Capability CardsA-layerField Extraction

搜推广能力抽取卡

把硬证据从资源存在推进到能力边界、硬事实、风险缺口和下一步抽取字段。

5cards

A 层资源能力卡

每张卡保留 TOP3 core ideas，并固定下一步抽取字段。

A

Yandex commerce ranking

Boundary: Personalized transaction ranking and recommender-style reranking for commerce surfaces.

Evidence: src_yandex_ecommerce_ranking, company_yandex_ranking

TOP3 core ideas

  1. Ranking quality depends on user activity and context, not only query relevance.
  2. Transformer-based ranking can be used in two-stage commerce recommendation/ranking flows.
  3. The resource should be evaluated through task definition, baseline, metric and online effect fields.

Hard facts

  • arXiv metadata ready
  • representative analysis ready
  • full paper extraction pending

Risk/gap: No metric table or full-text experiment extraction yet.

taskfeaturesmodelbaselineoffline metricsonline metrics
A

Yambda-5B dataset

Boundary: Large-scale multi-modal ranking and retrieval benchmark asset.

Evidence: src_yambda_5b, company_yandex_yambda

TOP3 core ideas

  1. Dataset depth is stronger than platform-name evidence.
  2. Multiple feedback types and temporal behavior are central to recommendation evaluation.
  3. A benchmark card must record scale, fields, tasks, license and download path.

Hard facts

  • metadata reports billion-scale interactions
  • representative analysis ready
  • dataset field extraction pending

Risk/gap: License, field dictionary and download endpoint still need hard extraction.

scalemodalitiesfeedback typessplit protocollicensedownload URL
A

VK-LSVD

Boundary: Industrial short-video recommendation benchmark and data asset.

Evidence: src_vk_lsvd, company_vk_lsvd

TOP3 core ideas

  1. Short-video recommendation needs sequence and interest-drift modeling.
  2. Industrial data should include user-video interactions, content features and context.
  3. Benchmark evidence gives VK a deeper role than generic social/video platform evidence.

Hard facts

  • arXiv metadata ready
  • representative analysis ready
  • challenge/data page pending

Risk/gap: Need full-text task and field extraction before calling it a complete benchmark card.

interaction schemavideo featurescontext fieldstasksmetricslicense
A

2GIS local discovery promotion

Boundary: Local intent, merchant discovery and map/directory promotion surface.

Evidence: src_2gis_ads_candidate, company_2gis_local_discovery

TOP3 core ideas

  1. Local discovery is a distinct search-promo intent layer.
  2. Merchant promotion must be decomposed into visibility, location and action conversion.
  3. Regional fit is part of the capability boundary, not an afterthought.

Hard facts

  • official page snapshot ready
  • representative analysis ready
  • format/case extraction pending

Risk/gap: No pricing, format, city coverage or customer case extraction yet.

product formatsmerchant workflowcoveragepricingcase studiesconversion actions
A-candidate

WB seller promotion

Boundary: Seller-facing marketplace promotion operation and product education surface.

Evidence: extra_wb_promo_instructions, company_wb_promotion

TOP3 core ideas

  1. Seller workflow is a hard fact for marketplace promotion capability.
  2. Dynamic pages may require CDP supplement rather than HTTP-only snapshots.
  3. Marketplace promotion should be paired with search-quality and fairness risk analysis.

Hard facts

  • official seller page ready
  • CDP supplement ready
  • representative analysis ready

Risk/gap: Need section-level extraction and comparison with other marketplace sources.

promotion typesentry pointsseller rulesquality controlscost/metric fieldsrisk notes