Skip to main content Link Menu Expand (external link) Document Search Copy Copied

Meeting Summary for CS 486-686 Lecture/Lab, Spring 2025

Date: Feb 13, 2025
Time: 07:58 AM Pacific Time (US and Canada)
Meeting ID: 893 0161 6954

Quick Recap

  • AI Tools & Systems Discussed:
    • Chat Assistant for Families
    • Mindfulness Chatbot for Mental Well-Being
    • AI Prompt Refiner Tool
    • Stock Market Analysis System
  • Team Projects Presented:
    • Blog Assistant Tool
    • Chinese Party Game
    • AI Code Tutor
  • New Initiatives:
    • AI Code Search System introduced by Greg
    • Shreya’s work on calculating the cost of LM calls

Next Steps

  • Greg’s Action Items:
    • Finalize details for the upcoming RAG project focused on source code search and augmentation.
    • Provide additional information about the fast HTML library for building Python-only web apps.
    • Consider sharing students’ projects with faculty members.
  • Students’ Action Items:
    • Prepare for the natural language-based code search and location augmentation RAG project.
    • Explore the fast HTML library for potential applications in future projects.
    • Continue refining coding skills, especially with AI coding assistants and incremental development techniques.

Detailed Discussion

1. Zoom Screen Sharing & Order Presentation

  • Greg addressed issues with Zoom screen sharing and navigating the system.
  • The topic shifted to an order for the campus choir, with a focus on Ian:
    • Recommendation: Ian should introduce his assistant, assign a name, explain its purpose, and conclude with a discussion of challenges or learnings from working with LLMs and APIs.

2. Greg’s Chat Assistant for Families

A chat assistant was presented, designed to accommodate families sharing one account by tailoring responses based on age groups.

  • Key Features:
    • Standard ChatGPT functionalities.
    • Customized responses for kids, teenagers, adults, and elderly individuals.
    • Support for asynchronous streaming from a lightweight language model.
  • Open Questions:
    • How does the assistant determine the current date?
    • Does it augment user input with additional context?

Mermaid Diagram – Chat Assistant for Families Flow:

flowchart TD
    A[User Input] --> B{Determine Age Group}
    B -- Kid --> C[Provide Kid-Friendly Response]
    B -- Teenager --> D[Provide Teen-Friendly Response]
    B -- Adult --> E[Provide Adult-Friendly Response]
    B -- Elderly --> F[Provide Senior-Friendly Response]
    C --> G[Asynchronous Streaming]
    D --> G
    E --> G
    F --> G

3. Mindfulness Chatbot for Mental Well-Being

The mindfulness chatbot was demonstrated as a tool to offer personalized mental health support.

  • Features:
    • Enables users to select models through a UI with predefined conversation starters.
    • Dynamically generates buttons based on user responses.
    • Summarizes the user’s journey with an inspiring quote, interaction summary, and mental health metrics.
    • Built entirely with pre-existing models and templates, removing the need for custom coding.

Mermaid Diagram – Mindfulness Chatbot Interaction Flow:

flowchart TD
    A[User Initiates Chat] --> B[Selects Model]
    B --> C[Chatbot Provides Predefined Messages]
    C --> D[User Responds]
    D --> E[Dynamic Button Generation]
    E --> F[Conversation Continues]
    F --> G[Generate Conversation Summary & Metrics]

4. AI Prompt Refiner Tool Demo

The AI prompt refiner tool was shown comparing unrefined versus refined prompt responses.

  • Workflow:
    • The initial prompt is sent concurrently to both an answering model and a refining model.
    • The refined prompt is then resubmitted to the answering model.
    • Observation: Smaller language models tend to struggle with prompt refinement, whereas instruction-tuned models provide improved results.
    • The UI was generated using root code that successfully adhered to the specified color scheme.

