NanoMesh

Distributed AI Orchestration System and Heterogeneous Micro-Model Infrastructure.
分布式 AI 调度系统与异构微模型基础设施网络。

English

1. Overview

NanoMesh is an adaptive, distributed orchestration system designed for heterogeneous micro-models and multi-modal task dispatching. The platform decomposes incoming requests into Directed Acyclic Graphs (DAGs) and stateful, multi-round conversation trees.

By establishing a tri-tier compute architecture, NanoMesh dynamically routes workloads across:

Client Tier: In-browser local model inference driven by WebGPU and WebAssembly.
Edge Tier: Distributed serverless compute nodes deployed on Cloudflare Workers.
Cloud Tier: Upstream large language model APIs for heavy generation tasks.

                                [ Client Request ]
                                        │
                         [ Unified Protocol Envelope ]
                                        │
           ┌────────────────────────────┼────────────────────────────┐
           ▼                            ▼                            ▼
   [ Client Tier ]               [ Edge Tier ]                [ Cloud Tier ]
 In-Browser WebGPU/Wasm      Cloudflare Workers Gateway     Upstream LLM APIs
  (SmolLM2-135M-ONNX)           (Routing & State Machine)   (Heavy Synthesis)

1.1 Visual Interface & Real-Time DAG Canvas

NanoMesh Interactive Workstation
Figure 1.1: NanoMesh Interactive Console & Workstation Interface

NanoMesh Real-Time DAG Canvas
Figure 1.2: Real-Time DAG Execution Canvas & Multi-Turn State Machine

2. Architecture & Core Concepts

2.1 Tri-Tier Compute Model

Lightweight Nodes (Client-Side): Executed locally in the user's browser using @huggingface/transformers with WebGPU or Wasm backends. Zero server compute consumption, low latency for deterministic subtasks.
Intermediate Nodes (Edge-Side): Lightweight routing, security classification, embedding generation, and state machine coordination running on Cloudflare Workers.
Heavyweight Nodes (Cloud-Side): Complex reasoning and multi-modal synthesis delegated to upstream model APIs.

2.2 Shift-Left Processing Pipeline

To minimize payload latency and gateway bandwidth:

Document Parsing: Handled in-browser via WebAssembly / pdf.js for PDF, DOCX, TXT, CSV, and Markdown. Extracted contents are populated as in_memory_text.
Image Optimization: Automated downsampling via HTML5 Canvas (maximum 2048px resolution, JPEG 0.85 compression), reducing raw asset size by 80%–90%.
Text Normalization: In-memory text extracted from documents is truncated at 30,000 characters if exceeded.
Direct Storage Integration: Assets requiring cloud persistence are uploaded directly to Cloudflare R2 storage using AWS SigV4 pre-signed PUT URLs.

3. Protocol Specification

3.1 Unified Protocol Envelope

Every client interaction is packaged into a strongly typed, immutable request envelope:

interface UnifiedProtocolEnvelope {
  envelope_id: string;             // Global UUID
  session_id: string;              // Active session identifier
  round_index: number;             // Monotonically increasing round counter
  instruction_payload: string;     // Cleaned natural language user instruction
  assets_manifest: AssetDescriptor[]; // Multi-modal asset metadata
  constraint_context: {
    target_language: string;       // Preferred output language ('zh' | 'en' | 'auto')
    output_style: string;          // Output format ('concise' | 'detailed' | 'structured')
  };
  metadata: {
    client_timestamp: number;      // Epoch timestamp in milliseconds
    char_count_total: number;      // Character count of the instruction
    estimated_input_tokens: number;// Estimated token consumption
    has_new_upload: boolean;       // Indicates if new assets were attached this round
    new_upload_modalities: string[];// Modality types of newly attached assets
  };
}

3.2 Asset Descriptor

interface AssetDescriptor {
  asset_id: string;                // Unique asset fingerprint
  modality: 'image' | 'audio' | 'text_doc' | 'tabular' | 'code' | 'pdf';
  mime_type: string;               // Original MIME type
  file_size: number;               // File size in bytes
  r2_key: string | null;           // Remote R2 object key
  in_memory_text: string | null;   // Deconstructed plain text or Markdown table
  in_memory_data: string | null;   // Optional base64 inline representation
  consumed: boolean;               // State cursor indicating upstream consumption
  source_digest: string;           // SHA-256 integrity digest
  is_new: boolean;                 // Indicates whether uploaded in current round
  filename: string;                // Original file name
}

Note: If a user submits a query without attachments, the gateway wraps the raw instruction into a virtual text asset to maintain DAG modality consistency.

