Most sentiment tools stop at a single LLM scoring a wall of scraped text. BettaFish treats public-opinion analysis as a debate problem instead: five specialized agents argue out an interpretation through a moderator-run forum, so no single model's blind spot dictates the conclusion. The bet is that disagreement between a search agent, a media-parsing agent, and a private-data agent surfaces signal that a lone model averages away.
What Sets It Apart
- A debate-moderator forum coordinates QueryEngine (web/news search), MediaEngine (video and image parsing), InsightEngine (private-database mining), and ReportEngine, so a claim has to survive cross-examination before it lands in the report.
- It reaches past text: MediaEngine deep-parses short-form video and pulls structured cards (weather, stocks, calendars) from images, which matters when discourse on Douyin or Xiaohongshu is mostly visual.
- The analysis layer mixes fine-tuned BERT, a GPT-2 LoRA, and small Qwen3 models alongside the LLM, rather than routing everything through one general model — cheaper inference on the narrow sentiment tasks.
- Secure APIs let you fuse a private business database with public crawl data, so internal metrics and external chatter sit in one report.
Who It's For
Great fit if you run brand, institutional, or policy monitoring across the Chinese social web and need explainable reports rather than a single sentiment number. Look elsewhere if you want a hosted SaaS or a quick score: this is a self-hosted Python framework that expects PostgreSQL/MySQL, your own LLM and search API keys, and WeasyPrint system dependencies for PDF export. It is also explicitly research- and education-scoped, not licensed for commercial use, and crawler reach varies by platform.