Mermaid Diagram – AI Prompt Refiner Workflow:

sequenceDiagram
    participant U as User
    participant AM as Answering Model
    participant RM as Refining Model
    U->>AM: Submit Initial Prompt
    U->>RM: Submit Initial Prompt
    RM->>U: Return Refined Prompt
    U->>AM: Submit Refined Prompt
    AM->>U: Return Enhanced Answer

5. Project Analysis and Stock Market Models

Greg discussed two related projects:

  • Play Analysis Project:
    • Utilizes OpenAI and Google image generation.
    • Encountered bugs were discussed, along with an application providing positive feedback regardless of input.
  • Stock Market Analysis System:
    • Employs two different models:
      • Generator: Provides the initial response.
      • Reviewer: Evaluates and refines the response.
    • The system requires manual deployment and careful framing of system messages to effectively engage the reviewer model.

Mermaid Diagram – Stock Market Analysis Process:

flowchart TD
    A[User Stock Query] --> B[Generator Model Produces a Response]
    B --> C[Reviewer Model Evaluates the Response]
    C --> D[Final Analyzed Output]

6. Integrating External Data and AI Coding

The integration of an external data repository into an AI project was discussed, highlighting several challenges:

  • Challenges Encountered:
    • Handling legacy code bases.
    • Coordinating between multiple code bases.
  • Discussion Points:
    • The potential for AI coding systems in incremental development scenarios.
    • The importance of human oversight for complex issues.
    • Using a fast HTML library for Python-only web app development.
  • Additional Presentations:
    • Eli: Demonstrated a blog assistant tool using a language model for summarizing and answering questions about blog posts.
    • Zhimin: Showcased a markdown-format project with a chatbot, albeit with slow performance.
    • Discussion on integrating web pages with AI coding systems and the utilization of Microsoft’s Co-Pilot for enhanced knowledge integration.

7. Exploring the Chinese Party Game with AI

Zhimin introduced a popular Chinese party game featuring four AI players using different models.

  • Game Mechanics:
    • Roles:
      • One undercover player receives a secret word.
      • The other three players receive civilian words.
    • Objective:
      • Identify the undercover player based on subtle clues.
    • The game proceeds in rounds until the undercover player is identified or eliminated.
  • Rule Clarification:
    • The game is played exclusively with AI players.

Mermaid Diagram – Chinese Party Game Flow:

flowchart TD
    A[Game Start] --> B[Assign Roles to 4 AI Players]
    B --> C[Distribute Civilian or Secret Words]
    C --> D[Players Provide Synonym Clues]
    D --> E{Is Undercover Identified?}
    E -- Yes --> F[End Game]
    E -- No  --> G[Proceed to Next Round]

8. Chat Assistants for Post-Graduation Support

Greg showcased a versatile chat assistant aimed at supporting post-graduation needs including:

  • Job hunting and career growth
  • Financial aid and student loans

  • Tax information and life advice

  • Demonstrated Capabilities:
    • Processing user prompts and generating tailored responses.
    • Explaining complex topics (e.g., student loan payments, tax implications) often through humorous or unconventional scenarios.
  • Shreya’s Assistant:
    • Maintains conversation context by storing all prompts and responses.
    • Detects when queries deviate from the ongoing conversation, prompting the user to either continue or restart.
    • Generates a summary of the current discussion.

9. AI Code Tutor and Search System

Greg discussed recent developments on two fronts:

  • AI Code Tutor:
    • Live on Render, it compares user code queries against a set rubric and provides suggestions for improvement.
  • AI Code Search System (New Project):
    • Utilizes a RAG (Retrieval-Augmented Generation) approach to accurately locate code files, classes, functions, and methods based on natural language queries.
  • Additional Work:
    • Shreyas shared progress on calculating the cost of LM calls and plans to implement user authentication for cost tracking.

Mermaid Diagram – AI Code Search System Workflow:

flowchart TD
    A[User's Natural Language Query] --> B[RAG Model Retrieves Relevant Code]
    B --> C[Identify Code Files/Classes/Functions]
    C --> D[Display Search Results]

Conclusion

The meeting provided comprehensive insights into various AI projects and tools. Discussions ranged from chat assistants tailored to different user groups to complex systems for refining prompts and analyzing stock market data. The outlined next steps emphasize the continued refinement of coding skills, exploration of new libraries, and further development of AI systems that integrate natural language processing with incremental coding improvements.