4. Model Arbitration & Routing Pipeline

The routing pipeline implements a three-stage arbitration algorithm:

[ Input Request ]
        │
        ▼
┌────────────────────────────────────────────────────────┐
│ Stage 1: Security Audit & Hard Filtering               │
│ - Shadow String Canonicalization (Unicode NFKC)        │
│ - Heuristic Pattern Sniffing                           │
│ - JEV Safety Binary Classification (Clef-Flash)        │
│ - Subsistence Law Template Filtering                   │
└────────────────────────────────────────────────────────┘
        │
        ▼
┌────────────────────────────────────────────────────────┐
│ Stage 2: JEV Scoring & Role Competence Resolution      │
│ - Candidate Template Scoring via Workers AI            │
│ - Fast-Failed Sentinel Check                           │
│ - Competence Law: Modality & Tag Intersection          │
│ - Dynamic Tag Fallback & Compensation Prompt Injection │
└────────────────────────────────────────────────────────┘
        │
        ▼
┌────────────────────────────────────────────────────────┐
│ Stage 3: Semantic Anchor Matching & Traffic Shaping    │
│ - Dense Vector Embedding via @cf/baai/bge-m3           │
│ - Cosine Similarity against Model Anchors              │
│ - Ultimate Scoring Formula with Quadratic Penalty      │
│ - Instantiation of Primary (Top 1) & Standby Models    │
└────────────────────────────────────────────────────────┘

4.1 Stage 1: Security Audit and Hard Filtering

Shadow String Preprocessing: Strips zero-width characters, control codes, spaces, and punctuation; applies Unicode NFKC normalization and case folding.
Heuristic Sniffing: Inspects patterns for unauthorized exploits, credential theft, denial-of-service tools, and destructive instructions. Benign queries pass immediately (0ms). High-risk red lines (e.g., CSAM, physical violence) are blocked directly.
JEV Safety Bypass Classification: Suspicious patterns are evaluated using @cf/cloudflare/clef-flash binary classification (is_malicious: noul) to distinguish genuine threats from defensive or academic inquiries.
Hard Model Filtering: Discards models failing any of the following constraints:
- User rating $R_m < 1.0$
- Active concurrency $> 300$ or active connections $\ge$ maximum concurrency limit
- Context window $<$ estimated token requirement
- Model status $\neq$ 'active'
- Client preferences (disables inference_target === 'browser' if local inference is toggled off)
Subsistence Law: Filters candidate DAG templates based on user asset modalities: $${ \text{Head In-Modalities} } \subseteq A_{\text{user}} \subseteq { \text{Head In-Modalities} } \cup { \text{Tail In-Modalities} }$$ In single-node DAGs ($\text{in} = 0, \text{out} = 0$), this enforces exact equivalence: $A_{\text{user}} \equiv N_{\text{in}}$.

4.2 Stage 2: JEV Scoring and Role Competence

Template Relevance Scoring: Evaluates candidate templates matching the Subsistence Law using @cf/cloudflare/clef-flash.
Fast-Failed Sentinel: If the top template confidence score is below threshold ($\tau < 0.35$ or normalized score $< 25$), the gateway instantiates a role_unreachable sentinel node, returning an advisory capability boundary notification without debiting user credits.
Role Competence Law: For each role in the selected template, candidates are resolved through set intersection:
- $\text{Role.in_modalities} \subseteq \text{Model.in_modalities}$
- $\text{Role.out_modalities} \subseteq \text{Model.out_modalities}$
- $\text{Role.requisition_tag} \subseteq \text{Model.tags}$
Graceful Tag Downgrade: If the intersection yields an empty set, the scheduler steps down to the next priority tag and attaches a compensation prompt to maintain output consistency.

4.3 Stage 3: Semantic Anchor Matching & Traffic Shaping

Eligible models undergo quantitative ranking:

Vector Semantic Matching: Generates user prompt embeddings via @cf/baai/bge-m3 and calculates cosine similarity against model scenario anchors: $$\text{sim} = \max_{a \in \text{Anchors}} \cos(\vec{v}_{\text{prompt}}, \vec{v}_a)$$
Ultimate Scoring Formula: $$S = \mu_0 + \alpha \cdot \max(0, \text{sim} - \tau) + \beta \cdot \frac{R_m}{5.0} - \gamma \cdot \left(\frac{C_{\text{active}}}{C_{\text{max}}}\right)^2$$
- $\mu_0 = 0.25$ (Generalist prior score; 0.0 for specialists)
- $\alpha = 0.50$ (Specialist activation boost weight)
- $\tau = 0.65$ (Specialist activation threshold)
- $\beta = 0.25$ (Quality rating weight, where $R_m \in [1, 5]$)
- $\gamma = 0.40$ (Quadratic concurrency penalty weight)
Topology Instantiation: The highest-scoring candidate (Top 1) is selected as the Primary Model; Top 2 and Top 3 are preserved as Standby Models for manual client switching or automatic failover.

