Ollama Chatbot Implementation

[Saran-Nag]

Chatbot Integration Setup & Overview

Setup Instructions

To run the chatbot locally, complete the following steps:

1. Install Docker

   • Ensure Docker is installed and running
   • Set up an environment for the general SPAC application

2. Install Ollama

   • Install Ollama on your system

   • Pull the required models:

     • gemma3:latest
     • nomic-embed-text:latest

3. Run the Application

   • Start the SPAC application as usual
   • The chatbot should initialize alongside the application

---

Chatbot Overview

The chatbot is designed to assist users by answering questions related to:

• The application background
• The user-loaded dataset
• The current tab/context within the application

Current Capabilities

• Can answer dataset-related queries

  • Example: “How much of X cell is in the dataset?”

• Has partial awareness of the user’s current tab

• Uses RAG (Retrieval-Augmented Generation) context provided by George Zaki:

  • Includes:

    • Plot information
    • Application background
    • Basic biological knowledge

Known Limitations

• Tab Tracking Issues

  • The chatbot frequently hallucinates the current tab
  • Root cause is unclear (could be model limitations or implementation issues)

• Plot Generation Inconsistency

  • Supports generating plots for:

    • Boxplot tab
    • Anno v Anno tab

  • However:

    • Parameter handling is inconsistent
    • Output reliability is low
    • Unclear if issues stem from prompt design, model behavior, or code

• Field Update Reliability

  • Chatbot occasionally fails to update fields properly
  • In some cases, no changes are made despite valid prompts

---

Future Work

Planned improvements include:

• Expand plot generation support to additional tabs

• Improve consistency and reliability of plot generation

• Reduce hallucinations, especially in:

  • Tab tracking
  • User context awareness

• Improve robustness of chatbot-driven field updates

• Investigate whether current issues are:

  • Model-related
  • Prompt-related
  • Implementation-related

---

Notes

• The system is functional but still experimental in several areas
• Further debugging and iteration are required to stabilize behavior

[georgezakinih]

@Saran-Nag , let's make the PR when the full implementation is in place.