{"id":14,"date":"2026-04-13T14:21:47","date_gmt":"2026-04-13T14:21:47","guid":{"rendered":"https:\/\/steve.id.vn\/?p=14"},"modified":"2026-04-13T14:21:47","modified_gmt":"2026-04-13T14:21:47","slug":"marketplace-pet-tool-huong-dan-su-dung-co-che-enrichment-ai","status":"publish","type":"post","link":"https:\/\/steve.id.vn\/?p=14","title":{"rendered":"Marketplace PET Tool \u2014 H\u01b0\u1edbng d\u1eabn s\u1eed d\u1ee5ng & C\u01a1 ch\u1ebf Enrichment AI"},"content":{"rendered":"

Marketplace PET (Product Enrichment Tool) l\u00e0 h\u1ec7 th\u1ed1ng n\u1ed9i b\u1ed9 t\u1ef1 \u0111\u1ed9ng h\u00f3a qu\u00e1 tr\u00ecnh l\u00e0m gi\u00e0u d\u1eef li\u1ec7u s\u1ea3n ph\u1ea9m (product attributes) cho c\u00e1c s\u00e0n marketplace \u2014 hi\u1ec7n t\u1ea1i \u0111ang v\u1eadn h\u00e0nh cho Target+<\/strong>, v\u1edbi roadmap m\u1edf r\u1ed9ng sang Walmart, Amazon v\u00e0 c\u00e1c s\u00e0n kh\u00e1c.<\/p>\n

\ud83d\udd17 Truy c\u1eadp n\u1ed9i b\u1ed9: https:\/\/steve.id.vn\/marketplace-pet<\/a><\/p>\n\n

\u2699\ufe0f C\u01a1 ch\u1ebf v\u1eadn h\u00e0nh \u2014 Pipeline 5 t\u1ea7ng<\/h2>\n

Thay v\u00ec d\u00f9ng AI cho m\u1ecdi th\u1ee9 (t\u1ed1n k\u00e9m, kh\u00f4ng nh\u1ea5t qu\u00e1n), PET s\u1eed d\u1ee5ng pipeline th\u00e1c n\u01b0\u1edbc 5 ngu\u1ed3n. M\u1ed7i t\u1ea7ng ch\u1ec9 \u0111\u01b0\u1ee3c g\u1ecdi khi t\u1ea7ng tr\u01b0\u1edbc kh\u00f4ng \u0111\u1ee7 \u0111\u1ed9 tin c\u1eady:<\/p>\n\n\n\n\n\n\n\n\n\n
#<\/th>\nNgu\u1ed3n<\/th>\nC\u01a1 ch\u1ebf<\/th>\n\u01afu \u0111i\u1ec3m<\/th>\n<\/tr>\n<\/thead>\n
1<\/strong><\/td>\nExistingData<\/strong><\/td>\nL\u1ea5y gi\u00e1 tr\u1ecb \u0111\u00e3 c\u00f3 t\u1eeb file upload g\u1ed1c<\/td>\nZero cost, t\u1ee9c th\u00ec<\/td>\n<\/tr>\n
2<\/strong><\/td>\nScraper<\/strong><\/td>\nFirecrawl scrape trang s\u1ea3n ph\u1ea9m th\u1ef1c t\u1ebf, AI extraction<\/td>\nD\u1eef li\u1ec7u tr\u1ef1c ti\u1ebfp t\u1eeb marketplace<\/td>\n<\/tr>\n
3<\/strong><\/td>\nRuleSignal<\/strong><\/td>\nKeyword \/ regex rules: “x-large” \u2192 X-Large, “polka dot” \u2192 Polka Dot<\/td>\n100% deterministic, kh\u00f4ng hallucinate<\/td>\n<\/tr>\n
4<\/strong><\/td>\nHistorical<\/strong><\/td>\nTra c\u1ee9u c\u00e1c QA \u0111\u00e3 approve tr\u01b0\u1edbc \u0111\u00f3 cho c\u00f9ng s\u1ea3n ph\u1ea9m\/lo\u1ea1i<\/td>\nT\u00e1i s\u1eed d\u1ee5ng d\u1eef li\u1ec7u \u0111\u00e3 ki\u1ec3m ch\u1ee9ng<\/td>\n<\/tr>\n
5<\/strong><\/td>\nAI (GPT-5.4-nano)<\/strong><\/td>\nSuy lu\u1eadn t\u1eeb t\u00ean s\u1ea3n ph\u1ea9m, m\u00f4 t\u1ea3, context \u0111\u00e3 scrape<\/td>\nFallback th\u00f4ng minh cho tr\u01b0\u1eddng h\u1ee3p kh\u00f4ng c\u00f3 rule<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

