Code Translation with Large Language Models using Ollama

This project is a prototype system that uses Large Language Models (LLMs) via Ollama to automatically translate C++ code into Java. It is designed to be modular, scalable, and suitable for enterprise integration, acting as a proof-of-concept for future automation workflows.

🚀 Technologies Used

• Python
• Docker - Service Containerization
• FastAPI - Lightweight web framework for REST APIs
• RabbitMQ - Message queue for asynchronous task handling
• Ollama - Local LLM backend
• javac - Java compiler

📦 Getting Started

1. Set key configuration values in .env

LLM_MODEL=qwen2.5-coder:7b  
MAX_ALLOWED_TOKENS=32768  # model specific

2. Build all services

docker compose build

3. Pull your preferred model

docker exec -it ollama ollama pull qwen2.5-coder:7b

4. Start all services

docker compose up

5. Start translation by sending a C++ file via HTTP request

POST Request Parameters

Info	Key	Type	Value
main file	files	File	legacyCode.cpp
header file	files	File	date.h
header file	files	File	currency.h
...
(optional)	custom_prompt	Text	The previous output missed a static nested helper class called Config. Ensure it’s static and public.
(WIP) test file	files	File	test_legacyCode.cpp

using POSTMAN

using CURL

curl -X POST http://localhost:8000/translate/ \
-F "files=@path/to/legacyCode.cpp" \
-F "files=@path/to/date.h" \
-F "files=@path/to/currency.h" \ 
-F "custom_prompt=The previous output missed a static nested helper class called Config. Ensure it’s static and public."
-F "files=@path/to/test_legacyCode.cpp" \

🧱 System Architecture

🔧 Components

FastAPI Service (Frontend Interface)

• Exposes an API endpoint at /translate/
• Accepts file uploads (.cpp, .h)
• Stores files at /fastapi/uploads/
• Sends jobs to RabbitMQ

RabbitMQ (Message Broker)

• Buffers and routes translation tasks
• Decouples file upload from translation processing
• Holds queued translation jobs until a worker is ready
• Enables reliable and scalable task dispatching

Translation Worker (Core Logic)

• Listens to the queue for new jobs
• Handles the complete translation pipeline:
  • Preprocessing of C++ files
  • Prompt-based translation via LLM
  • Compilation with javac
  • Retry logic using error feedback from java compiler
  • Outputs saved in /output/

Ollama (LLM Backend)

• Hosts the local LLM model (e.g., qwen2.5-coder:7b)
• Receives structured prompts via POST /api/generate
• Returns translated Java code
• Easily replaceable with other local models

Model Loader (LLM Preloader)

• Waits for Ollama to become available
• Sends warm-up prompt to preload the specified LLM model into memory
• Ensures fast first response by loading model tensors in advance

Example POST Request Parameters

Info	Key	Type	Value
.cpp file	files	File	legacyCode.cpp
.h file	files	File	date.h
.h file	files	File	currency.h
...	...	...	...
(optional)	custom_prompt	Text	The previous output missed a static nested helper class called Config. Ensure it’s static and public.
(WIP) test_.cpp file	files	File	test_legacyCode.cpp

📄 Notes

• Ollama currently has a default context window of 2048 tokens. To mitigate this, a estimate_token_count method is used, roughly estimating the tokens needed for a given prompt (currently word count * 2.8).

• Filenames are converted to PascalCase to follow Java naming conventions. Adjust as needed for other target languages.

• Language-specific pattern hints in output/profiles are appended to prompts to improve translation accuracy. Adjust via C++ Hints Extraction as needed.

• Existing translated files in the output folder will be overwritten when the translation process starts.

🛠 Debugging & Useful Commands

• General Docker commands:

  docker compose up --build -d              # detached mode
  docker logs translation_worker --follow   # show logs of specific service
  docker exec -it ollama sh                 # access Ollama container shell
  ollama list                               # list available models
  docker-compose restart ollama             # restart ollama to regain vram

• Test the LLM directly:

  curl -s -X POST http://localhost:11434/api/generate \
    -H "Content-Type: application/json" \
    -d '{"model": "qwen2.5-coder:7b", "prompt": "What is 1 + 1?", "stream": false}'

📚 Planned Features

• (WIP) Test Worker: Auto-generate unit tests post-translation for automatic quantitative evaluation

• (WIP) PMD Feedback Loop: Use static analysis to improve retry logic

• Support for additional language pairs (e.g., Python ⇄ Java)

• DB integration (e.g., PostgreSQL)

• Enterprise fine-tuning of model

• Web UI for job monitoring and status tracking

📄 License

This project is part of a Bachelor Thesis in collaboration with Oesterreichische Kontrollbank AG (OeKB). Licensed under the MIT License.