1688 → AliExpress AI Product Matcher (Vision + LLM Verified)

Give this Actor a 1688 product (title + image) and it returns ONE AI-verified AliExpress equivalent — the single product you can actually use as a supplier or competitor — not a raw list of look-alikes for you to sort through.

Every candidate is put through a two-stage AI pipeline: a vision model (OpenCLIP) scores visual + title similarity, then an LLM (DeepSeek) runs a two-pass judgment to confirm the match is genuinely the same or a close substitute. Only a product that survives both stages is returned.

Retrieval runs only through the official AliExpress Affiliate API — no scraping, no reverse-engineered endpoints, no CAPTCHA bypass, no proxy chains.

Built for cross-border sourcing agents, dropshippers, and resellers who need a decision, not a search result.

---

This is a matcher, not a scraper

Most 1688 / AliExpress Actors are scrapers: you give a keyword or an image and get back 20–100 candidates, then you spend your own time filtering out the false positives.

This Actor is a matcher. It does the filtering for you with AI and commits to an answer:

	Typical scraper / image-search Actor	This AI matcher
Output	20–100 candidates to sort yourself	ONE selectedCandidate (or an honest null)
Match verification	None — you decide	Vision (OpenCLIP) + LLM (DeepSeek) two-pass judgment
Chinese-title handling	Weak (queries on Chinese keywords)	English searchQuery override built in
Retrieval source	Often reverse-engineered mobile APIs / proxies	Official AliExpress Affiliate API only
What you do with it	More manual work	Drop the pair straight into sourcing

Use a scraper if you want candidates and will sort them yourself. Use this if you want one verified supplier equivalent per SKU and don't want to hunt through false positives.

---

What you get

For every source SKU you send in, the Actor returns one result row containing:

• selectedCandidate — the single AliExpress product accepted as a high-confidence match (or null if none passed the AI judgment).
• candidates — up to 20 AliExpress candidates retrieved for the SKU (id, title, image, product URL).
• openClipRanking — image+title similarity ranking (OpenCLIP ViT-B/32) for the candidates.
• top5Judgments — the AI dual-judgment verdict for the top 5 candidates.
• retrievalOutcome — success, no-result, failed, skipped, or stop-no-go.
• complete, stopReason, degradations — per-SKU status and diagnostics.

Run-level records are written to the key-value store: SUMMARY, CALL_LEDGER, CHARGE_LEDGER, and OUTPUT.

---

Coverage — read this before you run

Roughly half of 1688 products have a usable AliExpress equivalent. On a frozen 100-SKU cross-category test, about 50% of source SKUs produced a usable AliExpress candidate in the top 5.

This is not a limitation of the AI — 1688 and AliExpress are different catalogs. 1688 carries many white-label, factory, and China-domestic items that simply are not listed on AliExpress (a retail-facing subset). Coverage depends on whether the product actually exists on AliExpress.

What the AI does guarantee:

• Match quality — when a match is returned, it has passed a vision + LLM dual judgment, so you get very few false positives.
• Honest nulls — when no candidate is good enough, you get selectedCandidate: null with a clear reason instead of a forced, low-quality guess.
• Retrieval stability — the AliExpress query layer returns a terminal answer (a match, or a confirmed "no equivalent") for essentially every SKU.

We never claim "guaranteed same product" or a near-100% hit rate. You only pay for SKUs we actually deliver a result for (see Pricing).

AliExpress search works best in English. 1688 titles are usually Chinese, and a Chinese query returns loosely related results. For best recall, send a short English searchQuery (e.g. "baking tools") — it overrides the title as the search term. If you only send a Chinese title with no searchQuery, recall drops well below the ~50% above.

---

Input

{
  "products": [
    {
      "sourceSkuId": "1688-974333356844",
      "title": "Silicone Scraper Integrated Spatula Spatula Food Grade Baking Tools Cake Cutter Cream Scraper Home Shovel",
      "imageUrl": "https://cbu01.alicdn.com/img/ibank/O1CN01FLMKAx25xePqKFZFQ_!!2218359047593-0-cib.jpg",
      "searchQuery": "baking tools",
      "leafCategory": "baking-tools"
    }
  ]
}

Field	Required	Notes
products	yes	1–100 items.
products[].sourceSkuId	yes	Stable identifier for the source 1688 product.
products[].title	yes	Source product title (Chinese is fine — see the note above).
products[].imageUrl or products[].imageBase64	yes	Provide exactly one. imageUrl must be HTTPS.
products[].searchQuery	no	Overrides the query used to search AliExpress. Recommended for best recall.
products[].leafCategory	no	Category hint used by the category-fallback retrieval step.

maxTotalCostUsd is an internal safety ceiling for the run's own external cost — it is not your bill.

