AI Agents for Code Generation

Project Overview
The project uses a ReAct-style agent architecture to connect different AI components, each specialized in specific tasks:

Documentation Retrieval:
A QueryEngineTool reads API documentation from local files and other data sources.
Code Reading:
A custom tool reads existing code files to provide context and support code generation.
Code Generation:
Two LLMs are integrated:
Mistral: Serves as the primary language model for general query understanding.
CodeLlama: Specializes in generating valid code in response to user prompts.
The overall flow is as follows:

Data Ingestion:
The system loads documents (including PDFs) from a data directory using a SimpleDirectoryReader and processes them with a parser (LlamaParse). It then builds a vector index using an embedding model.
Tool Setup:
Two main tools are integrated:
API Documentation Tool: Uses the query engine to provide context from the loaded documents.
Code Reader Tool: Reads and returns the contents of code files.
Agent Interaction:
A ReActAgent is created using the code generation LLM (CodeLlama) along with the tools. The agent accepts user prompts, generates code along with a description, and outputs a filename suggestion.
Post-Processing:
A QueryPipeline with a PydanticOutputParser processes the generated response into structured JSON (fields include code, description, and filename) which is then printed and saved as a file.
Technologies Used
Python 3.x
The main programming language.
Flask
Demonstrated via a sample CRUD API (see test.py) to showcase integration with generated code.
llama_index and LlamaParse
For document reading, vector indexing, and natural language query processing.
ReActAgent (from llama_index)
To orchestrate the interactions between tools and LLMs.
Ollama LLM Interface:
Mistral: General language understanding and query processing.
CodeLlama: Specialized in code generation tasks.
Pydantic
For defining and validating structured output (e.g., code, description, filename).
dotenv
To manage environment variables.
Requests
Used in example scripts (e.g., create_item_script.py) for making HTTP requests.
Additional Dependencies:
See requirements.txt for a complete list, which includes libraries for asynchronous processing, embedding resolution, and more.
Models Used
Mistral:
Utilized via the Ollama API for handling general language processing.
CodeLlama:
Employed via the Ollama API for code generation, ensuring the produced code is syntactically valid.
BAAI/bge-m3:
A local embedding model (resolved as local:BAAI/bge-m3) used for document vectorization and indexing.
Getting Started
Prerequisites
Python 3.x installed on your system.

Flask (for running the provided CRUD API sample).

Install required Python packages:

pip install -r requirements.txt