Budget Breakdown:

  • $700 (35%): Retrieval Infrastructure R&D. Scraping and cleaning CKB core documents, building the hybrid retrieval foundation, and developing the dedicated MCP server.
  • $600 (30%): Agent Architecture Development. Developing the X-MAS multi-model routing system, Prompt tuning (eliminating hallucinations regarding CKB-specific terminology), and deploying automated cross-language patrol scripts.
  • $400 (20%): LLM API Costs (8-week estimate). Covers Embedding consumption for full data initialization, as well as the massive interactive token costs incurred during development, debugging, and the seed user beta testing phase.
  • $300 (15%): Servers & Delivery. Cloud server / Vector DB rental fees during the beta phase, final open-source documentation writing, and project closure delivery.

Estimated Timeline

**8 Weeks (approx. 2 months).**​Note: Adhering to the “small steps, fast iteration” principle of the Spark program, this phase is highly focused on developing the core engine and launching the multi-language push feature within a short cycle.

Clear To-Do List

Week 1-3: Data Acquisition and Retrieval Foundation Setup

  • Scrape and clean CKB whitepapers, RFCs, CCC docs, and historical top-tier long posts from Nervos Talk.
  • Deploy the Qdrant vector database, complete full-data Embedding, and establish indexes based on Payload metadata tags.
  • Develop Agent dynamic routing and multi-tool calling logic based on the LangGraph / ReAct architecture.

Week 4-6: Agent Engine Development and Cross-Language Feature Launch

  • Fine-tune dedicated Prompts and Chain of Thought (CoT) to resolve hallucination issues regarding CKB-specific terminology and underlying concepts.
  • Develop the dedicated Nervos Talk MCP server, and complete the beta environment deployment and integration testing of the basic Bot interaction framework.
  • Beta Prep: Deploy telemetry/tracking events and quick feedback mechanisms within the Bot interaction interface.

Week 7-8: Community Beta Testing, Iterative Optimization, and Open-Source Delivery

  • Launch Beta: Recruit 10+ experienced developers from Discord/TG/forums for real-world scenario testing.
  • Agile Iteration: Conduct weekly Triages (issue routing), review the Agent’s internal Reasoning Traces, and intensively optimize RAG weights and Prompts based on Bad Cases.
  • Questionnaire & Evaluation: Distribute evaluation questionnaires to beta users to compile statistics on core metrics like resolution rate, response latency, and CSAT.
  • Project Closure Delivery: Organize the open-source codebase, write exhaustive deployment/usage documentation, and submit a final report on Nervos Talk including detailed test data and conversion rates.

Alignment with the CKB Ecosystem

  • Meets Actual Ecosystem Needs: Directly addresses CKB’s current pain points of “high development barriers” and “bottlenecks in official human support.” Based on sample estimations of historical logs from the core CKB developer community, data shows a massive concentration of new developer pain points in highly repetitive “cold start” phases—such as environment configuration, testnet faucets, RPC errors, and fundamental concepts (e.g., the difference between Type/Lock Scripts, Cell capacity calculation). Estimating from data analysis and previous automated workflow projects, once deployed, the system is expected to intercept and automatically resolve over 50% of basic-to-intermediate development questions in the community. This will drastically reduce the “Time-to-First-Hello-World” for developers, directly improving conversion rates for hackathons and bounties.
  • Provides Unique Competitive Advantages: Tailored specifically for CKB’s unique Cell model and hardcore geek culture. Unlike generic LLMs on the market, Nervos Brain deeply integrates CKB’s underlying logic and historical community consensus. It is not just a Q&A tool, but an ecosystem hub capable of bridging language barriers, accumulating, and distributing CKB’s core technical assets.
  • Future Planning Note: If this Spark MVP project is successfully validated and receives positive community feedback, we plan to upgrade it and apply for long-term Community Fund DAO funding to cover continuous operations and governance tracking across the broader CKB ecosystem.

