Affiliate Product Catalog &
In-App Product Suggestions

Executive Summary

Cressi already suggests products to users, but those suggestions are pulled live from Google Shopping on every request — which means ongoing API cost, latency, no control over the catalog, and no affiliate revenue.

This project loads your existing affiliate product feeds (Mytheresa, Coutr, and ~1,700 further advertisers available on demand) into Cressi’s existing vector search, and serves suggestions catalog-first — falling back to the current live search only when the catalog has no good match. Every product link is an affiliate tracking link, so suggestions become a revenue stream rather than a cost centre.

The Opportunity

Source data verified: Mytheresa feed ≈ 37k unique products, Coutr feed ≈ 36k unique products (each delivered as size-level rows that we collapse into single products). A third file maps 1,696 additional advertisers we can onboard later.

How We’ll Build It

The solution is an extension of Cressi’s existing pipeline — the same vector database (Qdrant), the same embedding model, and the same product carousel the front-end already renders. We are not building a parallel system, which is what keeps the timeline short and the risk low.

Feed files (Excel / pipe-text)
   │   1 ·  parse → normalize → merge size variants → keep in-stock
   ▼
Common product records
   │   2 ·  generate embeddings (existing model) → upsert
   ▼
Qdrant "products" collection  (filterable: gender · category · price · brand · stock)
   │
User request ─▶ assistant ─▶ 3 · vector search + filters
                                  ├─ good matches?  ──▶ serve from catalog  (affiliate links)
                                  └─ too few?        ──▶ fall back to live search (today’s behaviour)
                                                        ▼
                                            Product carousel  (unchanged)

1 · Ingestion — a generic affiliate-feed parser

The feeds all follow the standard Rakuten layout but differ in delivery (headerless Excel vs. pipe-delimited text). We build a config-driven parser: a small column-map per feed, large files streamed row-by-row, size-level rows merged into one product carrying its available sizes, out-of-stock filtered out, and the affiliate URL preserved exactly. Onboarding a new merchant later is a config block, not new code.

2 · Embedding & catalog store

A dedicated products collection in the existing Qdrant, populated with the current embedding model. Products are batch-embedded and upserted with a stable ID, so re-loading a refreshed feed updates prices and stock in place rather than duplicating.

3 · Catalog-first serving with safety net

A new search service embeds the user’s request, runs a similarity search with filters (gender, category, price), and serves the matches. If the catalog can’t satisfy the request well, it transparently falls back to the existing live search — so coverage never gets worse than today. Results map onto the front-end’s existing product shape, so no UI changes are required.

4 · Refresh & merchant onboarding

A scheduled job re-ingests updated feeds (prices/stock are time-sensitive) and prunes sold-out items. Additional advertisers are added from the supplied directory by dropping in a feed and its column map.

Delivery Plan & Estimates

Phases run as parallel workstreams (ingestion/store alongside serving), so calendar time is shorter than the summed effort. One-time initial embedding of ~73k products is machine compute (a few hours), not developer time, and is included in Phase 2.

Investment

Billed at a blended engineering rate of $30 / hour (an 8-hour developer-day = $240/day). The estimate is based on a build effort of 2.5–3.5 developer-days.

Time-and-materials at $30/hour, billed against actual hours within the estimated range; a fixed-fee option is available on request. Excludes third-party costs (Qdrant hosting, embedding/LLM API usage, affiliate-network fees), which run on your existing accounts.

Assumptions, Dependencies & Risks

• Feed cadence must be agreed (daily/weekly). Prices and stock in the files are time-sensitive; without scheduled refresh they go stale.
• Assumes feeds continue to arrive in the standard Rakuten format. A genuinely new format is a small mapping add-on, not a rebuild.
• Builds on the existing Qdrant instance and embedding model already running in Cressi.
• Affiliate compliance: tracking links are preserved exactly, and a disclosure notice (e.g. affiliate-link disclosure) may be required — a brief legal check is recommended.
• Estimate assumes one developer; a second developer can parallelize Phases 1–2 against Phase 3 to compress the calendar further.

Scope

Next Steps