Scholarlink

This study investigates the automation of an embedding-based video recommendation system in Spanish through the application of Large Language Models (LLMs). The primary objective was to determine the optimal embedding model, search strategy, and chunking parameters for a system designed to deliver the top 5 most relevant videos in response to user queries. To achieve this, we developed an automated testing framework. A dataset of 400 randomly selected videos from a pool of 50,000 was used. GPT-4 was employed to generate 10 questions per video, which were subsequently refined to remove irrelevant queries. These questions served as the basis for evaluating and comparing the performance of various models and configurations. The evaluation metric measured the percentage of instances in which the system accurately retrieved the source video within the top 1 and top 5 recommendations. This evaluation was conducted across different embedding models, search strategies, and chunking parameters.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Optimizing Embedding-Based Video Recommendation in Spanish Using LLMs for Automated Model and Parameter Selection

  • 会议论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年

查看全文
  • Ignacio Despujol Zabala,
  • Carlos Turró Ribalta,
  • Jaime Busquets Mataix,
  • Sergio Puche García

摘要

This study investigates the automation of an embedding-based video recommendation system in Spanish through the application of Large Language Models (LLMs). The primary objective was to determine the optimal embedding model, search strategy, and chunking parameters for a system designed to deliver the top 5 most relevant videos in response to user queries. To achieve this, we developed an automated testing framework. A dataset of 400 randomly selected videos from a pool of 50,000 was used. GPT-4 was employed to generate 10 questions per video, which were subsequently refined to remove irrelevant queries. These questions served as the basis for evaluating and comparing the performance of various models and configurations. The evaluation metric measured the percentage of instances in which the system accurately retrieved the source video within the top 1 and top 5 recommendations. This evaluation was conducted across different embedding models, search strategies, and chunking parameters.