Community Beta Testing & MVP Validation Plan (User Study & Evaluation)

To ensure the practical utility of Nervos Brain and avoid building in a vacuum, during the Week 7-8 beta testing phase, we will reference academic standard User Study methodologies. We will employ a closed-loop mechanism of “Telemetry Data + Agile Feedback” to evaluate and optimize the system.

1. Core Success Metrics

We will define MVP success through the following quantitative and qualitative metrics:

  • Accuracy & Resolution: * Target: Issue resolution rate of over 50% without human intervention.
    • Measurement: Introduce a quick feedback mechanism at the end of Bot replies (e.g., :+1:/:-1: or :white_check_mark:/:cross_mark: buttons in Discord/TG). A :+1: is considered successfully resolved.
  • Latency: * Target: Basic retrieval response < 15 seconds; complex Agent tasks involving multi-hop reasoning and code generation < 60 seconds.
  • User Satisfaction: * Target: An overall Customer Satisfaction (CSAT) score of 3.0/5.0 or higher from beta users.

2. Testers & Channels

  • Scale: Recruit 10+ testers with relevant development experience.
  • Channels: Targeted invitations to active users in the Discord dev-chat channel, high-frequency posters on the Nervos Talk forum, and public beta invitations issued via official Telegram developer groups.

3. Feedback Loop

  • Implicit Feedback (Telemetry): The backend will record all conversations marked with a :-1: (Downvote). We will conduct weekly Triages to locate whether failures are due to “retrieval misses,” “Prompt hallucinations,” or “tool calling errors” by reviewing the Agent’s internal Reasoning Trace.
  • Explicit Feedback (User Interview/Survey): Before the end of Week 8, send a structured, minimalist questionnaire (taking < 3 minutes) to beta users to collect their subjective evaluations on “answer professionalism,” “tone naturalness,” and “code usability.”
  • Iterative Integration: Based on the collected Bad Cases, conduct a final round of RAG knowledge base weight fine-tuning and Prompt optimization before final delivery. The detailed test data and conversion rates will be included in the final project closure report.
5 Likes
2

项目名称

Nervos Brain —— 基于 Agentic RAG 的全球开发者 Onboarding 引擎与跨语言中枢

团队 / 个人简介及联系方式

  • 团队/负责人：苏云禾（独立开发者）

  • 背景简介：具备深厚 AI 技术背景的开源开发者。熟练掌握 Python、PyTorch 及前沿 AIGC 技术（LLM, Agent, RAG），曾独立开发基于大模型与 CoT（思维链）的自动化工作流，并作为开源作者在 Civitai 发布广受关注的 AI 模型与 ComfyUI 插件。深入理解 CKB 底层技术（UTXO 模型、Fiber Network)。

  • GitHub Profile: iris-Neko (irisNeko) · GitHub

  • Civitai Published Works: ControlNet ASCII Generator

  • Corresponding GitHub Open-Source Project: ComfyUI_ascii_art

  • 联系方式：

    • Telegram：@IrisNeko_QvQ

    • Discord: @iris_neko_

项目描述

问题（市场痛点与生态缺口）： 当前 CKB 生态正处于向大规模应用转型的关键期，但面临两个阻碍增长的硬性缺口：

  1. 开发者上手摩擦力大：CKB 的 UTXO/Cell 模型与 RISC-V VM 极其硬核，现有技术支持高度依赖 DevRel 团队的人工答疑，响应慢且无法规模化，导致潜在开发者流失。

  2. 技术讨论的语言与平台孤岛：大量深度的底层设计权衡（Trade-offs）探讨沉淀在 Nervos Talk 的历史长帖中。中英文社区缺乏实时同步，全球开发者无法高效共享高价值知识。

解决方案（Nervos Brain）： 本项目旨在开发一个集成于 Discord、Telegram 及技术论坛的主动型智能引擎：

  1. 专属 MCP 服务器：打破传统 Bot 限制，AI 直接读取 CKB 最新的 RFC、CCC (CKBer’s Codebase) 文档和历史 Bug 记录，辅助开发者编写代码。

  2. 多步推理 Agentic RAG：利用思维链（CoT）将复杂问题（如“如何在 Fiber 上开启通道”）拆解，综合多方文档和社区避坑指南，生成 Step-by-Step 实战教程。

