/AInsights

Two tasks are used to enhance business processes and gain insights using LLMs and computer vision.

Primary LanguagePython

Initial Task Solutions

Task 1: Converting CSV File to Database

The solution involves converting the CSV file into a database and decomposing the question into subqueries (query decomposition) that can be understood by the system using LLM. The steps are as follows:

  1. Split the question into subqueries:

    • Decompose the main query into smaller, manageable subqueries.

  2. Loop over the subqueries and query the database file:

    • Iterate through each subquery and execute it against the database.

  3. Retrieve the results and prompt the LLM to present them in a presentable form:

    • Collect the results and use LLM to format and present them.

  4. Ensure the system tolerates query errors and makes multiple attempts if results are not found:

    • Implement error handling to manage failed queries and retry if necessary.

Tools used:

  • LangChain
  • GPT-3.5 Turbo or GPT-4 (both work fine, but the former is more cost-effective)

to run the code cd to the submissionFolder create your own .env file and add API_KEY=YOUR_SECRET_KEY

then write

py peopleDetection.py

Task 2: Counting People in Each Frame

I used YOLOv8x to count the number of people in each frame, excluding the cashier area.

Steps:

  1. Load the video frames:

    • Extract frames from the video.

  2. Apply YOLOv8x model:

    • Use the YOLOv8x model to detect and count the number of people in each frame.

  3. Exclude the cashier area:

    • Implement logic to exclude detections in the cashier area.

Potential Enhancements

Task 1

  • Reducing cost
  • Use similarity embeddings to find similar queries to the user's question
  • Use query structuring to extract the variables you want to query in general queries

Task 2

  • Track the objects
  • Use the summation of confidence between frames of the tracked objects
  • Use a fine-tuned model
  • Try different models