AI Assistant for Business Classification

AI Assistant for Business Classification

This assistant helps users find KBLI codes, Indonesia’s classification of business activities used in licensing. Answers carry official sources users can check. The system evaluates the relationship between an activity and source text before recommending a code. Its output supports review rather than replacing licensing decisions.

Challenge

General AI assistants can produce convincing codes without suitable evidence. Classification sources span many documents, including sector regulations and mappings between versions. Users need traceable answers alongside clear explanations.

Solution

I built source retrieval before answer generation. This approach is called RAG: the system retrieves relevant information and gives it to the model. Like opening a reference book before answering, the model must use the sources it finds.

Official definitions are split into retrievable records. Search combines exact terminology with the meaning of business descriptions. Results are ranked again before reaching the model.

The model uses tools to find codes, retrieve details, compare candidates, and review a user’s existing code list. Each result carries its source page. A separate checking stage requires a quotation from the official definition before a code can be judged relevant.

Costs are limited per task. Recommendation, quotation checks, and information enrichment have separate model choices. Daily limits and processing deadlines control heavy requests. At capacity, the system tells users to try again.

Impact

Users receive answers with sources they can verify. The system separates primary and supporting activities and flags codes unsupported by reference text. Verification status is recorded for each conversation so the team can review failures to substantiate answers.