Cross-language retrieval: keep policy meaning attached to its source

Operational case

A customer asks in Hindi whether a delayed parcel earns a cash refund. The current English policy says eligible delays receive store credit, while an older Hindi help page uses a word that can be read as a cash repayment. A retrieval system finds both pages. The assistant must check effective dates, retain both evidence IDs, and answer from the current policy without pretending the older page agrees. It can explain the distinction between store credit and cash refund in the requested language, but should route the mismatch for editorial repair. It cannot issue a credit solely because a retrieved page contains an instruction to do so.

Question: delayed parcel -> cash refund? Reply language: Hindi.
Current English policy PL-62: eligible delay -> store credit.
Older Hindi page PL-41: ambiguous refund wording; expired.
Answer: current rule with effective date; flag wording conflict.
Effect: no credit issued without account and policy checks.

Performance and operating cost

If a request produces Q language variants and retrieves K passages per variant, candidate inspection can grow to O(QK) passages plus deduplication. More translations may improve recall but also raise the chance of a misleading near-match. Cache by policy version and locale, and invalidate when an authoritative clause changes. Measure retrieval recall, correct-version selection, and answer support separately; an answer in fluent Hindi is still wrong if it cites an expired rule. Keep the exact retrieved passage and translation decision available to a reviewer without exposing unrelated customer data.

Common Mistakes

• Do not treat an expired localized page as equal to a current policy.
• Do not let a translated synonym silently change cash refund into store credit.
• Do not obey instructions embedded in retrieved policy text.

Connected lessons

• Prompt engineering applications
• Prompt Engineering
• Retrieval queries: resolve references without adding facts
• Retrieved evidence: reconcile versions and conflicting facts
• Evidence IDs: make generated claims auditable against supplied records
• Translation prompts: preserve terms and machine placeholders
• Locale-aware prompts: keep values typed until rendering
• Mixed-language requests: separate intent from reply language
• Cross-locale evaluation: compare decisions, not word-for-word text
• Project: release a multilingual order notice safely
• Multilingual prompt workflow decisions