5. Execution State Machine & Streaming Protocol

5.1 Slicing and Consumption Rules

Head Node ($\text{in} = 0, \text{out} > 0$): Claims unconsumed matching assets, sets consumed = true. Masks user instruction payload (is_head_masked = true) to prevent perception bias.
Intermediate Node ($\text{in} > 0, \text{out} > 0$): Strictly consumes parent node outputs ($\sum \text{Out}_{\text{parent}}$). Does not access raw multi-modal assets.
Tail Node ($\text{out} = 0$): Consumes all upstream node outputs, remaining unconsumed matching assets, user instruction, and constraint context to synthesize the final delivery artifact.
Single-Node DAG ($\text{in} = 0, \text{out} = 0$): Directly receives both user prompt and assets.

5.2 Deterministic Node Fingerprinting

Nodes compute deterministic SHA-256 signatures for execution caching: $$\text{Signature} = \text{SHA256}(\text{Role} + \text{ModelID} + \text{InputSourceDigest} + \text{InjectedPrompt})$$ Identical executions under low temperature ($T \le 0.2$) are served from the D1 database cache table (dag_execution_cache).

5.3 Server-Sent Events (SSE) Protocol

Event Name	Description
`dag_plan`	Emits resolved DAG template metadata and node topology array.
`node_start`	Emits start of node execution with role, selected model, and target device.
`client_exec_request`	Dispatches execution payload to browser for local WebGPU inference.
`text_chunk`	Streams incremental text chunks (including `cache_hit` flag).
`node_complete`	Signals node completion with duration, token metrics, and output artifact.
`dag_complete`	Signals completion of the entire DAG pipeline with final payload.
`security_blocked`	Emitted when a security violation is detected and blocked.
`capability_unreachable`	Emitted when user request exceeds system capabilities (Fast-Failed).
`dag_error`	Emitted when an unrecoverable execution error occurs.

6. Multi-Turn Dialogue & Dynamic Task Trees

To prevent context bloat and hallucination, NanoMesh maintains a structured Session Asset Ledger rather than forwarding raw chat history:

[Session Asset Ledger]
1. Asset Inventory:
   - [Asset 1]: modality: image | filename: spec.png | round: 1
   - [Asset 2]: modality: text_doc | filename: api.md | round: 2
2. Branch Catalog:
   - [Branch main]: Topic: Architecture Design | Status: active
3. Current Focus:
   - Active Branch: [Branch main]
   - Latest Baseline Output: "System architecture defined..."
[User Instruction]: {currentPrompt}

The gateway routes subsequent interactions through a five-state convergence model:

                            User Submits Query
                                    │
               ┌────────────────────┴────────────────────┐
               ▼                                         ▼
     New Assets Uploaded                        No New Assets Uploaded
               │                                         │
       JEV Binary Choice                         JEV Four-Way Choice
      ┌────────┴────────┐               ┌────────┬───────┴────────┬────────┐
      ▼                 ▼               ▼        ▼                ▼        ▼
CROSS_FUSION      DETACH_RESET     ANCHOR_TAIL REBIND_ROOT   DETACH_RESET SWITCH_BRANCH
(Multi-Source)    (Independent)    (Refinement) (Re-inspect)  (New Task)   (Branch Switch)

State A (ANCHOR_TAIL): Instruction belongs to the active branch domain and its requirements are fully satisfied by the latest artifact summary. Logically strips raw media assets and appends a lightweight execution node to the active branch tail.
State B (REBIND_ROOT): Instruction belongs to the active branch domain, but requests information uncaptured or lost in the downstream summary, necessitating re-examination of the branch's underlying raw multi-modal physical media. Re-mounts raw media assets, resets consumed = false, and routes to multi-modal perception workflows.
State C (CROSS_FUSION): Joint analysis comparing newly uploaded media with historical context. Re-mounts historical + new assets (consumed = false).
State D (DETACH_RESET): Orthogonal new task with zero semantic dependency on existing branches or assets. Clears historical context and initializes an isolated net envelope.
State E (SWITCH_BRANCH): Instruction departs from the active branch domain and targets an inactive historical branch in the catalog. Structurally excludes self-loops (targets $B \in \text{Historical} \setminus { \text{Active} }$). Re-activates the target branch's focus and re-binds its root media assets if perceptual depth is required.

7. Client WebGPU Runtime

NanoMesh embeds an in-browser inference engine driven by @huggingface/transformers (v3.8.1):

Model Endpoint: onnx-community/SmolLM2-135M-Instruct-ONNX-GQA
Model Footprint: 135M parameters; Q4 quantized download size $\approx 82\text{ MB}$.
Hardware Acceleration: Probes WebGPU with shader-f16 / q4f16 precision; automatically falls back to WebAssembly SIMD CPU execution if WebGPU is unavailable.
Compatibility Polyfill: Integrates @petamoriken/float16 to ensure compatibility across older Chromium mobile kernels.
Inference Configuration:
- do_sample: true
- temperature: 0.6
- top_p: 0.9
- repetition_penalty: 1.25
- max_new_tokens: 384
Output Sanitization: Cleans chat template tokens (<|im_start|>, System:) in real time.