Pipeline d\u1eebng ngay khi m\u1ed9t ngu\u1ed3n \u0111\u1ea1t ng\u01b0\u1ee1ng confidence \u2265 0.7<\/strong> \u2014 ti\u1ebft ki\u1ec7m t\u1ed1i \u0111a chi ph\u00ed AI. K\u1ebft qu\u1ea3 t\u1eeb job g\u1ea7n nh\u1ea5t (1.094 SKUs):<\/p>\n

    \n
  • \ud83d\udcda Historical<\/strong>: 8.983 fields (~44%) \u2014 t\u00e1i d\u00f9ng d\u1eef li\u1ec7u c\u0169, zero AI cost<\/li>\n
  • \ud83d\udcd0 Rule-based<\/strong>: 5.479 fields (~27%) \u2014 deterministic, zero AI cost<\/li>\n
  • \ud83e\udd16 AI<\/strong>: 4.648 fields (~23%) \u2014 ch\u1ec9 g\u1ecdi khi c\u1ea7n thi\u1ebft<\/li>\n
  • \ud83d\udd17 Cross-attribute<\/strong>: 1.073 fields (~5%) \u2014 suy ra t\u1eeb attribute kh\u00e1c<\/li>\n
  • \ud83d\udd0d Scraper<\/strong>: 168 fields (~1%) \u2014 tr\u1ef1c ti\u1ebfp t\u1eeb trang s\u1ea3n ph\u1ea9m<\/li>\n<\/ul>\n

    \u2192 71% t\u1ed5ng s\u1ed1 fields \u0111\u01b0\u1ee3c fill ho\u00e0n to\u00e0n mi\u1ec5n ph\u00ed (kh\u00f4ng t\u1ed1n API call), ch\u1ec9 23% m\u1edbi c\u1ea7n \u0111\u1ebfn AI.<\/em><\/p>\n\n

    \ud83d\udcca Quy m\u00f4 x\u1eed l\u00fd<\/h2>\n\n\n\n\n\n\n\n\n\n\n\n
    Ch\u1ec9 s\u1ed1<\/th>\nCon s\u1ed1 th\u1ef1c t\u1ebf<\/th>\n<\/tr>\n<\/thead>\n
    SKUs \/ batch<\/td>\n1.094 s\u1ea3n ph\u1ea9m<\/td>\n<\/tr>\n
    Attributes \/ s\u1ea3n ph\u1ea9m<\/td>\n69 lo\u1ea1i attribute<\/td>\n<\/tr>\n
    Lo\u1ea1i s\u1ea3n ph\u1ea9m \u0111\u01b0\u1ee3c h\u1ed7 tr\u1ee3<\/td>\n28 product types<\/td>\n<\/tr>\n
    Fields enriched \/ batch<\/td>\n~20.205 fields<\/td>\n<\/tr>\n
    Average confidence score<\/td>\n0.851 \/ 1.00<\/td>\n<\/tr>\n
    Th\u1eddi gian x\u1eed l\u00fd 1.094 SKUs<\/td>\n~20\u201330 ph\u00fat<\/td>\n<\/tr>\n
    T\u1ed5ng rows \u0111\u00e3 x\u1eed l\u00fd (16 jobs)<\/td>\n17.504 rows<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