  3. 跨语言 TL; DR 引擎：自动巡逻核心社区与 GitHub，提取长篇硬核讨论的核心逻辑，生成中英双语摘要并推送，自主促成社区高质量探讨。

预期交付成果

MVP Core

  • 基础设施：一套永久开源的 CKB AI 知识库与Agent系统，可供后续生态内所有钱包或开发者工具复用。

    • vector DB + Graph Planning （核心知识库与图结构知识检索框架，包括Nervos核心文档与示例）

    • 基于 Agentic RAG 的Bot框架

    • Nervos Talk MCP工具，完成基础社交功能

    • Discord Bot / TG Bot （根据具体进展确定是否交付）

  • 代码仓库：交付完整的开源代码库。

  • 文档说明：详细的部署指南与用户使用文档。

Technical Details

摒弃传统重度图数据库带来的高昂开发与维护开销，采用“轻量级向量检索 + 大模型原生 Tool Calling”的敏捷架构，确保 8 周内高质量交付。

  • 向量数据库 (Vector DB)：采用 Qdrant (Serverless)。免运维、轻量且极快。我们不盲目依赖复杂的图谱连线，而是通过为 CKB 文档打上精准的 Payload 元数据标签与向量化索引。Agent 在调用检索工具时，可以自主传入过滤条件，实现精准的混合搜索。

  • Agent 推理与控制层 (Orchestration)：放弃被动式的“一次性检索+生成”，采用基于 LangGraph 或大模型原生 Function Calling 的 ReAct (Reason + Act) 架构。我们将 CKB 的知识拆分为不同的工具（如 Search_CKB_RFCs, Query_Nervos_Talk_History, Read_CCC_Codebase）。Agent 会主动拆解开发者的复杂问题，自主决定调用哪些工具、查阅哪些文档，并在信息不足时主动进行图结构搜索（多跳推理），从而最大化榨取知识库的价值。

  • Embedding 模型选型：采用 text-embedding-3-large (OpenAI)。单价极低，且具备出色的中英文双语跨语言表征与代码理解能力，是处理 CKB 中英混合技术讨论的最优解。

  • 交互层：一套可拓展的多路由 Agentic Bot 框架，以及与 Nervos Talk 论坛数据实时交互的MCP系统。

所需资金

为何申请顶格资助（$2,000）？

常规的 $1,000-$1,500 预算通常只够支撑“单模型+基础向量检索”的轻量级 Bot。Nervos Brain 作为生态基础设施，工程复杂度更高：需构建知识图谱 +向量数据库的混合检索底座，并开发专属 MCP 服务器。同时，解析 CKB 硬核长文档与跨语言实时提炼，会产生庞大的 Token 消耗。2,000 美元是保障深度技术攻坚与系统稳定落地的必要门槛。

LLM API 策略（基于 X-MAS 理论）：

本项目不采用本地部署开源模型，而是全面调用云端 API。基于 X-MAS（多模型协作 Agent 系统）理论，Agent 会动态路由任务以实现“效果-成本”的最优解：

  • 基础交互与意图识别：路由至高性价比模型（如 GPT-5-mini / Haiku 4.5），处理日常轻量请求。

  • 长文本与跨语言摘要：路由至长窗口模型（如 Gemini 3.1 Pro / Claude 4.6 Sonnet），专注解析 RFC/CCC 冗长文档与 Nervos Talk 历史长帖。

  • 深度推理与代码生成：路由至顶尖模型（如 GPT-5.3 / Gemini 3.1 Pro），利用 CoT 处理极客级的底层架构或合约代码问题。

资金 Breakdown：

  • $700 (35%)：检索基建研发。CKB 核心文档的抓取与清洗、混合检索底座搭建、以及专属 MCP 服务器开发。

  • $600 (30%)：Agent 架构开发。X-MAS 多模型路由系统开发、Prompt 调优（消除 CKB 专有名词幻觉）、自动化跨语言巡逻脚本部署。

  • $400 (20%)： LLM API 成本（8 周预估）。涵盖全量数据初始化的 Embedding 消耗，以及开发调试与种子用户内测期产生的大量交互 Token 成本。