8. Security, Rate Limiting & Quotas

8.1 Sanitization and HTTP Security

DOMPurify Sanitization: All rendered Markdown outputs are sanitized through DOMPurify 3.1.6 inside safeRenderMarkdown(). Strips <script>, <iframe>, <object>, <form> elements and inline event attributes (onerror, onload, onclick).
HTTP Security Headers: Injects X-Content-Type-Options: nosniff across all API responses.

8.2 Sliding-Window Token Bucket Limiter

Implemented via an in-memory map within Cloudflare Worker isolates:

Bucket Capacity: 3 tokens (burst tolerance for immediate follow-ups).
Refill Interval: 1 token per 5 seconds (~12 requests/minute steady state).
Rate Limit Response: Returns HTTP 429 with Retry-After: N.
Client Cooldown: Displays an interactive cooldown countdown modal (#rate-limit-modal). Preserves user input in the chat box and enables resumption once the countdown finishes.

8.3 Upload Quotas

File Count Limit: Maximum 5 files per interaction (MAX_FILES = 5).
Total Payload Limit: Aggregate file size capped at 20MB (MAX_TOTAL_SIZE = 20MB).
Single File Hard Limit: Cloud direct upload limit capped at 25MB (MAX_CHAT_FILE_SIZE = 25MB).
Balance Protection: Users with credits <= 0 are blocked from file pre-signing and chat submission (HTTP 403 / modal alert).

9. Billing and Operations

Registration Bonus: 100 ⚡ initial compute credits upon account creation.
Two-Way Referral System: Referrers receive 50 ⚡ per referral. Referred users receive 100 ⚡ (initial bonus) + 50 ⚡ (referral reward) = 150 ⚡ total.
Ledger Audit: All transactions, credit consumption, and redemptions are recorded in D1 table credit_ledger.
Administrative Console (/admin.html):
- Real-time model configuration, concurrency controls, and status toggles.
- Voucher creation, activation, and batch invalidation.
- System notification broadcast dispatch.

10. Getting Started & Deployment

10.1 Prerequisites

Node.js v18+
Cloudflare Wrangler CLI
Cloudflare D1 Database & R2 Bucket

10.2 Configuration

Create .dev.vars in the project root directory:

JWT_SECRET=your-secure-jwt-secret
GITHUB_CLIENT_ID=your-github-oauth-client-id
GITHUB_CLIENT_SECRET=your-github-oauth-client-secret
GOOGLE_CLIENT_ID=your-google-oauth-client-id
GOOGLE_CLIENT_SECRET=your-google-oauth-client-secret
FRONTEND_URL=https://your-domain.com

Configure wrangler.toml bindings:

name = "nanomesh"
main = "src/index.ts"
compatibility_date = "2024-09-23"

[[d1_databases]]
binding = "DB"
database_name = "nanomesh-db"
database_id = "your-database-id"

[[r2_buckets]]
binding = "R2"
bucket_name = "nanomesh-assets"

[ai]
binding = "AI"

10.3 Installation and Deployment

# Install dependencies
npm install

# Deploy to Cloudflare Workers
npm run deploy

简体中文

1. 项目概述

NanoMesh 是一套针对异构微模型与多模态任务调度的自适应分布式编排系统。系统将用户的多模态指令编译为可编辑的有向无环图（DAG），并在多轮交互中自适应维护有记忆的任务演化树。

系统构建了三层算力拓扑网络：

端侧轻节点：运行于浏览器内的 WebGPU / WebAssembly 微模型推理；
边缘中节点：部署于 Cloudflare Workers 的无服务器分布式网关与状态机；
云端重节点：接入上游大语言模型 API 执行高复杂度生成任务。

                                  [ 用户交互指令 ]
                                         │
                                  [ 统一协议信封 ]
                                         │
           ┌─────────────────────────────┼─────────────────────────────┐
           ▼                             ▼                             ▼
     [ 端侧轻节点 ]                [ 边缘中节点 ]                [ 云端重节点 ]
 浏览器 WebGPU/Wasm 本地推理     Cloudflare Workers 网关调度     上游大模型 API 直答
   (SmolLM2-135M-ONNX)           (安全审计与状态机流转)         (重度逻辑综合交付)

1.1 系统界面与实时 DAG 编排画布

图 1.1: NanoMesh 交互控制台与多模态工作区

NanoMesh 实时 DAG 编排画布
图 1.2: 实时 DAG 编排画布与多轮意图溯源状态机

2. 架构设计与核心机制

2.1 三层算力拓扑

轻节点（浏览器侧端推理）：基于 @huggingface/transformers，直接调用客户端 GPU（WebGPU）或 CPU（WASM SIMD）运行，不消耗云端算力，保障基础处理时延与隐私隔离。
中节点（边缘分布式计算）：基于 Cloudflare Workers 执行安全审计、向量嵌入匹配、模板评分及 DAG 状态机调度。
重节点（云端模型直答）：处理需要大参数量支撑的复杂代码生成、多模态综合推理等重任务。