    V\u1ec1 l\u00fd thuy\u1ebft, pipeline c\u00f3 th\u1ec3 x\u1eed l\u00fd kh\u00f4ng gi\u1edbi h\u1ea1n batch song song<\/strong> \u2014 ch\u1ec9 b\u1ecb gi\u1edbi h\u1ea1n b\u1edfi rate limit c\u1ee7a OpenAI API v\u00e0 Firecrawl (d\u1ec5 d\u00e0ng scale b\u1eb1ng c\u00e1ch t\u0103ng worker).<\/p>\n\n

    \ud83c\udfaf Logic Enrichment \u2014 L\u00e0m sao bi\u1ebft gi\u00e1 tr\u1ecb n\u00e0o l\u00e0 \u0111\u00fang?<\/h2>\n

    Valid Values Constraint<\/h3>\n

    M\u1ed7i attribute c\u00f3 danh s\u00e1ch valid_values<\/strong> h\u1ee3p l\u1ec7 \u0111\u01b0\u1ee3c c\u1ea5u h\u00ecnh s\u1eb5n theo t\u1eebng marketplace v\u00e0 product type. AI v\u00e0 rule-based \u0111\u1ec1u ph\u1ea3i match v\u00e0o danh s\u00e1ch n\u00e0y \u2014 kh\u00f4ng th\u1ec3 t\u1ef1 b\u1ecba ra gi\u00e1 tr\u1ecb m\u1edbi. V\u00ed d\u1ee5: Product_Size<\/code> ch\u1ec9 ch\u1ea5p nh\u1eadn: XS, S, M, L, XL, X-Large, X-Small…<\/p>\n

    Confidence Scoring<\/h3>\n

    M\u1ed7i k\u1ebft qu\u1ea3 \u0111\u01b0\u1ee3c g\u00e1n \u0111i\u1ec3m confidence 0\u20131. Ch\u1ec9 khi confidence \u2265 ng\u01b0\u1ee1ng m\u1edbi \u0111\u01b0\u1ee3c accept:<\/p>\n

      \n
    • High risk attribute<\/strong> (Pattern, Size): ng\u01b0\u1ee1ng 0.85<\/li>\n
    • Medium risk<\/strong>: ng\u01b0\u1ee1ng 0.70<\/li>\n
    • Low risk<\/strong>: ng\u01b0\u1ee1ng 0.55<\/li>\n<\/ul>\n

      Parent-code Propagation<\/h3>\n

      S\u1ea3n ph\u1ea9m c\u00f9ng m\u1ed9t parent SKU (variations m\u00e0u\/size) t\u1ef1 \u0111\u1ed9ng inherit attribute t\u1eeb nhau \u2014 ch\u1ec9 c\u1ea7n enrich 1 l\u1ea7n, to\u00e0n b\u1ed9 variant \u0111\u01b0\u1ee3c c\u1eadp nh\u1eadt.<\/p>\n

      Human-in-the-loop Review<\/h3>\n

      Fields c\u00f3 needs_review=true<\/code> (confidence th\u1ea5p, high-risk attribute) \u0111\u01b0\u1ee3c \u0111\u01b0a v\u00e0o queue review. Reviewer xem evidence, approve ho\u1eb7c s\u1eeda. D\u1eef li\u1ec7u \u0111\u00e3 approve \u0111\u01b0\u1ee3c l\u01b0u v\u00e0o Historical \u0111\u1ec3 d\u00f9ng cho l\u1ea7n sau.<\/p>\n\n