---

How the AI pipeline works

1. Retrieve — query the official AliExpress Affiliate API (aliexpress.affiliate.product.query), 1 request/sec, with a category-fallback step, capped at 20 candidates per SKU.
2. Rank (vision) — score every candidate against the source image and title with OpenCLIP ViT-B/32 (0.8 × image similarity + 0.2 × title overlap).
3. Judge (LLM) — send the top 5 to a vision-capable LLM for a two-pass judgment (first read + independent confirmation); only a candidate both passes agree on is marked "high".
4. Select — the first high-confidence candidate becomes selectedCandidate; if none qualifies, the result is an honest null.

The run checkpoints continuously and resumes safely after a migration or timeout without re-charging or re-doing completed SKUs.

---

Pricing (pay-per-event)

• No startup fee. Failed runs cost $0.
• On a successfully completed run you pay a small base fee once, plus a per-delivered-result fee.
• A SKU is a delivered result only when we return a terminal answer for it (a match, or a confirmed "no AliExpress equivalent"). SKUs we could not process are not charged.

Event	When it's charged
run-base-fee	Once per successful run that delivered at least one result.
sku-result	Once per delivered SKU result — this includes a confirmed "no AliExpress equivalent", because you are paying for the AI-verified decision (match or no-match), not for a match. SKUs we could not process are never charged.

You are paying for an AI-verified decision per SKU (vision + LLM), not for raw scraped rows. Charges settle only at the end of a completed run and are idempotent across resume.

---

Limits & compliance

• Up to 100 SKUs per run.
• AliExpress data comes only from the official Affiliate API.
• Source images must be HTTPS URLs or Base64.
• Results reflect the AliExpress catalog at run time; availability and prices change.

---

FAQ

How is this different from a reverse image search on AliExpress? Reverse image search hands you 20–100 visually similar products and leaves the filtering to you. This Actor uses image similarity as just one signal (OpenCLIP), then asks an LLM (DeepSeek) to actually verify whether a candidate is the same product — same features, same function, same target user — and returns a single verified match.

Why didn't SKU X find a match? The product likely isn't listed on AliExpress, or the title/image was too generic. Add a focused English searchQuery to improve recall.

Do I pay if nothing matches? You pay per delivered result (which includes a confirmed "no equivalent" — itself useful information). You never pay for a failed run or for SKUs that couldn't be processed.

Why can selectedSkuCount be lower than completeSkuCount? completeSkuCount counts every SKU that reached a reliable terminal answer: either a verified match or a confirmed no-equivalent. selectedSkuCount counts only the subset with a non-null selectedCandidate. For example, completeSkuCount: 1 with selectedSkuCount: 0 means the SKU was fully processed and no suitable AliExpress equivalent passed verification.

Is AliExpress scraped? No. Retrieval is 100% the official AliExpress Affiliate API — no scraping, no CAPTCHA bypass, no reverse-engineered endpoints.

My 1688 titles are in Chinese — will this work? Yes, that's the intended case. Pass the Chinese title and a short English searchQuery. The English query drives AliExpress retrieval; the Chinese title and image drive AI verification.

Can I send images directly? Yes — use imageBase64 instead of imageUrl.

How do I find the AliExpress equivalent of a 1688 product?

Send the 1688 product's title + image and the Actor returns the single AI-verified AliExpress equivalent (or a confirmed no-equivalent), not a candidate list to sort.

Can I match a 1688 product photo to AliExpress?

Yes. The photo drives OpenCLIP image ranking and the LLM confirmation. If you only have a photo and no 1688 source yet, the 1688 Image Search Scraper finds the 1688 product first; then this Actor maps it to AliExpress.

Can I bulk match 1688 SKUs to AliExpress?

Yes. Submit up to 100 source products in products. Each SKU receives an independent terminal result: one verified equivalent or a confirmed no-equivalent.

Can I check whether a private-label product already sells on AliExpress?

Yes. Provide the source product's title and image. A verified match indicates that the same or a genuinely substitutable product is already listed; a no-equivalent result means none of the retrieved candidates passed the dual judgment.

Does this return an AliExpress price band?

No. This Actor verifies product identity; it does not calculate a price band. The selected result includes the AliExpress product URL so you can inspect current listing prices directly.

Can I use this as an AliExpress product finder?

Yes, when you start with a known 1688 product. It is not a general keyword scraper: it finds and verifies the closest AliExpress equivalent to the supplied source product.