2.2 计算左移处理管线

为降低传输等待时间与网关计算压力，预处理全面左移至客户端执行：

文档就地解析：浏览器端使用 WebAssembly / pdf.js 直接解析 PDF、DOCX、TXT、CSV、Markdown 等文档，提取为 in_memory_text 纯文本或表格流。
图像自适应降采样：客户端通过 HTML5 Canvas API 按比例缩减分辨率（最大 2048px，JPEG 0.85 质量压缩），在人眼无损前提下缩减 80%~90% 上传体积。
文本软限制截断：文档提取出的文本超过 30,000 字符时，就地自动截断并给予提示。
直连对象存储：对于需要持久化的多模态介质，通过网关签发 AWS SigV4 预签名临时 PUT URL，文件由浏览器直传 Cloudflare R2 存储桶。

3. 通信协议规范

3.1 统一协议信封 (UnifiedProtocolEnvelope)

客户端在发起调度前，将非结构化输入封装为强类型只读信封：

interface UnifiedProtocolEnvelope {
  envelope_id: string;             // 全局唯一信封 UUID
  session_id: string;              // 当前会话 Session ID
  round_index: number;             // 当前会话轮次序号 (从 1 递增)
  instruction_payload: string;     // 清洗后的自然语言文本提示词
  assets_manifest: AssetDescriptor[]; // 关联多模态资产清单
  constraint_context: {            // 客户端请求的环境约束上下文
    target_language: string;       // 偏好输出语言 ('zh' | 'en' | 'auto')
    output_style: string;          // 输出风格 ('concise' | 'detailed' | 'structured')
  };
  metadata: {                      // 运行时元数据
    client_timestamp: number;      // 客户端发起的毫秒时间戳
    char_count_total: number;      // 提示词字符数
    estimated_input_tokens: number;// 预估消耗 Token 量
    has_new_upload: boolean;       // 本轮是否包含新引入资产
    new_upload_modalities: string[];// 本轮新引入资产的模态列表
  };
}

3.2 资产描述符 (AssetDescriptor)

interface AssetDescriptor {
  asset_id: string;                // 资产唯一指纹 ID
  modality: 'image' | 'audio' | 'text_doc' | 'tabular' | 'code' | 'pdf';
  mime_type: string;               // 原始 MIME 类型
  file_size: number;               // 字节大小
  r2_key: string | null;           // R2 寻址指针（若需云端存储）
  in_memory_text: string | null;   // 就地解构提取出的纯文本 / Markdown
  in_memory_data: string | null;   // Base64 短内联数据（可选）
  consumed: boolean;               // 状态游标：是否已被 DAG 某个上游节点独占消费
  source_digest: string;           // 资产完整性哈希指纹
  is_new: boolean;                 // 是否为本轮对话新上传的资产
  filename: string;                // 文件名
}

说明：若用户提交时未挂载任何附件，网关自动将提示词虚拟为文本资产挂载入资产清单，保证 DAG 模态匹配规则的一致性。

4. 模型仲裁与路由算法

系统采用三步仲裁法确定执行拓扑：

[ 用户输入请求 ]
       │
       ▼
┌────────────────────────────────────────────────────────┐
│ 第一步：安全审计与硬筛选                               │
│ - 影子字符串标准化 (Unicode NFKC 压平)                 │
│ - 启发式风险模式嗅探                                   │
│ - JEV Safety 旁路二元决策 (@cf/cloudflare/clef-flash) │
│ - 基于“温饱规则”的候选模板筛选                        │
└────────────────────────────────────────────────────────┘
       │
       ▼
┌────────────────────────────────────────────────────────┐
│ 第二步：JEV 评分与胜任力解析                           │
│ - Workers AI 候选模板胜任度打分                        │
│ - Fast-Failed 边界探针判定                             │
│ - 胜任定律：输入/输出模态与岗位标签交集运算            │
│ - 岗位标签降级策略与补偿 Prompt 注入                   │
└────────────────────────────────────────────────────────┘
       │
       ▼
┌────────────────────────────────────────────────────────┐
│ 第三步：锚点向量匹配与流量整形                         │
│ - @cf/baai/bge-m3 提取多语言密集语义向量               │
│ - 与模型高频场景锚点计算余弦相似度                     │
│ - 终极量化打分公式计算（含二次方并发惩罚）             │
│ - 实例化主执行模型 (Top 1) 与备选模型池 (Top 2 & 3)    │
└────────────────────────────────────────────────────────┘

