A multilingual workflow needs locale-specific inputs, policy text, examples, and evaluation cases. Translation alone does not prove equivalent behavior. Keep invariant business rules in structured fields and test the response in each supported language, including uncertainty and refusal states. Ask for the original evidence ID when summarizing a translated source. If the source language is uncertain or a key legal term has no agreed translation, report the ambiguity and send it for review instead of assuming an English interpretation fits every locale.
Make this comfortable
Multilingual prompts: test policy meaning across languages
Decision in practice
A benefits team receives claims in Hindi and English under one plan rule: dependent eligibility ends after the recorded cutoff date. The structured cutoff field is shared; each locale has reviewed wording and independent test cases. One Hindi example includes a date expressed in a local format, so the ingestion layer normalizes it before the model reasons about coverage. The team compares the decision and cited evidence, not just the fluency of the explanation. If the document's relationship term is ambiguous, both routes return needs-review.
Invariant fields: plan ID, cutoff date, relationship code.
Locale inputs: reviewed terms and date parsing.
Evaluate: approved, denied, unknown in each supported language.
Pass: same case facts produce the same decision and evidence IDs.Performance and operating cost
Each supported locale adds review and evaluation cost; N cases across L locales require at least O(NL) decisions for a full matrix. Prioritize high-volume and high-consequence slices, then add adversarial paraphrases for known ambiguities. Track decision parity separately from grammatical quality. A model may write polished text in a language while applying the wrong rule, so human reviewers with domain fluency should inspect disputed cases.
Common Mistakes
- Do not use fluent translation as proof of decision parity.
- Do not compare dates before normalizing their formats.
- Do not silently resolve an ambiguous relationship term.
Connected lessons
- Prompt Engineering
- Production prompt engineering
- Prompt inputs: normalize records before asking for conclusions
- Evaluation sets: measure the failure cases that matter
- Prompt privacy: send only the fields needed for the task
- Project: defend a retrieval and action workflow
- Prompt production decisions
Continue with: Cross-locale evaluation: compare decisions, not word-for-word text.
Continue with: Survey prompts: check wording against the intended construct.
Continue with: Localization prompts: preserve variables and plural branches.
prompt engineering
tutorial