How Yahoo enhances search retargeting using Amazon Bedrock 雅虎如何使用亚马逊Bedrock增强搜索再定位

TL;DR

• Yahoo replaced its legacy Word2Vec + Locality-Sensitive Hashing (LSH) keyword expansion system with a generative AI-powered solution using Amazon Bedrock to improve search retargeting accuracy and reach.
• Anthropic’s Claude 3.5 Sonnet v2 was selected as the optimal foundation model for generating semantically relevant keyword expansions, achieving a fivefold improvement in median broad expansion ratio and doubling maximum expansion under strict similarity thresholds.
• A verification step was introduced to mitigate LLM hallucinations by converting expanded keywords into embeddings and filtering those below a cosine similarity threshold with the original seed keywords, ensuring semantic fidelity and trustworthiness.

Why It Matters

This case study demonstrates how large language models can be effectively integrated into real-world advertising infrastructure to solve complex semantic matching problems at scale. For AI practitioners and ad-tech engineers, it highlights practical strategies for deploying generative AI in production environments while maintaining control over output quality, relevance, and compliance—critical considerations when handling sensitive user data and advertiser intent signals.

Technical Details

• The legacy system relied on Word2Vec embeddings combined with locality-sensitive hashing (LSH) for keyword expansion, which suffered from outdated vocabulary, phrase-level focus, syntactic bias, and frequent zero-expansion failures.
• The modernized architecture leverages Amazon Bedrock to access multiple foundation models (including Amazon Nova, Meta Llama, and Anthropic’s Claude) without custom infrastructure, enabling rapid experimentation and deployment.
• Claude 3.5 Sonnet v2 was chosen after evaluation via Amazon SageMaker Studio due to its superior balance between expansion breadth and semantic coherence relative to seed keywords.
• A post-generation validation layer converts all LLM-generated keywords into embeddings and computes cosine similarity against the original seed set; terms falling below a predefined threshold are discarded to filter out irrelevant or hallucinated outputs.
• Expanded keywords alongside originals are stored in an Amazon OpenSearch Service cluster for downstream batch scoring workflows that match users’ historical search behavior to advertiser-defined audience segments.

Industry Insight

Adopting managed LLM services like Amazon Bedrock allows enterprises to rapidly prototype and deploy AI-driven features such as dynamic keyword expansion without heavy engineering overhead, accelerating time-to-value for computational advertising use cases. Implementing guardrails—such as embedding-based similarity checks—is essential to maintain reliability and prevent degradation in targeting quality caused by model hallucinations, especially in high-stakes domains where precision directly impacts ROI and brand safety.

TL;DR

• 雅虎利用 Amazon Bedrock 和生成式 AI（特别是 Claude 3.5 Sonnet v2）重构了其搜索再营销（SRT）系统的关键词扩展能力，替代了传统的 Word2Vec + LSH 方法。
• 新方案解决了旧模型词汇过时、仅关注短语而非实体、依赖句法相似性以及零扩展等核心痛点，显著提升了语义理解能力和关键词覆盖广度。
• 通过引入基于嵌入相似度验证的 Guardrailing 机制，有效降低了大语言模型的幻觉风险，确保了广告内容的安全性与合规性。
• 部署后中位宽泛扩展比率提升了五倍，最大扩展比率翻倍，且在严格相似度阈值下仍能保持高产出，大幅优化了广告投放的精准度与受众规模。
• 该实践展示了如何利用无服务器化平台快速实验多种基础模型，并在计算广告领域平衡创新效率与隐私政策合规性。

为什么值得看

本文对数字广告从业者具有重要参考价值，它揭示了如何将前沿的大语言模型技术落地到实际业务场景中，解决传统机器学习方法在语义理解和动态适应性上的局限。同时，其关于 Guardrailing 设计和多模型评估的方法论，为其他希望在保障安全前提下引入 GenAI 的团队提供了可复用的工程范式。

技术解析

• 架构演进：原系统采用 Word2Vec 生成词向量并结合局部敏感哈希（LSH）进行关键词扩展，流程包括标准化、向量化、索引、搜索、过滤与排序；新架构则完全迁移至基于 Amazon Bedrock 的生成式 AI 管道，由 LLM 直接输出语义相关的扩展关键词集。
• 模型选型：在 SageMaker Studio 中测试了 Amazon Titan、Meta Llama 及 Anthropic Claude 等多个模型，最终选定 Claude 3.5 Sonnet v2，因其在扩展数量与语义相关性之间取得最佳平衡。
• 质量控制机制：为防止 LLM 产生无关或错误关键词，新增验证步骤——将生成的关键词转为嵌入向量，并与原始种子关键词计算余弦相似度，低于设定阈值的项被自动剔除，确保输出质量。
• 性能指标对比：相比旧方法，新方案在相同条件下实现中位扩展比提升 5 倍，最大扩展比翻倍，且能在原本无法产生任何扩展词的案例中成功生成数百甚至数千个相关关键词。
• 安全与合规集成：系统设计内置敏感词过滤、用户偏好屏蔽、黑名单匹配等功能模块，确保所有扩展结果符合广告行业的数据保护法规及品牌安全标准。

行业启示

• 生成式 AI 正在重塑程序化广告中的用户意图识别方式，从静态规则驱动转向动态语义推理，建议广告技术厂商积极布局 LLM 赋能的实时扩词与人群画像构建能力。
• 在引入大模型时必须同步建立“护栏”机制（如后处理校验、置信度评分、人工审核接口），以避免不可控输出带来的品牌风险或法律隐患，这是规模化落地的关键前提。
• 云厂商提供的托管型 AI 服务（如 Amazon Bedrock）显著降低了企业调用复杂模型的技术门槛与运维成本，使得中小规模团队也能高效完成原型验证与生产迭代，推动整个生态的技术民主化进程。