  • $300 (15%)：服务器与交付。内测期的云服务器/向量数据库租赁费用，以及最终的开源文档编写与结项交付。

预计完成时间

8 周（约 2 个月） 注：遵循 Spark 计划“小步快跑”原则，本阶段高度聚焦于短周期内核心引擎的开发与多语言推送的跑通。

清晰的 To-do list

Week 1-3：数据获取与检索基建搭建

  • 抓取并清洗 CKB 白皮书、RFCs、CCC 文档及 Nervos Talk 历年精华帖子长文。

  • 部署 Qdrant 向量数据库，完成全量数据的 Embedding，并建立基于 Payload 元数据标签的索引。

  • 基于 LangGraph / ReAct 架构，开发 Agent 动态路由与多工具调用逻辑。

Week 4-6：Agent 引擎开发与跨语言功能上线

  • 调优专属 Prompt 与思维链（CoT），解决 CKB 专有名词及底层概念的幻觉问题。

  • 开发专属 Nervos Talk MCP 服务器，并完成基础版 Bot 交互框架的内测环境部署与联调。

  • 内测准备：在 Bot 交互界面部署埋点与快捷反馈机制。

Week 7-8：社区内测、迭代优化与开源交付

  • 启动内测：从 Discord/TG/论坛 招募 10 名以上有经验的开发者进行真实场景测试。

  • 敏捷迭代：每周进行 Triage（问题分流），复盘 Agent 内部推理轨迹（Reasoning Trace），针对 Bad Cases 集中优化 RAG 权重与 Prompt。

  • 问卷与评估：向内测用户发放评估问卷，统计解决率、响应延迟及 CSAT 等核心指标达成情况。

  • 结项交付：整理开源代码库，撰写详尽的部署/使用文档，并在 Nervos Talk 提交包含详细测试数据与转化率的结项报告。

与 CKB 生态的关联性

  • 满足 CKB 生态实际需求：直击 CKB 当前“开发门槛高”、“官方支持人力遇颈”的痛点。对现阶段 CKB 核心开发者社群历史记录的抽样估算。数据显示，新开发者的大量痛点高度集中在环境配置、测试网水龙头、RPC 报错以及基础概念（如 Type/Lock Script 区别、Cell 容量计算）等重复性极高的“冷启动”环节。根据数据分析结果与先前自动化工作流项目情况估算，预计系统上线后，可拦截并自动解决社区中 50% 以上的基础/中级开发问题，大幅缩短开发者成功运行第一个 CKB Demo 的时间（Time-to-First-Hello-World），直接提升黑客松和悬赏任务的转化率。

  • 提供独特竞争优势：针对 CKB 独有的 Cell 模型与硬核极客文化量身定制。有别于市面上的通用大模型，Nervos Brain 深度融合了 CKB 的底层逻辑与历史社区共识，不仅是一个问答工具，更是一个能跨越语言障碍、沉淀并分发 CKB 核心技术资产的生态中枢。

  • 未来规划备注：若本次 Spark MVP 项目验证成功并取得良好社区反馈，我们计划将其升级，并申请长期的 Community Fund DAO 资助，以覆盖整个 CKB 生态长期的运营与治理追踪。

社区内测与 MVP 验证方案 (User Study & Evaluation)

为了确保 Nervos Brain 的实际效用并避免“闭门造车”，在 Week 7-8 的内测阶段，我们将参考学术论文标准的用户研究（User Study）方法，采用“数据埋点 + 敏捷反馈”的闭环机制来评估和优化系统。

1. 核心成功指标 (Success Metrics) 我们将通过以下量化与质化指标来定义 MVP 的成功：

  • 准确率与采纳率 (Accuracy & Resolution)：

    • 目标：无人工干预情况下的问题解决率达到 50% 以上。

    • 测量方式：在 Bot 回复末尾引入快捷反馈机制（如 Discord/TG 的 :+1:/:-1: 或 :white_check_mark:/:cross_mark: 按钮）。标记为 :+1: 即视为有效解决。

