Kigyou-list pipeline tags 1,000+ B2B categories

The work is framed as a response to a specific gap: Japan's standard industry classification (JSIC) is too coarse to distinguish modern B2B businesses. Startups in LLM deployment support, supply-chain DX or virtual human development all land in the same broad bucket, "information services," so finding the right companies meant manually inspecting hundreds of websites.

The proposed fix is a three-stage cascade rather than sending every company to an LLM. Keyword and regex filters run first, then vector similarity against a predefined taxonomy, and only ambiguous or novel cases reach a fast LLM through the Groq API with LLaMA-3. The taxonomy itself is hierarchical, spanning broad categories down to narrow ones, so both wide and pinpoint searches work.

Operating at a scale of roughly five million companies, the build reports that caching website hashes to skip unchanged firms cut monthly crawl and reclassification costs by 85%. Layering in public-sector open data such as IT subsidy awards and patent categories as training signals pushed tag confidence above 98%. The system stores tags in PostgreSQL arrays and uses a GIN index for millisecond filtering.