Skip to content
CompletedAugust 2025 - December 2025

SimRAG Reproduction Study

Reproduction study of a paper implementing similarity-based RAG with two-stage fine-tuning on consumer hardware.

PythonRAGQdrantSentence TransformersOllamaPurdue GenAI APIPyTorchDockerPoetry
GitHubRead Paper

Challenge

Reproduce and understand the SimRAG paper's similarity-based RAG techniques, implementing on consumer hardware to learn RAG fundamentals.

Solution

•

Built provider-agnostic interface supporting local (Ollama) and cloud LLMs

•

Implemented two-stage fine-tuning: instruction following, then domain adaptation

•

Created test suite with mocked dependencies for reproducible testing

Result

Successfully trained and tested on consumer hardware (RTX 3080). Fine-tuning showed limited improvement because the 1.5B model I used was smaller than the paper's 8B/27B models, highlighting the importance of model capacity in RAG systems.