4.1 第一步：安全审计与硬筛选

影子字符串预处理：剥离零宽字符与不可见格式控制符，执行 Unicode NFKC 标准化，压平空白与标点符号，统一转小写。
启发式嗅探打标：对恶意代码、未授权渗透、网络拒绝服务工具及凭据窃取等风险特征进行模式扫描。常规良性请求 0ms 直接放行；高危红线（如违禁暴力）直接熔断阻断。
JEV Safety 旁路拦截：仅对命中可疑特征的输入，调用 @cf/cloudflare/clef-flash 执行 is_malicious: noul 二元分类，区分实际恶意利用与良性安全防护配置。
模型硬筛选：从元数据库剔除不符合以下条件的模型：
- 综合评分 $R_m < 1.0$；
- 实时并发数 $> 300$ 或当前并发达到自身上限（满载）；
- 上下文窗口容量小于当前预估所需 Token；
- 模型状态非激活（status !== 'active'）；
- 满足客户端本地推理偏好（若用户开启“不使用本地模型推理”，屏蔽 inference_target === 'browser'）。
温饱规则（模板收敛）： $${ \text{头节点输入模态} } \subseteq A_{\text{用户资产}} \subseteq { \text{头节点输入模态} } \cup { \text{尾节点输入模态} }$$ 对于单节点 DAG（$\text{in} = 0, \text{out} = 0$），强制等价收敛：$A_{\text{用户资产}} \equiv N_{\text{输入模态}}$。

4.2 第二步：JEV 评分与岗位胜任力

候选模板评估：调用 @cf/cloudflare/clef-flash 对满足温饱定律的候选模板进行打分评估。
Fast-Failed 机制：若最高模板匹配得分低于置信阈值（$\tau < 0.35$ 或标准分 $< 25$），直接装配 role_unreachable 单节点，以 0 算力消耗返回能力边界建议面板。
胜任定律交集运算：从模板提取各个岗位，对候选模型库执行交集筛选：
- 岗位输入模态 $\subseteq$ 模型输入模态；
- 岗位输出模态 $\subseteq$ 模型输出模态；
- 岗位征调标签 $\subseteq$ 模型标签。
标签平滑降级：若交集为空，按“岗位-标签”预置表向下检索备选标签重新匹配，并在该节点挂载补偿 Prompt 稳定输出。

4.3 第三步：锚点向量匹配与流量整形

在精选模型池内执行最终量化选拔：

语义向量匹配：调用 @cf/baai/bge-m3 提取提示词向量，与模型预设的高频场景锚点做批量余弦相似度计算： $$\text{sim} = \max_{a \in \text{Anchors}} \cos(\vec{v}_{\text{prompt}}, \vec{v}_a)$$
终极量化打分公式： $$S = \mu_0 + \alpha \cdot \max(0, \text{sim} - \tau) + \beta \cdot \frac{R_m}{5.0} - \gamma \cdot \left(\frac{C_{\text{active}}}{C_{\text{max}}}\right)^2$$
- $\mu_0 = 0.25$（通才先验分；专才模型为 0.0）
- $\alpha = 0.50$（专才权重系数）
- $\tau = 0.65$（专才门控激活阈值）
- $\beta = 0.25$（综合评分权重，其中 $R_m \in [1, 5]$）
- $\gamma = 0.40$（二次方并发惩罚系数，负载接近满载时抑制分配）
拓扑实例化：评分第 1 名作为主执行模型（Primary Model）；第 2、3 名写入备选模型池（Standby Models），供前端随时热切换或网关故障转移。

5. 执行状态机与流式协议

5.1 信封切片规则（不贪吃原则与提示词遮蔽）

首节点（$\text{in} = 0, \text{out} > 0$）：认领模态匹配且 consumed === false 的资产，置位 consumed = true。严格遮蔽用户主提示词（is_head_masked = true），避免下游语义污染。
中间节点（$\text{in} > 0, \text{out} > 0$）：严格仅接收所有父节点的汇总输出产物（$\sum \text{Out}_{\text{parent}}$），不触碰主信封原始介质。
尾节点（$\text{out} = 0$）：接收上游汇总产物、剩余未被消费资产、用户原始提示词与环境约束，生成最终交付物。
单节点 DAG（$\text{in} = 0, \text{out} = 0$）：同时接收原始资产与提示词。

5.2 确定性节点指纹与缓存

各节点依据输入要素计算 SHA-256 哈希签名： $$\text{Signature} = \text{SHA256}(\text{Role} + \text{ModelID} + \text{InputSourceDigest} + \text{InjectedPrompt})$$ 在低采样温度（$T \le 0.2$）下，通过 D1 表 dag_execution_cache 命中产物直接返回。

