Week 3: CodeGenBot – Retrieval-Augmented Code Generation Assistant

This week’s project extends the Cellula AI internship into code intelligence by building CodeGenBot: a retrieval-augmented Python code generation assistant. The system leverages semantic search over the HumanEval dataset and a state-of-the-art LLM to generate Python code from user prompts, grounded in real coding examples. The app is delivered as a modular, production-ready Streamlit chatbot.

• Retrieval-Augmented Generation (RAG): Combines semantic search with LLM code synthesis for accurate, context-aware code generation.
• Semantic Search: Uses Sentence Transformers to embed and retrieve similar coding problems from HumanEval.
• LLM Integration: Utilizes DeepSeek-R1-Distill-Qwen-1.5B via HuggingFace Inference API for high-quality Python code generation.
• Conversational UI: Streamlit-based chat interface with code formatting, chat history, and error handling.
• Modular Pipeline: Clean separation of embedding, retrieval, and generation logic for easy extension and maintenance.

• Dataset: HumanEval – real-world Python coding problems and solutions.
• Embedding: Prompts are embedded using all-MiniLM-L6-v2 (Sentence Transformers) for semantic similarity search.
• Retrieval: Vector search retrieves the most relevant coding examples to provide context for the LLM.
• Code Generation: The LLM receives both the user prompt and retrieved context, improving code relevance and correctness.
• UI: Streamlit app manages user interaction, chat state, and code display.

1. Clone the repository:

   git clone https://github.com/NightPrinceY/Cellula-NLP-Engineer-Intern.git
   cd Cellula-NLP-Engineer-Intern/Week3

2. Install dependencies:

   pip install -r requirements.txt

3. Download HumanEval dataset:

   • The required parquet file is included in data/.

4. Set up HuggingFace API key:

   • Export your HuggingFace API key as an environment variable or update codegen.py as needed.

Run the Streamlit app:

streamlit run app.py

• Interact with CodeGenBot in your browser.
• Enter a Python coding prompt (e.g., "Write a function that returns the factorial of a number").
• The bot will retrieve similar problems, generate code, and display the result in a chat interface.

Week3/
├── app.py                # Streamlit UI
├── pipeline.py           # Main pipeline logic
├── codegen.py            # LLM code generation wrapper
├── embedding.py          # Embedding utility
├── retrieval.py          # Vector search and retrieval
├── data/
│   └── test-00000-of-00001.parquet  # HumanEval dataset
├── requirements.txt      # Dependencies
├── CodeGen.ipynb         # Experiments and prototyping

• app.py: Streamlit app for user interaction and chat UI.
• pipeline.py: Orchestrates embedding, retrieval, and code generation.
• codegen.py: Handles LLM API calls and prompt construction.
• embedding.py: Provides the Embedder class for semantic encoding.
• retrieval.py: Implements vector search and context retrieval.
• data/: Contains the HumanEval dataset in parquet format.
• CodeGen.ipynb: Jupyter notebook for prototyping and experiments.
• requirements.txt: Python dependencies.

from pipeline import CodeGenPipeline
pipeline = CodeGenPipeline("hf://datasets/openai/openai_humaneval/openai_humaneval/test-00000-of-00001.parquet")
result = pipeline.generate_code_from_prompt("Write a function that returns the factorial of a number")
print(result)

• Add support for multi-language code generation
• Integrate more datasets and problem types
• Enhance retrieval ranking and context selection
• Deploy as a web service or IDE extension

Prepared by: Yahya Alnwsany
Cellula AI Intern – Week 3
Portfolio | Week 3 Repo