Scholarlink

This paper describes the design and implementation of the Smart Psychologist Assistant, a data-centric web application that automates routine clinical workflows and augments the decision-making of mental-health professionals. The system leverages a modular client–server architecture in which large language models (LLMs)—GPT-4 (Turbo) from OpenAI and high-throughput Groq deployments of open-weight models—are orchestrated through function calling and tool-use pipelines. Text embeddings generated by the LLMs are stored in PostgreSQL with the PGVector extension, enabling sub-second semantic search across patient dialogues, therapist notes, and knowledge-base articles. Benchmarking on a synthetic workload that mirrors medium-size practices (10 therapists, 500 patients) demonstrates average end-to-end response times below 400 ms and horizontal scalability to thousands of concurrent sessions. The open, service-oriented design facilitates future extensions such as emotion-state inference, multimodal data ingestion, or the integration of additional evidence-based therapeutic frameworks. The proposed approach illustrates how LLM-driven vector databases combined with modern web technologies can create practical, secure, and extensible tools that advance data-centric innovation in psychological support services.

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

Data-Centric Innovations for Psychological Support: Integrating LLMs and Vector Search in Professional Web Applications

  • 章节

  • 发表日期：

    2026年

查看全文
  • Inna Baraban,
  • Serhii Baraban,
  • Olha Voitsekhovska,
  • Mariia Baraban,
  • Nazarii Shportun

摘要

This paper describes the design and implementation of the Smart Psychologist Assistant, a data-centric web application that automates routine clinical workflows and augments the decision-making of mental-health professionals. The system leverages a modular client–server architecture in which large language models (LLMs)—GPT-4 (Turbo) from OpenAI and high-throughput Groq deployments of open-weight models—are orchestrated through function calling and tool-use pipelines. Text embeddings generated by the LLMs are stored in PostgreSQL with the PGVector extension, enabling sub-second semantic search across patient dialogues, therapist notes, and knowledge-base articles. Benchmarking on a synthetic workload that mirrors medium-size practices (10 therapists, 500 patients) demonstrates average end-to-end response times below 400 ms and horizontal scalability to thousands of concurrent sessions. The open, service-oriented design facilitates future extensions such as emotion-state inference, multimodal data ingestion, or the integration of additional evidence-based therapeutic frameworks. The proposed approach illustrates how LLM-driven vector databases combined with modern web technologies can create practical, secure, and extensible tools that advance data-centric innovation in psychological support services.