Prompt-Based Gemini QA Assistant

This project demonstrates how to build a lightweight, prompt-driven question-answering assistant using Google’s Gemini API. It uses structured or semi-structured textual documents to extract answers using multiple prompt engineering techniques.

Project Goals

• Explore prompt engineering strategies (zero-shot, few-shot, chain-of-thought).
• Use LLMs (via Gemini API) to extract information from textual documents.
• Log results for history, evaluation, or general response comparison.

---

Features

• Load and read mock textual documents.
• Apply different prompting techniques dynamically.
• Save responses to a .csv file.
• Swap between mock output and real API calls.
• CLI-based loop to simulate multi-query usage.

---

Prompting Techniques Used

• Zero-Shot: Direct question → response with no context.
• Few-Shot: Injects 2–3 example QA pairs.
• Chain-of-Thought (CoT): Encourages intermediate reasoning steps.

---

Directory Structure

Prompt-Based-Gemini-QA-Assistant/ │ ├── sample_docs/ # Contains mock text document │ └── loan_agreement_summary.txt ├── scripts/ │ └── run_qa.py # Main entrypoint for QA assistant ├── src/ │ ├── load_text.py # Loads document from file │ ├── prompt_templates.py # Generates prompt strings │ ├── qa_engine.py # Gemini and mock generators │ └── evaluate.py # Result logging to CSV ├── qa_results.csv # Stores prompt-type → answer logs ├── requirements.txt └── README.md

---

Getting Started

1. Install dependencies

pip install -r requirements.txt

2. Configure .env

Create a .env file in the root directory:

GOOGLE_API_KEY=your_api_key_here

3. Run the assistant

python -m scripts.run_qa

You’ll be prompted to ask analytical questions like:

• "What is the loan amount?"

• "When is the first payment due?"

Type "Goodbye!" to exit.

Example Output

[ZERO-SHOT] What is the interest rate?
1.2% annually, fixed.

[FEW-SHOT] What is the interest rate?
The interest rate is 1.2% annually, fixed.

[CHAIN OF THOUGHT] What is the interest rate?
I need to find the section that discusses the interest rate...

---

Skills Demonstrated

• Prompt Engineering

  • Designed and tested zero-shot, few-shot, and chain-of-thought (CoT) prompt formats to evaluate Gemini's performance on document-based QA tasks.

• Google Gemini API

  • Integrated Gemini's LLM via API calls to generate responses based on document content and prompt instructions.

• Large Language Models (LLM)

  • Applied LLM reasoning techniques without fine-tuning, focusing on structured input-output handling and comparative prompt analysis.

• Natural Language Processing (NLP)

  • Processed unstructured text inputs and extracted relevant information using LLM-driven NLP workflows.

• Machine Learning (ML) Foundations

  • While no training or model fitting was performed, the project demonstrates core ML skills including evaluation mindset, reproducibility, and modular experimentation.

• Python Programming

  • Implemented a clean, modular Python codebase for managing prompt formats, user inputs, and LLM API interactions.

---

Future Directions

• Add Streamlit/Gradio UI
• Support PDF chunking with LangChain
• Score prompts using GPTCritic or Gemini RAG evaluation
• Deploy with usage limits

---

Author

Owen Wienczkowski — ML/NLP Engineer-in-training, focused on LLM and NLP applications GitHub: @owenwienczkowski