← all repositories
DemonDamon/FinnewsHunter

When your trading bot needs a committee meeting first

A multi-agent system that crawls Chinese financial news, debates sentiment, and spits out alpha signals—assuming you can keep all the services running.

1.4k stars Python AgentsDomain Apps
FinnewsHunter
Velocity · 7d
+0.5
★ / day
Trend
steady
star history

What it does FinnewsHunter is a full-stack platform that scrapes real-time financial news from Chinese sources (Sina Finance, Securities Times, etc.), vectorizes articles, and routes them through LLM-powered agents for sentiment analysis and market-impact scoring. It bundles a React frontend, FastAPI backend, PostgreSQL/Milvus/Redis storage, and Celery workers into Docker Compose. There’s also K-line charting via akshare and fuzzy search across 5,000+ A-share stocks.

The interesting bit The “multi-agent” angle isn’t just branding. The system explicitly orchestrates agent roles—NewsAnalyst, Researcher, etc.—through the author’s own AgenticX framework, with a “debate workflow” layer. It’s essentially a quant shop’s research pipeline rebuilt as opinionated infrastructure, complete with one-click LLM provider switching between Bailian, DeepSeek, Kimi, OpenAI, and Zhipu.

Key highlights

  • Built on the author’s AgenticX framework; reuses its BailianEmbeddingProvider and MilvusStorage rather than rolling custom vector plumbing
  • Supports 5 LLM providers with frontend toggle; defaults to Alibaba Cloud’s Bailian for China latency
  • Async vectorization pipeline: articles get embedded in background without blocking analysis
  • Docker Compose deployment with explicit service dependency chain (Postgres → Redis → Milvus → Celery)
  • Includes 4 different database reset methods, which tells you something about iteration velocity

Caveats

  • README install path hardcodes /Users/damon/myWork/AgenticX for the framework dependency; you’ll need to adapt this
  • The “knowledge graphs” and “alpha factor mining” mentioned in the intro aren’t detailed in the visible README sections—actual signal generation logic is unclear
  • Frontend cache invalidation requires manual hard-refresh; the docs spend surprising length explaining browser DevTools

Verdict Worth a look if you’re building quant-adjacent NLP pipelines for Chinese equities and want a reference architecture rather than a library. Skip it if you need a clean, minimal API or aren’t prepared to babysit a 6-service Docker fleet.

heatdrop uses Google Analytics to see which pages get read — nothing else. Your call. How we handle data.