Skip to end of metadata
Go to start of metadata

You are viewing an old version of this content. View the current version.

Compare with Current View Version History

« Previous Version 5 Next »

Overview

As we seek to improve and streamline user experiences across departments of the University, chatbots have been proposed by various departments as a means to accomplish this goal. While there are many different chatbot offerings in the market today, TTS seeks to create a single, unified chatbot in-house to serve all departments, leveraging Generative AI.

Benefits

  • Unified experience: Users only have to engage with one chatbot, regardless of the department they seek to engage with.

  • Centralized Design/Development: The chatbot will follow a singular design pattern and set of best practices.

  • Fine-tuned for Tufts: Models can be fine tuned to Tufts data and provide a more reliable/accurate experience.

  • Cost Savings: Chat bots in general can yield savings by making certains roles or systems obsolete. Additionally, in-house development can yield cost savings over off-the-shelf chatbots and models.

  • Time savings: Staff can spend less time on mundane, repetitive tasks, and spend more time on meaningful, interesting projects.

Risks

  • Data: “Garbage in, garbage out” The training of any model hinges on accurate, high quality data in an optimal format. If a department lacks this, it can hinder development and even worse, lead to unreliable model output. Data can broken down into the following aspects:

    • Accuracy: The data for a process is accurate and up-to-date

    • Volume: There’s an adequate amount of data to train, per model requirements

    • Format: Data is in an adequate format for training

  • Hallucinations: If a model lacks adequate data on a particular topic it can potentially provide false information. Tuning of model parameters (e.g temperature) can help minimize hallucinations.

  • Infrastucture Costs

Design

The chatbot will be designed around a modular architecture, with each department being assigned it’s own model trained and fine-tuned on it’s specific data. This will allow for modular development and training of models, minimizing the impact to models of other departments.

On top of the department-specific models will be a routing model, trained on a set of distinct chat scenarios for each department for the purposes of routing chat to the correct department model.

User will interact with the AI ensemble via chat interface which can be embedded on University web pages or made available via mobile app.

All code for the models and chat app will be centrally housed in a GitHub repository (e.g jumbo-bot).

  • No labels