The author developed a local development tool named TeaRAGs to address inefficiencies in agentic coding workflows. The project was initiated after existing solutions failed to handle a Ruby-based monorepo exceeding 3.5 million lines of code. The tool is designed to keep all code, indices, embeddings, and call graphs local, ensuring data privacy.
Current agentic tools often struggle with large codebases, leading to issues like long indexing times, poor support for dynamic languages like Ruby, and context degradation. Research cited, such as the 2025 Chroma study on context rot and the 2024 TACL paper on the 'Lost in the Middle' phenomenon, highlights how irrelevant context fragments degrade model performance.
The author's internal testing demonstrated significant performance gains with TeaRAGs. When tasked with resolving a bug based solely on a ticket description, the agent using TeaRAGs completed the search and report generation in 40 seconds, compared to over 10 minutes for an agent relying on standard grep-based methods.
Industry trends show a shift in how codebases are indexed. While some providers like Sourcegraph have moved away from embeddings due to privacy concerns and costs, others like Cursor continue to index code on their own servers. The author emphasizes that while semantic search is not a universal fix, it remains essential for effective discovery in large-scale repositories.