  • 响应速度 (Latency)：

    • 目标：基础检索响应 < 15 秒；涉及多跳推理与代码生成的复杂 Agent 任务 < 60 秒。

  • 用户满意度 (User Satisfaction)：

    • 目标：内测用户整体满意度评分（CSAT）达到 3.0/5.0 以上。

2. 测试样本与招募渠道 (Testers & Channels)

  • 规模：招募 10 名以上具有相关开发经验的测试者。

  • 渠道：定向邀请 Discord dev-chat 频道的活跃用户、Nervos Talk 论坛高频发帖者，以及通过官方 Telegram 开发者群组发布内测邀请。

3. 反馈收集与整合机制 (Feedback Loop)

  • 隐式反馈 (Telemetry)：后台会记录所有被标记为 :-1: (Downvote) 的对话。我们会每周进行 Triage（问题分流），通过复盘 Agent 的内部推理轨迹（Reasoning Trace），定位是“检索遗漏”、“Prompt 幻觉”还是“工具调用错误”。

  • 显式反馈 (User Interview/Survey)：在 Week 8 结束前，向内测用户发送一份结构化的极简问卷（耗时 < 3分钟），收集他们对“回答专业度”、“语气自然度”及“代码可用性”的主观评价。

  • 迭代整合：基于上述收集到的 Bad Cases（失败案例），在最终交付前集中进行最后一轮的 RAG 知识库权重微调与提示词（Prompt）优化，并将详细的测试数据与转化率写入最终的结项报告中。

3 Likes
3

Hi @IrisNeko, thanks for submitting the Nervos Brain proposal. Before we move to committee review, please revise and clarify the following:
感谢您提交 Nervos Brain 项目提案。在提交委员会审核之前，请您修改并澄清以下内容：

1. Scope & Priority
MCP Server + Agentic RAG + cross-language TL;DR engine + Discord Bot + TG Bot + knowledge graph + vector DB, this is a lot for one person in 8 weeks. Please define your MVP core and rank your deliverables by priority.
一个人8周内同时交付这么多模块，工作量不小。请明确 MVP 核心以及对交付物做优先级排序。

2. $2,000 Justification
The top-tier grant requires stronger justification. Why is $1,000-$1,500 insufficient? And for a LLM project, your budget breakdown has no line item for LLM API costs. Which model are you using? Self-hosted or API calls? What’s the estimated cost over 8 weeks?
顶格资助需要更充分的论证。为什么 $1,000-$1,500 不够？另外，对于一个LLM项目，Budget 里缺少 LLM API 调用成本这一项。用什么模型？自部署还是调 API？8周预估费用多少？

3. Technical Details

  • The “60% auto-resolution” claim, what’s the basis? Data analysis from previous project or estimate?
  • Tech stack for knowledge graph, vector DB, embedding model?
  • “自动解决60%问题” 的依据是什么？基于过往项目的数据分析还是估算？
  • 知识图谱、向量数据库、embedding model 的具体技术选型是什么？

4. Testing Plan
Your budget mentions “community beta testing” but doesn’t elaborate. How do you plan to evaluate whether Nervos Brain actually works? Specifically: what metrics define success (e.g. answer accuracy, response time, user satisfaction)? How many testers, from which channels? How will you collect and incorporate feedback before final delivery?
Budget 中提到了"社区内测"但没有展开。你打算如何验证 Nervos Brain 是否真正有效？具体来说：用什么指标衡量成功（如回答准确率、响应速度、用户满意度）？多少测试用户、从哪些渠道招募？如何在最终交付前收集和整合反馈？

5. Sustainability
Who covers server and API costs after the grant ends? What’s the plan after the Spark?
Grant 结束后服务器和 API 费用谁承担？Spark 结项后该项目的计划是什么？

6. Background Details
Please add links to your GitHub profile and Civitai published works.
请补充 GitHub 和 Civitai 作品链接。

7. Payment Preference
This cycle, Spark grants can be paid in 100% CKB or 100% USDI. Please specify your preference.
本期 Spark 资助支持 100% CKB 或 100% USDI 支付，请明确您的选择。

