Content Digital Employee Platform v5.0 Design Proposal

Product Positioning

A general-purpose digital content management platform for individuals and enterprises — an AI-driven content production system integrating 3 years of self-media experience.

Core Architecture (Reference: AIHOT)

``

┌─────────────────────────────────────────────────────────────┐

│ 内容数字员工平台 │

├─────────────────────────────────────────────────────────────┤

│ │

│ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │

│ │ 信源管理 │───▶│ AI 处理中心 │───▶│ 内容输出 │ │

│ └─────────────┘ └─────────────┘ └─────────────┘ │

│ │ │ │ │

│ ▼ ▼ ▼ │

│ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │

│ │ T1 Official Sources │ │ Pre-screening AI Relevance │ │ Curated Recommendations │ │

│ │ T1.5 Official Social Media │ │ Multi-dimensional Scoring │ │ Daily Report Generation │ │

│ │ T2 KOL/Media │ │ Event Clustering │ │ Multi-platform Publishing │ │

│ │ T3 Industry Vertical │ │ Translation + Summarization │ │ CMS Integration │ │

│ └─────────────┘ └─────────────┘ └─────────────┘ │

│ │

└─────────────────────────────────────────────────────────────┘

`

1. Information Source Management System

1.1 Information Source Tiering (inspired by AIHOT)

1.2 Connected Sources

• ✅ Weibo Hot Search (T2)
• ✅ Zhihu Hot List (T2)
• ✅ Reddit Hot Topics (T2-T3)
• ✅ Unitree Embodied Intelligence Community (T3)
• ? OpenAI Blog (T1)
• ? Anthropic Blog (T1)
• ? Hugging Face (T1.5)
• ? 36Kr / Machine Intelligence (T2)
• ? Product Hunt (T2)

1.3 Source Configuration File

`json

{

"sources": [

{

"id": "openai-blog",

"name": "OpenAI Blog",

"tier": 1,

"type": "rss",

"url": "https://openai.com/blog/rss.xml",

"category": ["AI", "LLM"],

"enabled": true

},

{

"id": "weibo-hot",

"name": "Weibo Hot Search",

"tier": 2,

"type": "api",

"url": "https://weibo.com/ajax/side/hotSearch",

"category": ["General", "Society"],

"enabled": true

}

]

}

`

2. AI Processing Center

2.1 Processing Pipeline (Based on AIHOT 11th Edition Iteration Experience)

`

Raw Information ──▶ Pre-filter (AI Relevance) ──▶ Multi-dimensional Scoring ──▶ Weight Calculation ──▶ Selection Judgment

│ │ │ │ │

│ ▼ ▼ ▼ ▼

│ DeepSeek V3 DeepSeek V4 Code Formula Threshold Filter

│ (Low-cost Pre-filter) (Five-dimension Scoring) (Controllable & Tunable) (Category Threshold)

│

└──▶ Translation + Summary ──▶ Event Clustering ──▶ Storage

`

2.2 Multi-dimensional Scoring System

Five-dimension Scoring (Handled by LLM, 200-line Prompt):

| Dimension | Description | Weight |

|------|------|------|

| Novelty | Whether it's the latest release / first report | 20% |

| Importance | Scope of impact, industry influence | 25% |

| Relevance | Relevance to target domain | 20% |

| Readability | Content quality, information density | 15% |

| Virality | Topic appeal, discussion potential | 20% |

Final Weight Calculation (Handled by Code, Quantifiable & Tunable):

`javascript

finalScore = (

baseScore tierWeight // Source weight

noveltyScore * 0.20 + // Novelty

importanceScore * 0.25 + // Importance

relevanceScore * 0.20 + // Relevance

readabilityScore * 0.15 + // Readability

viralScore * 0.20 // Virality

) * categoryMultiplier // Category multiplier

`

2.3 Event Clustering System

Problem: The same event is reported by multiple sources → duplicates on the selection page

Solution:

1. Embedding semantic similarity calculation
2. Time window aggregation (within 24 hours)
3. 权威源优先原则：官网 > 官推 > KOL > 媒体

`javascript

// Event cluster structure

{

"clusterId": "evt_20260507_gpt55",

"mainItem": "OpenAI Blog Release", // most authoritative source

"relatedItems": [

"36Kr Report",

"Machine Heart Interpretation",

"Altman's tweet"

