LLMs, GenAI, RAG, agents, evaluation · insight-0528
Give every extracted field an explicit unknown value so the model is not forced to invent
- Evidence type
- practitioner heuristic
- Difficulty
- easy
- Roles
- Applied AI, ML engineer, Data analyst
- Industries
- aviation, insurance
Summary: situation, action, reason, boundary
Situation
An extraction schema requires a booking reference, a fare class and a travel date for every complaint email.
Action
Make fields nullable or add an explicit not-stated option, instruct the model to use it when the text lacks the value, and add eval cases where each field is genuinely missing.
Reason
Required fields with no way to say absent turn missing information into a generation task, which the model completes with plausible values.
Boundary
Document and email extraction, form filling, CRM enrichment.
Symptom or risk
Emails that never mention a fare class still produce one, because a required string must be filled, and the invented values look identical to real ones.
Exceptions
Fields that are always present by construction, such as a message ID supplied by your system, can stay strictly required.
Trade-offs
Nullable fields push missing-data handling downstream, which must then be designed.
Pitfalls
Measuring extraction accuracy only on emails where every field is present.
Example
In a synthetic test of 200 complaint emails, the fare-class field would be checked separately on the subset where the email never states a class; any non-null value there is an invention.