Please update your proposal. Feel free to reach out if anything is unclear.
请更新您的 Proposal，有疑问请随时沟通。

cc @Hanssen @yixiu.ckbfans.bit @xingtianchunyan

6 Likes
4

Hi @zz_tovarishch, 感谢您的细致审核与宝贵建议！针对您提出的问题，我已经对主贴的提案内容进行了重新梳理和同步更新。以下是具体的澄清与答复：

1. Scope & Priority (MVP 核心与交付物优先级排版)

针对 8 周的开发周期，我已经明确了 MVP 的核心边界与优先级。现在的预期交付成果（MVP Core）如下：

  • 基础设施： 一套永久开源的 CKB AI 知识库与 Agent 系统，可供后续生态内所有钱包或开发者工具复用。
    • P0： Vector DB + Graph Planning（核心知识库与图结构知识检索框架，包括 Nervos 核心文档与示例）
    • P0： 基于 Agentic RAG 的 Bot 框架
    • P1： Nervos Talk MCP 工具，完成基础社交功能
    • P2： Discord Bot / TG Bot（根据具体进展确定是否在 8 周内完整交付）
  • 代码仓库： 交付完整的开源代码库。
  • 文档说明： 详细的部署指南与用户使用文档。

2. $2,000 Justification (资金预算与 LLM API 成本说明)

Nervos Brain 作为生态基础设施，工程复杂度远超“单模型+基础向量检索”的轻量级 Bot。我们需要构建混合检索底座、开发专属 MCP 服务器，并处理 CKB 硬核长文档与跨语言提炼，这会产生庞大的 Token 消耗。$2,000 是保障系统稳定落地与深度技术攻坚的必要门槛。

LLM API 策略（基于 X-MAS 理论）：
本项目不采用本地部署开源模型，而是全面调用云端 API。Agent 会动态路由任务以实现“效果-成本”的最优解：

  • 基础交互与意图识别： 路由至高性价比模型（如 GPT-5-mini / Haiku 4.5）。
  • 长文本与跨语言摘要： 路由至长窗口模型（如 Gemini 3.1 Pro / Claude 4.6 Sonnet），处理冗长文档与论坛长帖。
  • 深度推理与代码生成： 路由至顶尖模型（如 GPT-5.3 / Gemini 3.1 Pro），利用 CoT 处理底层架构或合约代码问题。

资金 Breakdown：

  • $700 (35%) - 检索基建研发： CKB 核心文档抓取与清洗、混合检索底座搭建、专属 MCP 服务器开发。
  • $600 (30%) - Agent 架构开发： 多模型路由系统开发、Prompt 调优（消除专有名词幻觉）、自动化跨语言巡逻脚本部署。
  • $400 (20%) - LLM API 成本（8周预估）： 涵盖全量数据初始化的 Embedding 消耗，以及开发调试与种子用户内测期产生的大量交互 Token 成本。
  • $300 (15%) - 服务器与交付： 内测期的云服务器/向量数据库租赁费用，以及开源文档编写与结项交付。

3. Technical Details (自动解决依据与技术选型)

关于“自动解决问题”的依据：
这基于对现阶段 CKB 核心开发者社群历史记录的抽样估算。数据显示，新开发者的大量痛点高度集中在环境配置、水龙头、RPC 报错及基础概念等重复性极高的“冷启动”环节。根据数据分析结果与先前自动化工作流项目情况估算，预计系统上线后可拦截并自动解决社区中 50% 以上的基础/中级开发问题，大幅缩短 Time-to-First-Hello-World，提升黑客松和悬赏转化率。

具体技术选型：

  • 向量数据库 (Vector DB)： 采用 Qdrant (Serverless)。免运维、轻量极快。通过为 CKB 文档打上精准 Payload 元数据标签与向量化索引，实现精准混合搜索。
  • Agent 推理与控制层： 采用基于 LangGraph 或大模型原生 Function Calling 的 ReAct 架构。Agent 会主动拆解问题，自主调用工具并进行图结构多跳推理。
  • Embedding 模型选型： 采用 text-embedding-3-large (OpenAI)。单价低，中英文双语跨语言表征与代码理解能力出色。

