AI Knowledge Base Agent

中文文档

Automatically collect AI content from multiple sources (GitHub Trending, arXiv/blog RSS, Hacker News), intelligently analyze and score it, and generate a structured knowledge base with a static web UI.

Features

• Pluggable Multi-Source Collection — GitHub Trending + RSS (arXiv / vendor blogs) + Hacker News, unified into a generic Item model. Adding a source = adding one file; enabled via AKB_SOURCES
• Source-Aware 3D Scoring — Technical Advancement / Practicality / Community Activity, 1-10 scale; the three axes are reinterpreted per source via per-source prompts (e.g. for articles: insight depth / applicability / timeliness)
• Dynamic Weight Adjustment — Auto-adjusts weights based on human review feedback
• Bonus Scoring — Breakthrough innovation projects can receive extra points (up to +2)
• Date-based Storage — Entries organized under knowledge/YYYY-MM-DD/ subdirectories
• Auto Cleanup — Automatically removes directories older than 15 days on each pipeline run
• Static Web UI — Dark-themed SPA with search, tag filtering, and date navigation, deployed on Cloudflare Pages
• GitHub Actions — Daily automated collection with auto-commit of results

Project Structure

ai-knowledge-base/
├── .github/workflows/     # GitHub Actions scheduled tasks
├── app/
│   ├── sources/           # Pluggable sources (github / rss / hackernews) + registry
│   ├── crawler/           # GitHub Trending crawler (used by sources/github_trending)
│   ├── agent/             # AI scoring agent (LLM-powered) + per-source prompts
│   ├── storage/           # Knowledge entry storage
│   ├── review/            # Human review management
│   ├── api/               # FastAPI web interface
│   └── main.py            # Main entry point
├── scripts/
│   ├── build_static.py    # Generate data.json from markdown
│   └── build.sh           # Cloudflare Pages build script
├── site/
│   ├── index.html         # Static SPA frontend
│   └── data/              # Generated data.json
├── templates/             # HTML templates (FastAPI)
├── knowledge/             # Generated knowledge entries
│   ├── index.md           # Master index
│   └── 2026-05-17/        # Date-based subdirectory
│       ├── project-a_9.0_2026-05-17.md
│       └── project-b_8.5_2026-05-17.md
└── pyproject.toml

Quick Start

# Install
python3 -m venv .venv && source .venv/bin/activate
pip install -e .

# Configure
cp .env.example .env
# Edit .env and fill in LLM_API_KEY

# Start web server (with scheduler)
akb serve

# Run a manual crawl + analysis
akb crawl

# Adjust weights based on review feedback
akb adjust-weights

# Build static site data
python3 scripts/build_static.py ./knowledge ./site/data

Environment Variables

Variable	Description	Default
LLM_API_URL	LLM API endpoint	OpenRouter
LLM_API_KEY	API Key	Required
LLM_MODEL	Model name	minimax/minimax-m3:free
LLM_CONCURRENCY	Concurrent LLM analysis requests (higher = faster, too high may hit rate limits)	5
GITHUB_TOKEN	GitHub Token (optional, increases rate limit)	-
GITHUB_MAX_PROJECTS	Max GitHub Trending projects per run	10
AKB_SOURCES	Enabled sources, comma-separated (github,rss,hackernews)	github,rss,hackernews
RSS_FEEDS	RSS/Atom feeds, comma-separated (empty = arXiv cs.AI/cs.CL + HF blog)	(defaults)
RSS_MAX_PER_FEED	Max entries per RSS feed	10
HN_QUERY	Hacker News search query	AI OR LLM OR agent
HN_MIN_POINTS	Min HN points to include	50
HN_MAX_ITEMS	Max HN items per run	20
CRAWL_SCHEDULE	Cron schedule for crawling	0 0 * * *
API_HOST	FastAPI listen address	127.0.0.1
API_PORT	FastAPI port	8900
KNOWLEDGE_DIR	Knowledge base directory	./knowledge

Scoring System

Three dimensions, each scored 1-10, with initial equal weights of 33.3%:

Dimension	Evaluation Criteria
Technical Advancement	Tech stack sophistication, innovation level, technical depth, cutting-edge relevance
Practicality	Problem-solving capability, use cases, scalability, commercial value
Community Activity	Star growth, Issue response, PR processing, documentation quality

Entries with a total score below 6 are automatically flagged as "pending review".

Data Retention

• Knowledge entries are stored in date-based subdirectories (knowledge/YYYY-MM-DD/)
• Each pipeline run automatically cleans up directories older than 15 days
• The master index (knowledge/index.md) is regenerated on every run

GitHub Actions

The included workflow (.github/workflows/daily-collect.yml) runs daily at UTC 00:00 (Beijing 08:00):

1. Crawls GitHub Trending AI projects
2. Analyzes and scores each project via LLM
3. Saves results to date-based subdirectories
4. Builds static site data (site/data/data.json)
5. Auto-commits and pushes new entries

Required repository secrets: LLM_API_URL, LLM_API_KEY, LLM_MODEL

Static Site Deployment

The site/ directory contains a self-contained SPA that reads data.json generated by build_static.py. Deploy to Cloudflare Pages with:

• Build command: bash scripts/build.sh
• Output directory: site
• Root directory: /

Live site: https://ai-knowledge-base-22f.pages.dev

License

MIT