glam/V5_QUICK_REFERENCE.md

V5 Extraction - Quick Reference

Status: ✅ 75% PRECISION ACHIEVED

Architecture

Conversation Text → Subagent NER → V5 Validation → Clean Institutions (75% precision)

What Works

• ✅ Subagent NER: Clean, accurate names (no mangling)
• ✅ V5 Validation: 3 filters (country, organization, proper name)
• ✅ 75% precision: 3/4 correct (up from V4's 50%)

What Doesn't Work

• ❌ Pattern-based extraction: 0% precision (names mangled)

Commands

Run V5 demonstration:

bash /Users/kempersc/apps/glam/scripts/demo_v5_success.sh

Test subagent + V5 integration:

python /Users/kempersc/apps/glam/scripts/test_subagent_v5_integration.py

Subagent NER Prompt Template

Extract ALL heritage institutions from the following text.

Return JSON array with:
{
  "name": "Full institution name",
  "institution_type": "MUSEUM | ARCHIVE | LIBRARY | GALLERY",
  "city": "City name",
  "country": "2-letter ISO code",
  "isil_code": "ISIL code if mentioned",
  "confidence": 0.0-1.0
}

Rules:
1. Preserve full names (e.g., "Van Abbemuseum", not "Abbemuseum")
2. Classify by primary function
3. Determine country from city names or context
4. Exclude: organizations, networks, generic descriptors

Next Steps for Production

1. Implement extract_from_text_subagent() in InstitutionExtractor
2. Update batch extraction scripts
3. Process 139 conversation files

Files

• Documentation: output/V5_VALIDATION_SUMMARY.md
• Session Summary: SESSION_SUMMARY_V5.md
• Test Scripts: scripts/test_subagent_v5_integration.py
• Demo: scripts/demo_v5_success.sh

---

Result: V5 achieves 75% precision via subagent NER + validation filters