4. Testing Plan (社区内测与 MVP 验证方案)

为了避免“闭门造车”，我们将在 Week 7-8 采用“数据埋点 + 敏捷反馈”的闭环机制来评估和优化系统：

  • 1. 核心成功指标 (Metrics)： * 准确率与采纳率：无人工干预情况下的基础问题解决率 > 50%（通过回复末尾的 :+1:/:-1: 按钮测量）。
    • 响应速度：基础检索 < 15 秒；复杂 Agent 任务 < 60 秒。
    • 用户满意度：内测用户 CSAT > 3.0/5.0。
  • 2. 测试样本与招募： 招募 10 名以上有经验的测试者。通过 Discord dev-chat、Nervos Talk 及 TG 开发者群组定向邀请。
  • 3. 反馈收集与整合： * 隐式反馈：记录所有 :-1: 对话，每周 Triage 复盘 Agent 推理轨迹。
    • 显式反馈：Week 8 发放结构化极简问卷，收集专业度、可用性等主观评价。
    • 集中优化：针对 Bad Cases 集中微调 RAG 权重与 Prompt，最终数据将写入结项报告。

5. Sustainability (后期维护与计划)

  • 结束后的费用承担： Spark 结束后，服务器和 API 费用将由 Nervos 公司承担。
  • Spark 结项后的计划： 首要计划是补全进阶功能并实现生态的深度集成。首先，全面部署并长期维护 Discord 与 TG Bot，实现 24/7 自动化答疑；其次，上线跨语言巡逻脚本，主动推送 GitHub 与社区长帖的双语摘要；同时深化图结构多跳推理。在取得良好的社区真实数据验证后，我们将以此为基础申请长期的 Community Fund DAO 资助，持续推进 CKB 生态百科的迭代。

6. Background Details (作品链接)

7. Payment Preference (支付偏好)

本次 Spark 资助我选择 100% USDI 支付。

5 Likes
6

是因为没有到L1, 目前已经可以了，请把链接加入吧

1 Like
7

您好，打扰了。

上周五我针对 @zz_tovarishch 提出的宝贵建议，对 Nervos Brain 的提案进行了全面的梳理与更新（主要补充了 MVP 核心边界、API 预算拆解、技术选型以及测试方案等细节）。

考虑到刚刚度过周末，不知道各位目前是否有时间查看更新后的内容？非常希望能听听几位对最新方案的看法与意见。如果目前的 MVP 规划或资金预算还有任何需要进一步解释或调整的地方，我随时可以配合修改。

另外，也想向各位同步了解一下目前 Spark Program 提案的审批进度以及接下来的流程。

感谢各位的时间与对本提案的关注，期待你们的反馈！
@zz_tovarishch @Hanssen @yixiu.ckbfans.bit @xingtianchunyan

8

Hi Irisneko,

Spark committee每周一会进行会议评审讨论项目，
会后会尽快与申请人、项目方在Talk联系反馈，请稍等

2 Likes
9

@zz_tovarishch 傍晚好，打扰了。考虑到昨天咱们 Spark committee 已经完成了周一的评审，想赶在今天下班前冒昧跟进一下：不知道 Nervos Brain 的提案目前是否有初步的反馈或结论了？

主要是想看看是否有哪些地方需要我这边进一步修改或补充，如果有的话，我可以提前规划时间尽快调整，以免耽误咱们整体的审批流程。辛苦各位，期待您的回复！

1 Like
10

Hi @IrisNeko, apologies for the delayed response. Our staff member responsible for syncing outcomes was out sick today. Here’s the update.

The Spark Committee has reviewed the Nervos Brain proposal. The committee recognizes the value of using AI to integrate existing ecosystem resources and bridge community knowledge silos. After discussion, the proposal is currently Pending, with the following feedback for revision:

1. Deliverables need to be more concrete
The current description remains abstract on what the system actually does from an end-user perspective. Please clarify:

  • What specific interactions can users have with the Bot? (e.g. what types of questions it handles, what a typical conversation flow looks like)
  • How would an external developer experience and evaluate the deliverables?

