Spotlight: Qodo Innovates Efficient Code Search with NVIDIA DGX

Overview

Qodo, a multi-agent code integrity platform, built its AI agents on retrieval-augmented generation powered by a state-of-the-art code embedding model trained on NVIDIA DGX. Qodo fine-tuned two embedding models, Qodo-Embed-1-1.5B and Qodo-Embed-1-7B (based on Qwen), achieving state-of-the-art accuracy on the Hugging Face MTEB CoIR leaderboard in their size categories. In a collaboration with NVIDIA, Qodo's code indexer, RAG retriever, and embedding model were substituted into NVIDIA's internal RAG solution (Genie) for searching private code repositories, integrated into NVIDIA's internal Slack system, yielding more detailed and accurate responses to technical questions from expert C++ developers than the original pipeline, evaluated using Ragas-generated synthetic questions against RTXDI, RTXGI and RTXPT SDK repositories.

The challenge

Large language models help developers write code faster, but face limitations understanding the nuances of programming languages, complex dependencies, and codebase-specific context, which can lead to lower-quality code and bottlenecks; general-purpose embedding models focus on language patterns rather than code-specific elements like syntax, variable dependencies, control flow and API usage, leading to imprecise code retrieval.

The solution

Qodo built its AI agents on retrieval-augmented generation (RAG), a code-specific indexing pipeline with language-specific static analysis for chunking, and a state-of-the-art code embedding model trained on an NVIDIA DGX 8x A100 80GB node using bfloat16 precision and large micro-batch sizes. Qodo fine-tuned two embedding models, Qodo-Embed-1-1.5B and Qodo-Embed-1-7B, based on the Qwen open-source LLM. In collaboration with NVIDIA, Qodo's code indexer, RAG retriever and embedding model were substituted into NVIDIA's internal Genie RAG pipeline for searching private code repositories, integrated into NVIDIA's internal Slack system.

Reported business value

Qodo's fine-tuned embedding models achieved state-of-the-art accuracy, leading the Hugging Face MTEB CoIR leaderboard in their respective size categories. In a case study evaluated with Ragas-generated synthetic questions against the RTXDI, RTXGI and RTXPT SDK repositories, the Qodo-based pipeline yielded more detailed and accurate responses to technical questions from expert C++ developers than NVIDIA's original Genie pipeline.