      \ud83d\udcb0 Saving & Hi\u1ec7u qu\u1ea3<\/h2>\n

      N\u1ebfu l\u00e0m th\u1ee7 c\u00f4ng:<\/h3>\n
        \n
      • 1 ng\u01b0\u1eddi data entry: ~5\u20137 ph\u00fat\/SKU \u00d7 69 attributes = ~100 gi\u1edd c\u00f4ng cho 1.094 SKUs<\/strong><\/li>\n
      • Chi ph\u00ed nh\u00e2n s\u1ef1 (outsource data entry ~$3\u20135\/gi\u1edd): $300\u2013$500 \/ batch<\/strong><\/li>\n
      • Error rate th\u1ee7 c\u00f4ng: ~8\u201315%<\/li>\n<\/ul>\n

        V\u1edbi PET Tool:<\/h3>\n
          \n
        • Th\u1eddi gian: 20\u201330 ph\u00fat<\/strong> (ch\u1ea1y t\u1ef1 \u0111\u1ed9ng, kh\u00f4ng c\u1ea7n gi\u00e1m s\u00e1t)<\/li>\n
        • Chi ph\u00ed API (OpenAI + Firecrawl): ~$2\u20135 \/ batch 1.094 SKUs<\/strong><\/li>\n
        • Accuracy: ~88.7% t\u1ef1 \u0111\u1ed9ng<\/strong>, ph\u1ea7n c\u00f2n l\u1ea1i qua review queue<\/li>\n
        • Consistency: 100% \u2014 c\u00f9ng input \u2192 c\u00f9ng output, kh\u00f4ng sai do human error<\/li>\n<\/ul>\n\n\n\n\n\n\n\n\n
          <\/th>\nTh\u1ee7 c\u00f4ng<\/th>\nPET Tool<\/th>\nSaving<\/th>\n<\/tr>\n<\/thead>\n
          Th\u1eddi gian<\/strong><\/td>\n100 gi\u1edd<\/td>\n30 ph\u00fat<\/td>\n-99.5%<\/strong><\/td>\n<\/tr>\n
          Chi ph\u00ed<\/strong><\/td>\n$300\u2013500<\/td>\n$2\u20135<\/td>\n-99%+<\/strong><\/td>\n<\/tr>\n
          Accuracy<\/strong><\/td>\n85\u201392%<\/td>\n88.7% auto + review<\/td>\n\u2248 t\u01b0\u01a1ng \u0111\u01b0\u01a1ng ho\u1eb7c t\u1ed1t h\u01a1n<\/td>\n<\/tr>\n
          Scale<\/strong><\/td>\nLinear (th\u00eam ng\u01b0\u1eddi)<\/td>\nSub-linear (th\u00eam worker)<\/td>\nKh\u00f4ng gi\u1edbi h\u1ea1n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n\n

          \ud83d\uddfa\ufe0f Next Steps \u2014 Roadmap ti\u1ebfp theo<\/h2>\n

          Ng\u1eafn h\u1ea1n (Q2 2026)<\/h3>\n
            \n
          • \u2705 GPT-5.4-nano<\/strong> \u2014 v\u1eeba n\u00e2ng c\u1ea5p t\u1eeb GPT-4.1-nano, c\u1ea3i thi\u1ec7n classification accuracy ~1\u20132%<\/li>\n
          • \ud83d\udd04 Fix scraping pipeline<\/strong> \u2014 t\u0103ng t\u1ef7 l\u1ec7 extract th\u00e0nh c\u00f4ng material specs v\u00e0 gender t\u1eeb trang s\u1ea3n ph\u1ea9m (+3\u20135% accuracy cho Apparel_Material, Size_Grouping)<\/li>\n
          • \ud83d\udd04 M\u1edf r\u1ed9ng signal rules<\/strong> \u2014 th\u00eam rules cho Garment_neckline_type, Top_Style, Swimwear_Style<\/li>\n<\/ul>\n

