Key facts
- Augment Code utilizes semantic retrieval with embeddings, retrieval models, and vector databases for its AI coding tools.
- This approach is claimed to be more effective for large, private codebases compared to grep-based methods.
- Augment Code reports a 33% improvement in token efficiency over Claude Code in specific benchmarks.
- The company emphasizes the importance of the entire context engine, not just the retrieval system, for quality outcomes.
The development of AI coding tools is increasingly focused on the software that manages AI models, rather than solely on the models themselves. Anthropic's Claude Code team advocates for a 'lean harness' approach, believing that rapid model improvements make it impractical to build overly opinionated features. They prioritize a flexible system that allows developers to add their own tools.
In contrast, Augment Code has developed a context engine that pre-indexes code repositories using embeddings, retrieval models, and a vector database to retrieve conceptually relevant code. Vinay Perneti, Augment Code's VP of Engineering, argues that this semantic retrieval approach offers significant advantages, particularly for large, private codebases where AI models have not previously encountered the code. He contrasts this with grep-based methods used by other agents.
Perneti claims Augment Code's method leads to better token efficiency, citing a benchmark where their tool was 33% more efficient than Claude Code while achieving similar accuracy. He attributes potential discrepancies in benchmark results to variations in retrieval systems and the overall context engine's design. Augment Code has dedicated substantial research to optimizing retrieval and embedding models for large codebases since its founding in 2022.