Agent Code Generator

This project sets up a ReAct‑style agent that can:

• Read user prompts
• Use a vector index over your data/ folder (including PDFs) via QueryEngineTool
• Read raw code files from data/ via code_reader
• Generate new code with a dedicated LLM (codellama)
• Parse the LLM’s output into JSON (with code, description, filename)
• Save the generated code into output/

Architecture

flowchart LR
  subgraph User
    A[Enter Prompt]
  end

  subgraph Agent
    direction TB
    B[ReActAgent] 
    B -->|calls| C(code_reader)
    B -->|calls| D(QueryEngineTool)
    C --> E[reads data/<file>]
    D --> F[VectorStoreIndex over data/]
    F --> G["LLM (llama3.2) for docs"]
    B --> H["LLM (codellama) for code"]
  end

  subgraph Post-Processing
    H --> I[PydanticOutputParser]
    I --> J["JSON {code, description, filename}"]
    J --> K[write to output/]
  end

  A --> B

Loading

Setup

# create and activate a venv
python3 -m venv .venv
source .venv/bin/activate    # on Windows use: .venv\Scripts\activate

# install dependencies
pip install -r requirements.txt

Usage

1. Populate data/ with code files or PDFs.
2. Ensure your .env has LLAMA_CLOUD_API_KEY.
3. Run python main.py, enter prompts when prompted.
4. Generated code appears in output/ with the returned filename.

Prompt Suggestions

Copy‑paste the prompt below into the running agent to generate a client script:

Read the contents of data/test.py and write a Python script that calls the POST /items endpoint to create a new item.