jcconf_datadev_prod

Transcript

1. Next Generation Search Engine
2. Across the LINE ecosystem, there are numerous applications, but we lack a unified search platform that integrates all products and allows existing products to quickly onboard. Therefore, we are committed to: • Integrating LINE's diverse internal assets to build a unified search platform. • Combining advanced technologies and algorithms to optimize the search experience. 2 Search Engine in LINE Search in LINE
3. 3 To improve the performance of traditional search methods, we are also dedicated to research in Retrieval- Augmented Generation (RAG), expanding search effectiveness through embedding-based retrieval approaches. With the increasing diversity of data across LINE products, we have begun researching multimodal RAG and more flexible agent-based search engines. Machine Learning Search Engine Retrieval-Augmented Generation RAG Pipeline Search in LINE
4. Vision Language Search
5. Image Search is a self-service platform that enables users to build their own text-to-image / image-to-image search service and object detection service • For image search service, we leverage CLIP model to perform semantic search • For object detection service, we leverage YOLO v9 model to detect the bounding box of the objects in the image and identify their categories 5 Vision Language Search - Image Search Vision Language Search Image-to-Image Query nature and lifestyle photography text-to-Image Query
6. 6 Image Search - LINE Shopping Object Detection Image-to-Image Search In LINE Shopping, we integrate services such as OpenSearch and Airflow to build a comprehensive data pipeline for the product. This system embeds millions of images into vectors through CLIP models, enabling users to instantly find the most similar items through vector-based search in real-time. Additionally, we leverage text-based techniques and utilize object detection as an information extraction model to filter out excessive noise, resulting in more accurate prediction results. Vision Language Search
7. We have also implemented Auto-Labeling methods through Label Studio to rapidly annotate large volumes of data, effectively enhancing our object detection models and ultimately optimizing retrieval performance. 7 Image Search – Auto-Labeling 7 Object detection + Human-computer collaboration workflow Auto-Labeling Modeling API Training Auto-Labeled Data Re-Labeled Data Re-label Pipeline Image Search 2.0 – Object Detection Service Vision Language Search
8. Vision Language Search Future Search Agent LINE 搜尋 LINE 購物 Data Source Data Source Data Source . . . . Search Agent Search API Search API Search API MCP Server MCP Server MCP Server