enrich-catalog

Enrich Catalog Metadata

The catalog's style field is empty until enriched — the build-time normalizers can map category/color/material from structured data, but style is a judgment call. This skill drives the enrich CLI: export products missing style, classify them yourself, import the results. No re-embedding happens — only metadata is updated in place.

Run where the catalog DB lives (the directory containing catalog_vector_db/), with chromadb installed. In Cowork the host-side MCP server owns the DB; run this skill in local Claude Code on that host machine instead.

Steps

1. Export products missing style:

   python3 "${CLAUDE_PLUGIN_ROOT}/scripts/catalog/catalog_vectordb.py" enrich export --missing style --out pending.jsonl
   

   Each line: id, name, category, description excerpt.

2. Classify in batches of ~30 lines. Read a batch, judge each product's style from the controlled vocab:

   scandinavian, midcentury, modern, industrial, rustic, traditional, coastal, japandi, mediterranean, bohemian, eclectic

   Judge from name/category/description signals — leg style (tapered light wood → scandinavian/midcentury; hairpin/black metal → industrial; turned/carved → traditional), line language (low-slung straight → midcentury; curved minimal → modern/japandi; ornate → traditional), materials (rattan/jute → coastal/bohemian; raw pine/reclaimed wood → rustic; terracotta/wrought iron → mediterranean).

   When truly ambiguous, leave style empty rather than guessing — wrong metadata silently corrupts every later filtered search; missing metadata just falls back to judgment at selection time.

3. Optionally fix color_family / primary_material in the same pass when the build-time mapping is clearly wrong for a product you're already looking at.

4. Write enriched.jsonl — one JSON object per line:

   {"id": "...", "style": "japandi"}
   {"id": "...", "style": "midcentury", "color_family": "natural_wood"}
   {"id": "...", "style": "", "color_family": "beige"}
   

   Optional override keys: color_family, primary_material. An empty style leaves the style unset (deliberately unclassified) while still applying the other overrides; for products with nothing to change, omit the line entirely.

5. Import:

   python3 "${CLAUDE_PLUGIN_ROOT}/scripts/catalog/catalog_vectordb.py" enrich import --in enriched.jsonl
   

6. Report counts: exported / classified / left empty, plus a per-style breakdown. Spot-check with catalog_search using a style filter.

Notes

• Runtime scales with catalog size — thousands of products means many batches. Work one category at a time (group the exported JSONL lines by their category field, or pass a category as the argument and enrich just those), and import after each category so progress persists if you stop.
• A category's products share signals (all sofas, all rugs) — judging them together is faster and more consistent than mixed batches.
• Re-running is safe: --missing style only exports products still unclassified, so deliberately-empty styles will re-export — skip ids you already judged ambiguous.