5.3 SSE 实时流式事件协议

事件名称	描述
`dag_plan`	广播征调模板元信息与 DAG 拓扑节点结构数组。
`node_start`	广播开始执行的节点序号、岗位名称、选中模型及设备目标。
`client_exec_request`	当节点调度至浏览器本地时，向前端派发本地推理指令 payload。
`text_chunk`	尾节点流式输出文本增量 Token（附带 `cache_hit` 标识）。
`node_complete`	单个节点执行完成，回传耗时、Token 统计与产物。
`dag_complete`	整个 DAG 编排图执行完毕，推送最终综合交付物。
`security_blocked`	触发安全策略拦截时推送。
`capability_unreachable`	触发能力边界不可达（Fast-Failed）时推送。
`dag_error`	发生不可恢复的执行异常时推送。

6. 多轮交互与动态任务树

网关不盲目透传整段历史，而是维护一份结构化的“会话事实账本”：

【会话事实账本】
1. 资产原件介质表:
   - [资产 1]: 模态: image | 文件名: spec.png | 引入轮次: 第 1 轮
   - [资产 2]: 模态: text_doc | 文件名: api.md | 引入轮次: 第 2 轮
2. 任务树分支目录:
   - 【分支 main: 初始系统分析】 状态: 活跃分支
3. 当前活跃焦点:
   - 当前正在进行的分支: 【分支 main】
   - 当前分支最新产物基底: "系统架构定义已完成..."
[用户最新指令]: {currentPrompt}

JEV 依据上下文进行五状态收敛判定：

                            用户在聊天框发送指令
                                     │
               ┌─────────────────────┴─────────────────────┐
               ▼                                           ▼
      本轮包含新上传附件                           本轮无新上传附件
               │                                           │
         JEV 二选一判定                              JEV 四选一判定
       ┌───────┴───────┐                 ┌────────┬────────┴────────┬────────┐
       ▼               ▼                 ▼        ▼                 ▼        ▼
 CROSS_FUSION    DETACH_RESET       ANCHOR_TAIL REBIND_ROOT   DETACH_RESET SWITCH_BRANCH
 (新旧融合对比)  (独立全新任务)       (轻量修饰)  (回溯重探)    (新独立任务) (历史分支切换)

状态 A (ANCHOR_TAIL)：指令属于当前活跃分支脉络，且所需事实信息完全由当前最新产物基底即可自洽满足。逻辑剥离原始大型媒体资产，以文本基底在活跃分支末端生长单步执行节点。
状态 B (REBIND_ROOT)：指令属于当前活跃分支脉络，但所需信息未在最新产物基底中包含或在下游抽象中已丢失，仅凭文字基底无法充分解答，必须穿透回溯底层原始多模态物理介质本身。全量复位该分支原始物理资产（consumed = false）并装配多模态感知工作流。
状态 C (CROSS_FUSION)：联合分析新上传资产与历史既有资产。全量重置资产消费游标（consumed = false）并在当前分支创建融合节点。
状态 D (DETACH_RESET)：指令开启完全独立的新任务，与当前活跃分支及所有历史分支无信息依赖。彻底清空历史挂载，构造独立的“零历史净信封”。
状态 E (SWITCH_BRANCH)：指令脱离当前活跃分支领域，转而依赖某一非活跃的历史分支。目标分支集合在图结构上严格排除当前分支自环（目标分支 $B \neq B_{\text{active}}$）；切换焦点指针至目标分支，若目标分支关联多模态物理介质且需要溯源，连带复位该分支原始资产。

7. 端侧 WebGPU 微模型引擎

NanoMesh 内置基于 @huggingface/transformers (v3.8.1) 的浏览器端微模型推理引擎：

选型端点：onnx-community/SmolLM2-135M-Instruct-ONNX-GQA
模型规格：约 1.35 亿参数；Q4 量化下载体积约 82MB。
硬件适配：优先探测 WebGPU 与 shader-f16 / q4f16 精度；若不支持则自动降级为 Wasm SIMD CPU 执行。
环境垫片：针对部分移动端 Chromium 内核缺少原生 Float16Array 的问题，引入 @petamoriken/float16 垫片。
运行采样参数：
- do_sample: true
- temperature: 0.6
- top_p: 0.9
- repetition_penalty: 1.25
- max_new_tokens: 384
输出过滤：实时过滤 <|im_start|>、System: 等模板泄露标记。

8. 安全机制、频控与配额

8.1 渲染安全与响应头

DOMPurify 净化管道：所有 Markdown 及模型流式输出均在 safeRenderMarkdown() 内通过 DOMPurify 3.1.6 净化。移除 <script>、<iframe>、<object>、<form> 及 onerror、onload、onclick 等内联属性。
安全响应头：所有 API 响应注入 X-Content-Type-Options: nosniff。

8.2 边缘滑动窗口令牌桶限流

基于 Cloudflare Worker Isolate 内存 Map 实现的自适应令牌桶：

桶容量：3（支持短时快速重试）。
补充速率：每 5 秒补充 1 个令牌（稳态等效持续约 12 请求/分钟）。
超频响应：返回 HTTP 429 与 Retry-After: N 响应。
交互保护：前端弹出自定义倒计时模态弹窗（#rate-limit-modal），自动保留并回填用户输入文本，倒计时结束后解除锁定。

