Problem

When extracting dates from Japanese Wikipedia articles, the LLM returns dates in Japanese format like 16世紀 (16th century) instead of the expected YYYY format, causing validation errors:

Error extracting [<PropertyType.BIRTH_DATE: 'P569'>, <PropertyType.DEATH_DATE: 'P570'>] with LLM: 2 validation errors for PropertyExtractionResult
properties.0.value
  Value error, Invalid date format: '16世紀'. Must be YYYY, YYYY-MM, or YYYY-MM-DD

Current Behavior

The prompts in poliloom/prompts.py do specify the expected format:

- birth_date: Use format YYYY-MM-DD, or YYYY-MM, YYYY for incomplete dates
- death_date: Use format YYYY-MM-DD, or YYYY-MM, YYYY for incomplete dates

However, the LLM sometimes:

1. Returns dates in non-Western formats (Japanese characters instead of Arabic numerals)
2. Returns century-level precision which has no valid format

Proposed Solution

Strengthen the date extraction prompts to:

1. Explicitly require Arabic numerals (0-9) for all date values
2. Specify behavior when only century-level precision is available (e.g., skip extraction or use approximate year like "1550" for 16th century)

Affected Files

• poliloom/poliloom/prompts.py - DATES_EXTRACTION_SYSTEM_PROMPT
• Potentially also POSITIONS_EXTRACTION_SYSTEM_PROMPT for position start/end dates