Jessica Dong
Exploring Retrieval-Augmented Code Generation for Procedural Modelling in Houdini
Summary
This project explores retrieval-augmented code generation for procedural modelling in Houdini. The prototype connects Houdini to an external FastAPI-based RAG service, a manually curated procedural knowledge base, and a locally hosted Qwen2.5-Coder-7B model served through llama.cpp.
The project investigates how retrieved Houdini-specific knowledge can support generation of executable Python and SOP node networks, particularly when the model lacks specialised API or procedural conventions.The results show that retrieval can support reuse of verified procedural patterns, but that successful generation still depends on the model’s ability to interpret and compose the retrieved knowledge. The project also compares selected local and cloud-model runs and examines issues including retrieval quality, chunking, context size, execution correctness, and geometric correctness.
GitHub: https://github.com/adalek/MasterProject