Dolly Expert Lite

Dolly Expert Lite is an open-source application that leverages the power of the Dolly-v2-3b model, an open-source large language model (LLM), to provide domain-specific question answering using retrieval from custom-made vector databases.

Dolly Expert Builder

This version of our project focuses on answering construction-related questions based on the Ontario Building Code.

The main aim of Dolly Expert Lite is to showcase the potential of open-source LLM applications and demonstrate that they can provide valuable assistance to various professionals, such as tradespeople, engineers, architects, and more. By utilizing Dolly-v2-3b and the LangChain framework, we have developed a lightweight prototype that can retrieve information from the Ontario Building Code and provide coherent and factually correct guidance to users.

To build Dolly Expert Lite, we utilized the following components:

• Model: Dolly-v2-3b for lightweight development and deployment.
• Framework: LangChain, which integrates with ChromaDB, to create the question answering chain.
• Text Embeddings: We used Text Embeddings by Weakly-Supervised Contrastive Pre-training (E5-base-v2), known for its high score in retrieval tasks.
• The Ontario Building Code: We extracted information from the official government code and converted it into a vector store for retrieval.
• Vector Store: ChromaDB for efficient storage and retrieval of vectors.
• Deployment: We deployed Dolly Expert Lite on HuggingFace Spaces and used Gradio for the user interface.

During the development process, we encountered several challenges, such as working with new tools and dealing with divergences from standard use cases. However, we were able to overcome these challenges through troubleshooting and implementing workarounds. Despite the difficulties, we are proud to have created a functional lightweight prototype of an LLM application and successfully deployed it as a HuggingFace Space. This accomplishment demonstrates that open-source models like Dolly can be just as capable as closed-source alternatives.