8.3 附件上传配额

文件个数限制：单次最多上传 5 个文件（MAX_FILES = 5）。
总文件大小限制：多附件总大小上限 20MB（MAX_TOTAL_SIZE = 20MB）。
单文件上传硬限制：单文件云端上传硬限制 25MB（MAX_CHAT_FILE_SIZE = 25MB）。
算力余额保护：用户可用余额 credits <= 0 时，禁止附件预签名与任务提交（HTTP 403 / 弹出充值提示）。

9. 财务与运营

注册体验包：新用户注册获得 100 ⚡ 初始算力包。
双向邀请机制：老用户成功邀请新用户，老用户获赠 50 ⚡，新用户获得初始 100 ⚡ + 邀请奖励 50 ⚡（共 150 ⚡）。
账目审计体系：D1 表 credit_ledger 全量记录任务扣费、注册赠予、卡密充值与网格收益流水。
管理控制台 (/admin.html)：
- 在线修改模型状态、并发上限与权重；
- 批量生成、查询与作废充值卡密；
- 向客户端派发全员或定向系统通知。

10. 快速部署

10.1 环境要求

Node.js v18+
Cloudflare Wrangler CLI
Cloudflare D1 数据库与 R2 存储桶

10.2 环境变量配置

在项目根目录下配置 .dev.vars：

JWT_SECRET=your-secure-jwt-secret
GITHUB_CLIENT_ID=your-github-oauth-client-id
GITHUB_CLIENT_SECRET=your-github-oauth-client-secret
GOOGLE_CLIENT_ID=your-google-oauth-client-id
GOOGLE_CLIENT_SECRET=your-google-oauth-client-secret
FRONTEND_URL=https://your-domain.com

配置 wrangler.toml 绑定信息：

name = "nanomesh"
main = "src/index.ts"
compatibility_date = "2024-09-23"

[[d1_databases]]
binding = "DB"
database_name = "nanomesh-db"
database_id = "your-database-id"

[[r2_buckets]]
binding = "R2"
bucket_name = "nanomesh-assets"

[ai]
binding = "AI"

10.3 构建与部署

# 安装依赖
npm install

# 部署至 Cloudflare Workers
npm run deploy

License

Distributed under the MIT License. See LICENSE for more information.

NanoMesh

项目介绍

NanoMesh

English

1. Overview

1.1 Visual Interface & Real-Time DAG Canvas

2. Architecture & Core Concepts

2.1 Tri-Tier Compute Model

2.2 Shift-Left Processing Pipeline

3. Protocol Specification

3.1 Unified Protocol Envelope

3.2 Asset Descriptor

4. Model Arbitration & Routing Pipeline

4.1 Stage 1: Security Audit and Hard Filtering

4.2 Stage 2: JEV Scoring and Role Competence

4.3 Stage 3: Semantic Anchor Matching & Traffic Shaping

5. Execution State Machine & Streaming Protocol

5.1 Slicing and Consumption Rules

5.2 Deterministic Node Fingerprinting

5.3 Server-Sent Events (SSE) Protocol

6. Multi-Turn Dialogue & Dynamic Task Trees

7. Client WebGPU Runtime

8. Security, Rate Limiting & Quotas

8.1 Sanitization and HTTP Security

8.2 Sliding-Window Token Bucket Limiter

8.3 Upload Quotas

9. Billing and Operations

10. Getting Started & Deployment

10.1 Prerequisites

10.2 Configuration

10.3 Installation and Deployment

简体中文

1. 项目概述

1.1 系统界面与实时 DAG 编排画布

2. 架构设计与核心机制

2.1 三层算力拓扑

2.2 计算左移处理管线

3. 通信协议规范

3.1 统一协议信封 (UnifiedProtocolEnvelope)

3.2 资产描述符 (AssetDescriptor)

4. 模型仲裁与路由算法

4.1 第一步：安全审计与硬筛选

4.2 第二步：JEV 评分与岗位胜任力

4.3 第三步：锚点向量匹配与流量整形

5. 执行状态机与流式协议

5.1 信封切片规则（不贪吃原则与提示词遮蔽）

5.2 确定性节点指纹与缓存

5.3 SSE 实时流式事件协议

6. 多轮交互与动态任务树

7. 端侧 WebGPU 微模型引擎

8. 安全机制、频控与配额

8.1 渲染安全与响应头

8.2 边缘滑动窗口令牌桶限流

8.3 附件上传配额

9. 财务与运营

10. 快速部署

10.1 环境要求

10.2 环境变量配置

10.3 构建与部署

License

讨论与反馈

提 示

站务通知

需要登录

提交投诉

提示