Brian has two primary goals: providing fast and easily accessible information about Web 3 and assisting everyone in interacting with DApps within the ecosystem. How does he achieve this?
During this project's development, we faced challenges in extracting user intent and parameters from queries using NLP techniques. Our goal was to use local large-scale language models to guarantee data privacy. We researched open-source models like Falcon, Flan-T5, CodeStar, and exploited libraries like LangChain, Kor, HuggingFace, and PandasAI. However, tested models didn't provide accurate parameters. To overcome this, we manually created a high-quality annotated dataset leveraging our web3 expertise. We expect improved model performance with this dataset, delivering precise and consistent results. Our commitment is to seek innovative solutions to advance web3-focused natural language processing.
Another challenge we encountered during the project was converting user intents and parameters generated from NLP into a format that could be used to generate an ethers call to the blockchain and execute the transaction for the user. We explored various methods, including parsing the generated output and mapping it to the appropriate function and parameter format required by ethers.
By addressing this challenge, we aimed to provide users with a prompt experience, where they can simply express their intentions in natural language and have Brian generate the corresponding ethers call to execute the transaction seamlessly. This approach eliminates the need for users to manually construct complex transaction calls.
Tracks Applied (2)
Discussion