progress
SCOPING: Shopify app store reviews for ReferralCandy.
DATA SOURCE:
- Shopify app store page: publicly visible reviews
- Shopify Partner API has a reviews endpoint (GraphQL)
- Alternative: apps like Appfigures or manual CSV export from partner dashboard
PROPOSED ARCHITECTURE:
- reviews/ directory in merchant-knowledge repo
- reviews/reviews.jsonl — one JSON object per review
- Schema: {id, store_name, store_url, reviewer_name, rating, date, title, body, rc_response, rc_response_date, linked_merchant_id}
- reviews/index.md — summary stats (avg rating, rating distribution, top themes)
EMBEDDING STRATEGY:
- Add reviews to the same embeddings.npz (unified vector space)
- Each review becomes a document: 'Review by [store] ([rating] stars): [title]. [body]'
- Tag metadata with source='review' so semantic search can distinguish from call transcripts
- Estimated: if 1000 reviews, adds ~1000 docs to the 400 existing = 1400 total
- Build time still <1 min with local model
MCP TOOL OPTIONS:
1. Extend merchant_semantic_search — reviews show up alongside transcript matches (preferred — one search surface)
2. Add merchant_reviews(query, rating_filter, limit) — dedicated review search tool
3. Both — unified search + dedicated review tool for filtering by rating
PROFILE LINKING:
- Match review store_name to existing merchant profiles via fuzzy matching
- Add linked_merchant_id to review record
- When viewing a merchant_profile, could show associated reviews
NEXT STEPS:
1. Check Partner API access for review export
2. If no API, scrape public review page or use partner dashboard CSV
3. Build ingest script
4. Add to embedding pipeline
5. Extend MCP server