Classifying Sales-Call Text with Embeddings for Portable CRM Actions (and Trade-offs)

To classify sales call text using embeddings, generate vector representations for a small set of labeled examples. For each new transcript or text chunk, create a vector and compare it to the labeled examples using cosine similarity. This comparison is a decision problem, not a label prediction. Store the decision separately from the provider client, allowing for flexibility in switching embedding providers without altering tagging rules or CRM workflows.

Use a gateway with a single API key and billing plan to minimize credential and invoice complexity. The pipeline consists of labeled examples (shelves) with their corresponding call notes (texts). A new text is placed on the shelf with the closest examples in terms of vector similarity. Cosine similarity is a practical first step, but treat it as evidence, not a definitive answer.

Validate decisions against real labeled examples and maintain an audit record of the top labels, their scores, the embedding model, and the version of the example set. This approach allows for inspecting disagreements between the classifier and sales/operations teams. The provided Node.js classifier demonstrates how to use the OpenAI embeddings API with Infrai's openai npm package, enabling easy integration with various backend services.