Grounding recipes

Grounding hard nouns

Put the exact spelling in instructions and the model only has to get close. A name or a drug it has no reason to expect gets transcribed as whatever it sounds most like. Give it the list your callers will actually draw from, spelled the way you want it recognized and said back.

instructions: ground names and terms against a known list

The caller may ask for one of these doctors: Dr. Nikola Mrksic, Dr. Amara
Okafor, Dr. Liu Wei. The caller may name one of these medications:
Metoprolol, Lisinopril, Atorvastatin. If what you hear is close to one of
these names, resolve it to that name and say it back the way it's spelled
here, regardless of how it sounded.

Adapt the two lists to whatever your callers actually name: branches, plans, product lines, staff. Keep the list to names that are realistically confusable and don’t change every call, a roster or a formulary rather than one caller’s own contacts; for a list that’s genuinely different per caller, pass it in as part of that call’s setup instead of growing a static list that only some callers ever need.

Why this works: fused transcription and generation resolving toward what instructions already said was plausible.

Overview

Quickstart

Prompt guide

Prompt cookbook

Evaluations

Reference

Guides

Grounding hard nouns

Next steps

Prompt guide

Reference

​Grounding hard nouns

​Next steps

Prompt guide

Reference

Grounding hard nouns

Next steps