],

"category": "Model Release",

"score": 89

}

`

2.4 Pre-screening Strategy (Cost Optimization)

`javascript

// Phase 1: Rule pre-screening (free)

if (isSpam(item) || isDuplicate(item) || isIrrelevantCategory(item)) {

return; // discard directly

}

// Phase 2: AI pre-screening (low-cost model)

const isAIRelated = await preScreenWithAI(item); // DeepSeek V3

if (!isAIRelated) {

store(item); // persist to DB without scoring

return;

}

// Phase 3: Deep scoring (high-intelligence model)

const scores = await scoreWithAI(item); // DeepSeek V4 Pro

`

3. Content Output System

3.1 Featured Recommendations

• Filter by category thresholds (thresholds vary by category)
• Real-time timeline updates
• Supports user-defined interest tags

3.2 AI Daily Report

• Auto-generated at 8:00 daily
• Five sections: model releases, product updates, industry news, research papers, insights and tips
• Processed data is directly aggregated, without requiring a second AI call

3.3 Multi-Platform Publishing

四、数据存储

4.1 核心表结构

`sql

-- 信源表

CREATE TABLE sources (

id VARCHAR(32) PRIMARY KEY,

name VARCHAR(100),

tier TINYINT,

type ENUM('rss', 'api', 'html', 'twitter'),

url VARCHAR(500),

category JSON,

enabled BOOLEAN DEFAULT TRUE,

last_fetch DATETIME

);

-- 内容表

CREATE TABLE contents (

id VARCHAR(32) PRIMARY KEY,

source_id VARCHAR(32),

title VARCHAR(500),

summary TEXT,

original_url VARCHAR(500),

publish_time DATETIME,

fetch_time DATETIME,

-- AI 处理结果

is_ai_related BOOLEAN,

scores JSON, -- 五维评分

final_score DECIMAL(5,2),

is_featured BOOLEAN, -- 是否精选

cluster_id VARCHAR(32),

-- Category tag

category VARCHAR(50),

tags JSON,

INDEX idx_score (final_score DESC),

INDEX idx_featured (is_featured, publish_time DESC)

);

-- Event cluster table

CREATE TABLE event_clusters (

cluster_id VARCHAR(32) PRIMARY KEY,

main_content_id VARCHAR(32),

category VARCHAR(50),

score DECIMAL(5,2),

created_at DATETIME

);

`

5. API Endpoints

5.1 Core Endpoints

`

GET /api/featured # Featured content list

GET /api/daily # Daily report

GET /api/trending # Trending topics

POST /api/sources # Add source

GET /api/sources # Source list

POST /api/score # Manual scoring

GET /api/search?q=xxx # Content search

`

5.2 Webhook

`

POST /webhook/new-featured # New featured notification

POST /webhook/daily-ready # Daily report ready

``

6. Tech Stack

7. Development Plan

Phase 1: Core Framework (1 week)

• [ ] Source configuration system
• [ ] Collection scheduler
• [ ] AI scoring module
• [ ] Database table schema

Phase 2: AI Processing (1 week)

• [ ] Pre-filter prompt design
• [ ] Five-dimension scoring prompt
• [ ] Event clustering algorithm
• [ ] Daily report generator

Phase 3: Output Layer (1 week)

• [ ] Featured API
• [ ] Daily report API
• [ ] Web admin interface
• [ ] Webhook notifications

Phase 4: Integration and Optimization (1 week)

• [ ] CMS integration
• [ ] PLM data integration
• [ ] Performance optimization
• [ ] Monitoring and alerting

8. Cost Estimation

AI call cost (500 items/day):

• Pre-filter: 500 × 0.5K tokens × ¥0.001/K = ¥0.25
• Scoring: 250 × 2K tokens × ¥0.01/K = ¥5.00
• Translation summary: 250 × 1K tokens × ¥0.01/K = ¥2.50
• Daily cost: ¥8 / day ≈ ¥240 / month

Optimization plan:

1. Local model pre-filter (free)
2. Batch processing to reduce call count
3. Cache similar content scores

Core philosophy:

Use code for control, let AI do what it does best (understanding semantics), and calculate everything else with explicit formulas.

Controllable, tunable, quantifiable, and backtestable.