Technical Report · GoWave Automation Lab · 2026

Scan-to-Link: inferring the product behind an unknown bottle barcode with prefix k-NN

A lightweight, training-free recommender that turns the digits of an unscanned EAN-13/UPC-A liquor barcode into a ranked list of catalog products — so a shopkeeper can capture real barcodes by tapping a suggestion instead of typing. It exploits the GS1 company-prefix structure and a labeled corpus that grows every time a link is confirmed. Built into LiKAR's point-of-sale.

1. Problem

Indian liquor lists (e.g. the West Bengal Excise registry) publish brand name, pack size and MRP — but no barcodes. Public barcode databases (Open Food Facts, etc.) barely cover Indian IMFL, and when they do they return the foreign-market code (a French Heineken 871…) rather than the Indian-retail 890… bottle. The authoritative source (GS1 India) sells a company prefix to issue your owncodes (₹48k+/yr) — useless for looking up someone else's products. So a liquor POS starts with thousands of products and almost no scannable barcodes.

The only 100%-correct source is the bottle itself. The question this work answers: when a clerk scans a bottle we have never seen, can we guess which catalog product it is — from the digits alone — well enough to make capture a single tap?

2. Signal in the digits

An EAN-13 is not random. Under GS1, the leading digits are the company prefix (the brand owner), followed by an item reference the company assigns to each SKU — typically in contiguous blocks, so size variants of one product sit on adjacent numbers.

Therefore the longest common prefix (LCP)between an unknown code and a known one is a direct proxy for “same maker / same brand / same SKU family.” This is the kernel of our model.

3. Model

We use an instance-based (lazy) k-Nearest-Neighbors classifierwith a custom prefix-similarity kernel — no parameters are trained by gradient descent; the “model” isthe labeled corpus, and it updates online. Concretely:

• Similarity kernel: sim(x, c) = LCP(x, c) / 13 over the digit strings, with item-reference numeric distance as a tie-break.
• Neighbourhood: the k=25 corpus codes with the highest LCP (requiring LCP ≥ 6).
• Voting: neighbours vote for a brand and a category, weighted by similarity → the predicted manufacturer prefix (first 7 digits, used only when a neighbour shares ≥7), brand and category.
• Candidate scoring: each catalog product without a real barcode is scored — 1.0 × LCP/13 if it shares the brand of a neighbour (strong), or 0.5 × LCP/13 if only the manufacturer + category match (weak) — then ranked.
• Online bootstrapping (self-supervision): every confirmed link adds a new labeled instance, so accuracy and coverage rise with use. This is a form of human-in-the-loop active learning.

4. Dataset

The corpus is 775 real EAN-13/UPC-A codes with known product, brand and category, drawn from the official Assam price master and West-Bengal cross-matches shipped in LiKAR. Evaluation uses the 782 labeled Assam codes.

5. Evaluation & results

Protocol: leave-one-out cross-validation. Each code is removed, then ranked against the remaining corpus; we record the rank of its true manufacturer and true brand. We report top-kaccuracy on all codes and on the recoverable subset (where ≥2 codes share the label, i.e. a non-cold-start case). The gap between them is pure data sparsity, which the online corpus closes over time.

Reading the numbers. The manufacturer is identified top-1 in 91.4% of cases (100% whenever the maker is not a singleton). Because the manufacturer is known, the clerk sees a short, correct shortlist. The true brand is the single best guess 82.7% of the time on recoverable codes, rising to 97.8% within the top-10— i.e. on a normal scan the right product is almost always one tap away. The lower “all” brand numbers reflect cold-start singletons (a brand seen for the first time has no sibling to match) — exactly the cases the bootstrapping loop fixes.

6. Worked example

7. How it is built into LiKAR

The corpus is the union of the baked seed file and every product in the shop's database that already carries a real numeric barcode— so each confirmed link (and every state's data, e.g. Assam helping West Bengal) compounds into better suggestions. Scanning of already-known barcodes is unchanged; this only assists the unknown case.

8. Related work & prior research

We surveyed the landscape before claiming novelty. Three lines of prior work touch this problem:

• Authoritative lookup (GS1). GEPIR — now Verified by GS1 (2024) — resolves a GTIN to its registered brand-owner and attributes. It is reverse lookup and depends on the brand having uploaded data; GS1 itself notes new, discontinued or small-manufacturer codes are often absent. For Indian liquor it is sparse and, when present, returns the origin-market code, not the Indian retail bottle.
• Commercial / crowd databases. Go-UPC, EAN-Search, Barcode Spider, Open Food Facts etc. aggregate GTIN→product records. Same reverse direction, same coverage gap for Indian IMFL (empirically confirmed in §1).
• Prefix-based brand validation (patents). Prior patents (e.g. US 9,679,321 / 10,872,366 on product-identification validation) use the observation that a brand's products share a common UPC prefix — but to validate whether a newly submitted product legitimately belongs to a brand, given an existing catalog. The prefix↔brand association is therefore established prior art; we do not claim it.

9. Novelty & positioning

Being precise about the delta: the GS1 prefix↔brand-owner association and prefix / k-NN string matching are known (see §8) — we invent neither. Our contribution is the composition: turning prefix k-NN into a forward barcode→product recommender (unknown code → ranked catalog candidates, not a yes/no validation or a registry lookup), wired into a self-bootstrapping scan-to-link loop that runs with no external database, no API and no GS1 fee, for a domain (Indian liquor retail) that the lookup services above do not cover. We did notfind this specific forward-recommender-plus-bootstrapping workflow in the prior art or in any shipping Indian liquor POS we examined; to our knowledge it is novel in this domain — but we state that as a surveyed claim, not an unverifiable “world-first.” The measured top-k results in §5 are the substantive contribution.

10. Limitations

• Cold start. A brand/maker absent from the corpus yields no suggestion until one bottle is linked manually (then it self-heals).
• Prefix length varies. GS1 issues shorter prefixes to large firms and pooled GTINs to small ones; we use longest-common-prefix + nearest-neighbour rather than a fixed cut to absorb this.
• Brand ≠ SKU. The model is confident on maker and brand; the exact size variant relies on the human tap (a 2–4 item shortlist).
• Regional makers. WB-local country-liquor producers are under-represented in the Assam-seeded corpus and start cold.
• Evaluated on a 775/782-code corpus; absolute numbers will shift as the corpus scales, but the structure (high manufacturer, top-k brand) is stable.

11. Reproducibility

All artifacts ship in the LiKAR repository: the algorithm (lib/barcode-suggest.ts), the corpus (data/barcode-corpus.json), and the APIs/UI above. The evaluation is leave-one-out over the labeled corpus; the metrics in §5 are produced directly from it. The live tool is at /dashboard/inventory/scan-link.