Gemini Query Fanout (Streamlit)

A Streamlit GUI that replicates the logic from Query_fanout.ipynb:

• Generates 10 semantically related questions using Gemini generate API
• Uses Gemini embeddings to compute cosine similarities and rank questions
• Allows downloading the top results as a JSON file

Features

🤖 Model Selection

• Choose from available Gemini models (default: gemini-3-flash-preview)
• Refresh button (🔄) to fetch latest models from Gemini API
• Supports models: gemini-3-flash-preview, gemini-2.0-flash-exp, gemini-1.5-flash, gemini-1.5-pro, and more

🔑 API Key Management

• Enter your Gemini API key directly in the app
• API key is not saved between sessions for security
• The key is only sent to Google's Gemini API endpoints

⚙️ Generation Configuration

Customize the AI generation parameters:

• Temperature (0-1): Controls randomness - lower is more focused, higher is more creative (default: 0.7)
• Top K (1-100): Limits sampling to top K tokens (default: 40)
• Top P (0-1): Nucleus sampling probability threshold (default: 0.95)
• Max Output Tokens (1-8196): Maximum length of generated response (default: 2048)

🌍 Language Support

• Specify any language for keyword analysis and output
• Default: Polish

Quick start (using uv)

1. Install uv (see https://docs.astral.sh/uv/)
2. In the project folder:

uv add streamlit requests numpy
uv run streamlit run app.py

Quick start (pip)

python -m venv .venv
source .venv/bin/activate  # or `.venv\\Scripts\\activate` on Windows
pip install -r requirements.txt
streamlit run app.py

Usage

1. Enter a keyword to analyze
2. Select language (default: Polish)
3. Enter your Gemini API key
4. Choose AI model (refresh to get latest models)
5. Adjust Top N results slider (1-10 questions)
6. Optional: Expand Generation Config to customize AI parameters
7. Click Generate to create query fanout
8. Download results as JSON

Security Notes

• API key is not stored locally and must be entered each session
• The API key is sent to Google's Gemini API endpoints for generation and embedding requests
• Keep your API key secure and do not share it

Notes

• If you want to use a project environment manager, uv is supported and recommended for faster installs.
• Model list is cached for 5 minutes to reduce API calls
• Embeddings and query results are cached to improve performance