Financial News Digest Agent

Problem

Chinese financial briefs are scattered across Cailianshe, East Money, Sina Finance, and 36Kr. Reading them one by one is costly, and it is hard to quickly assess market sentiment, key figures, event types, and impact direction. Engineers also need a locally runnable, repeatable workflow that turns raw briefs into structured data that downstream systems can consume.

How It Works

The financial news digest agent runs through commands such as today, date YYYY-MM-DD, fetch, and summarize. It first fetches briefs from four Chinese financial sources. Each parser tries the API first, falls back to HTML parsing, and then uses a generic fallback if needed. It then uses an LLM to extract entity, key figures, event type, impact direction, a 50–100-character Chinese summary, keywords, and an importance score from 1 to 5. It also outputs positive, negative, or neutral sentiment labels with confidence and a numeric score.
The output includes a Markdown digest, a self-contained HTML page, and JSON data. The Markdown digest is grouped into headlines, market moves, policy updates, and company news. The HTML page uses embedded CSS to draw a sentiment pie chart and a keyword cloud. The JSON file can feed summarization, monitoring, or archival pipelines. A local JSON cache defaults to 30 minutes to reduce repeated requests.

Boundaries

This skill is designed for Chinese financial content and does not support English news sources, real-time streaming, or a web UI. Sentiment analysis relies on LLM semantic judgment and does not cover volatility or technical indicators from financial time-series models. Adding a new source requires updating sources.yaml and implementing the corresponding parser; adjusting summary style can be done by changing the temperature or system prompt.