Structured Chatbot Search for Gaming Supplier Invoices Prioritizes Field Accuracy

A developer argues that a gaming company's chatbot for finding supplier invoices must return correct, specific fields rather than plausible-sounding answers. The proposed design pairs embedding-based semantic search to retrieve relevant invoice passages with a chat model that outputs schema-constrained JSON. Keeping retrieval and extraction separate makes it easier to tell whether errors come from finding the wrong evidence or mapping it wrongly. The article recommends preserving source identifiers, validating output, and marking absent fields as missing, and it proposes evaluating exact field accuracy on held-out invoices. It also compares OpenAI, Google Vertex AI, Cohere, Pinecone, and an OpenAI-compatible gateway by setup cost and fit.