2. Add a demo video/document to deliverables
To help testers and community members understand how to use and evaluate the system, please add a functional demo video or walkthrough document as a committed deliverable.

3. If requesting $2,000, P2 should be delivered
The committee’s position: if approved at the $2,000 cap, the Discord/TG Bot deployment (currently listed as P2, delivery TBD) should be included as a firm commitment, not optional.

Please revise your proposal accordingly.

Hi @IrisNeko，抱歉回复晚了，负责同步结果的同事今天生病了，这边更新一下进展。

Spark Committee 已经讨论了 Nervos Brain 提案。委员会认可该项目借助 AI 整合生态现有资源、打通社区信息壁垒的价值。经讨论，提案目前状态为 Pending，需要根据以下反馈进行修改：

1. 交付物需要更具体
当前描述在最终用户视角上仍然偏抽象，请补充说明：

  • 用户可以和 Bot 进行哪些具体交互？（比如能处理哪类问题、典型对话流程是什么样的）
  • 外部开发者如何体验和评估交付物？

2. 交付物中增加功能演示视频/文档
为了帮助测试者和社区成员了解如何使用和评估系统，请将功能演示视频或使用说明文档作为正式交付承诺。

3. 如果申请 $2,000，P2 应承诺交付
委员会的立场是：如果批准 $2,000 顶格资助，Discord/TG Bot 部署（目前列为 P2，视进度而定）应作为确定的交付承诺，是不是可选项。

请据此修改提案。

元宵节快乐！

Zhouzhou
On behalf of the Spark Program Committee

2 Likes
11

个人感觉俩痛点都没那么痛……而且通用人工智能已经能而且肯定能越来越满足基本需求。

2 Likes
12

通用 AI 的能力确实在快速提升。但目前通用模型在 CKB 特定问题上的表现还是有明显短板的。比如去问 ChatGPT 怎么构造一个 CKB 交易、Type Script 和 Lock Script 的区别、或者 Fiber 通道怎么开，得到的回答大概率是过时的、不准确的，甚至是编造的。核心原因是 CKB 的技术文档体量小、更新快，通用模型的训练数据覆盖不足。

把 CKB 的 RFC、开发者文档、Nervos Talk 历史讨论这些散落在各处的专有知识整合成一个可检索、可推理的知识库，让 AI 的回答有据可查。这跟通用模型的能力提升不矛盾，反而是互补的：通用模型越强，配上高质量的专有知识源，效果越好。

另外跨语言摘要这个点，目前中英文社区的讨论确实存在割裂，这不是通用 AI 能自动解决的问题，需要有人把管道搭起来。

当然，这个项目最终能不能真正解决问题，还要看交付质量。而且作为试验田，让开发者在Spark尝试想法，如果有推广到整个生态的潜力，再去申请DAO，也是Spark设置的初心。

General AI capabilities are indeed improving fast. But current general-purpose models still have clear gaps when it comes to CKB-specific questions. Ask ChatGPT how to construct a CKB transaction, the difference between Type Script and Lock Script, or how to open a Fiber channel, and the answers are likely outdated, inaccurate, or fabricated. The core reason is that CKB’s technical documentation is small in volume and updates frequently, so general models have insufficient training data coverage.

Integrating CKB’s RFCs, developer docs, and historical Nervos Talk discussions into a searchable, reasoning-capable knowledge base so that AI answers are grounded in verifiable sources is not at odds with general model improvements. It’s complementary: the stronger general models get, the better they perform when paired with high-quality domain-specific knowledge.

On the cross-language summary point, the English and Chinese community discussions are genuinely fragmented. That’s not something general AI solves on its own; someone needs to build the pipeline.

Of course, whether this project actually delivers on these goals comes down to execution quality. And as an experimental ground, letting developers test ideas through Spark and then pursue ecosystem-wide adoption via a DAO proposal if the potential is there, that’s exactly what Spark was designed for.

2 Likes