Vecna / AIC51 - Multimodal Video Search Engine

Multimodal video keyframe search engine built for the AI Challenge (AIC). The system integrates CLIP (PE-Core-L-14-336), SigLIP (ViT-SO400M-14-SigLIP-384), Groq LLM Query Expander, and Milvus Vector Database.

Contributors

• Lê Tuấn Kiệt (@ProfK602170)
• Nguyễn Lê Minh Hoàng (@nlmhoagn)
• Cao Chí Minh (@JimmyK300)
• Ngô Đắc Minh (@kinus-is-coding)
• Phan Hiếu Minh (@PhanHieuMinh)

Prerequisites

• Python 3.11+
• Docker Desktop (Running for Milvus Vector Database)
• Node.js 18+ and npm

Setup & Running the Server

Open PowerShell in the project root directory:

# 1. Activate Virtual Environment
.\.venv\Scripts\activate

# 2. Start AIC51 Server (automatically launches Milvus Docker containers and Web UI)
aic51-cli --dev serve

Access the web interface at: http://localhost:6900

API Key Configuration (Groq LLM)

Query expansion and spelling correction require a Groq Cloud API Key.

Method 1: Environment Variable (Recommended)

Set the environment variable in PowerShell prior to starting the server:

$env:GROQ_API_KEY="YOUR_GROQ_API_KEY_HERE"

Method 2: Configuration File (config.yaml)

Open config.yaml in the project root directory and update the llm section:

llm:
  enable: true
  provider: "groq"
  model_name: "openai/gpt-oss-120b"
  api_key: "YOUR_GROQ_API_KEY_HERE"

Security Note: Never commit raw API keys to Git. GitHub Push Protection automatically blocks commits containing active secret keys. Use environment variables or keep api_key: "" when pushing to remote repositories.