supply-training-personas

Supplying training personas

The mechanism being changed

Training utterances are written by an LLM role-playing a persona. By default that persona is drawn uniformly from the whole PersonaHub corpus (~200k rows) and conditioned on nothing — a retail workflow and an identity-governance workflow get the same population of astrophysicists, sommeliers and marine biologists. Most drawn personas have no relationship to the vocabulary the application's users actually employ.

personas.json is how an application states who its users are. It changes the pool; the sampling itself is unchanged and stays deterministic.

Two modes, one file

personas.json contains	Source	Effect
a non-empty personas list	app_supplied	Your personas are used verbatim. PersonaHub is never consulted or downloaded.
domain_keywords (or domain) and no personas	domain_conditioned	PersonaHub is filtered to rows mentioning those keywords; the usual deterministic sample is taken from the survivors.
neither	—	error, not a fallback

An explicit personas list wins over domain_keywords — supplying personas is the stronger statement, and honouring the keywords as well would mix generic rows into a curated set.

The file

Discovered at <workflow>/personas.json. Both forms parse:

{
  "schema_version": 1,
  "domain": "retail order management",
  "domain_keywords": ["retail", "e-commerce", "shopper", "customer service"],
  "personas": [
    {"id": "impatient-shopper", "persona": "A shopper who orders often, tracks every delivery, and types in short fragments."},
    {"id": "support-agent", "persona": "A support agent handling returns and refunds all day; uses internal jargon."}
  ]
}

• A bare JSON array of personas is accepted, so a quick file needs no wrapper.
• An entry may be a bare string instead of an object; its id becomes its position.
• An object entry takes its text from persona or text, and its id from id or its position.
• schema_version must be 1.
• domain prose with no domain_keywords is usable on its own — its own words (longer than two characters, lowercased) become the keywords. That is the least-effort form of the feature.