Skip to main content Skip to navigation

SlicerChat: Building a Local Chatbot for 3D Slicer

Project Overview

The document examines the implementation of SlicerChat, an innovative local chatbot designed to assist users of the 3D Slicer software by leveraging generative AI to navigate its complex documentation. It addresses the significant challenges faced by new users who often struggle with the software's intricacies and highlights the potential of large language models (LLMs) to offer timely and relevant assistance. The project delves into strategies for enhancing the chatbot's effectiveness, including fine-tuning LLMs, utilizing Retrieval Augmented Generation (RAG), and optimizing overall model performance. However, the document also acknowledges critical issues associated with generative AI, such as tendencies for hallucination, dependence on outdated information, and privacy concerns, which need to be managed to ensure the chatbot’s reliability and user trust. Overall, the findings suggest that while generative AI holds great promise for improving educational tools like SlicerChat, careful consideration of its limitations is essential for successful implementation in educational contexts.

Key Applications

SlicerChat - a local chatbot for 3D Slicer

Context: Educational context for users of 3D Slicer, including both new users and experienced developers.

Implementation: SlicerChat was built as a custom extension in 3D Slicer using the CodeLlama Instruct architecture, optimized through fine-tuning and RAG techniques.

Outcomes: Improved access to 3D Slicer documentation, faster and more relevant answers to user queries, and support for both high-level and technical questions.

Challenges: Issues with hallucination (fabricating non-existent methods), reliance on outdated training data, and privacy concerns related to data use.

Implementation Barriers

Technical

Hallucination in LLMs, leading to inaccurate or fabricated responses.

Proposed Solutions: Integrate high-quality documentation sources and optimize RAG to minimize hallucination.

Data Privacy

Concerns regarding data and intellectual property privacy with the use of LLMs.

Proposed Solutions: Utilize open-source models that can be run locally to maintain control over data.

Computational Resources

High computational demands of larger models can limit accessibility for users.

Proposed Solutions: Implement parameter-efficient methods like Low Rank Adaptation for fine-tuning.

Project Team

Colton Barr

Researcher

Contact Information

For information about the paper, please contact the authors.

Authors: Colton Barr

Source Publication: View Original PaperLink opens in a new window

Project Contact: Dr. Jianhua Yang

LLM Model Version: gpt-4o-mini-2024-07-18

Analysis Provider: Openai

Let us know you agree to cookies