            Trung h\u1ea1n (Q3 2026)<\/h3>\n
              \n
            • \ud83d\uddbc\ufe0f Vision model cho pattern detection<\/strong> \u2014 d\u00f9ng AI nh\u00ecn \u1ea3nh s\u1ea3n ph\u1ea9m \u0111\u1ec3 ph\u00e2n bi\u1ec7t Wave vs Polka Dot, Solid vs Color Block (hi\u1ec7n t\u1ea1i l\u00e0 265 errors kh\u00f3 fix b\u1eb1ng text)<\/li>\n
            • \ud83c\udfea Walmart integration<\/strong> \u2014 m\u1edf r\u1ed9ng sang marketplace th\u1ee9 2, reuse to\u00e0n b\u1ed9 pipeline, ch\u1ec9 c\u1ea7n add attribute specs m\u1edbi<\/li>\n
            • \ud83d\udcc8 Target accuracy 92%+<\/strong> \u2014 k\u1ebft h\u1ee3p vision + scraping improvements<\/li>\n<\/ul>\n

              D\u00e0i h\u1ea1n (Q4 2026+)<\/h3>\n
                \n
              • \ud83e\udd16 Feedback loop<\/strong> \u2014 review actions c\u1ee7a human t\u1ef1 \u0111\u1ed9ng train l\u1ea1i signal rules<\/li>\n
              • \ud83c\udf10 Multi-marketplace dashboard<\/strong> \u2014 manage enrichment jobs cho Amazon, Walmart, Target+ trong m\u1ed9t UI<\/li>\n
              • \ud83d\udd0c API integration<\/strong> \u2014 k\u1ebft n\u1ed1i tr\u1ef1c ti\u1ebfp v\u1edbi PIM systems (Salsify, Syndigo) \u0111\u1ec3 enrich realtime khi c\u00f3 SKU m\u1edbi<\/li>\n<\/ul>\n\n

                B\u00e0i vi\u1ebft \u0111\u01b0\u1ee3c t\u1ed5ng h\u1ee3p t\u1eeb d\u1eef li\u1ec7u th\u1ef1c t\u1ebf c\u1ee7a 16 enrichment jobs v\u1edbi 17.504 rows \u0111\u00e3 x\u1eed l\u00fd \u2014 c\u1eadp nh\u1eadt th\u00e1ng 4\/2026.<\/em><\/p>\n","protected":false},"excerpt":{"rendered":"

                Marketplace PET (Product Enrichment Tool) l\u00e0 h\u1ec7 th\u1ed1ng n\u1ed9i b\u1ed9 t\u1ef1 \u0111\u1ed9ng h\u00f3a qu\u00e1 tr\u00ecnh l\u00e0m gi\u00e0u d\u1eef li\u1ec7u s\u1ea3n ph\u1ea9m (product attributes) cho c\u00e1c s\u00e0n marketplace \u2014 hi\u1ec7n t\u1ea1i \u0111ang v\u1eadn h\u00e0nh cho Target+, v\u1edbi roadmap m\u1edf r\u1ed9ng sang Walmart, Amazon v\u00e0 c\u00e1c s\u00e0n kh\u00e1c. \ud83d\udd17 Truy c\u1eadp n\u1ed9i b\u1ed9: https:\/\/steve.id.vn\/marketplace-pet \u2699\ufe0f C\u01a1 ch\u1ebf […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-14","post","type-post","status-publish","format-standard","hentry","category-chua-phan-loai"],"blocksy_meta":[],"_links":{"self":[{"href":"https:\/\/steve.id.vn\/index.php?rest_route=\/wp\/v2\/posts\/14","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/steve.id.vn\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/steve.id.vn\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/steve.id.vn\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/steve.id.vn\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=14"}],"version-history":[{"count":0,"href":"https:\/\/steve.id.vn\/index.php?rest_route=\/wp\/v2\/posts\/14\/revisions"}],"wp:attachment":[{"href":"https:\/\/steve.id.vn\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=14"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/steve.id.vn\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=14"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/steve.id.vn\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=14"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}