Maintainer triage is often a game of memory, but a new open-source project called "Already Filed" is replacing that mental load with local LLM inference. Developed by Vanshaj Poonia for the Hacktoberfest Open-Source AI Challenge, the tool uses Google's EmbeddingGemma model to scan GitHub repositories and identify duplicate issues directly in the browser. The project addresses a critical pain point in large open-source ecosystems: the inability to recall thousands of past tickets. By automating the detection of repeated bugs, it allows maintainers to issue the infamous "Duplicate of #..." command with data-backed confidence rather than guesswork.

The Technical Recipe for Accuracy

The core innovation here isn't just using an embedding model, but how it is tuned. Poonia's initial implementation using EmbeddingGemma achieved only a 42% top-three accuracy rate, barely outperforming standard keyword search. The breakthrough came from three specific optimizations. First, the search window was restricted to issues that were open or closed within the last 30 days, recognizing that old closed issues are rarely the target of new duplicates. Second, the developer switched from a generic "sentence similarity" prompt to a specific retrieval prompt, which boosted top-three accuracy to 61%. Finally, a weighted mix of body, title, and keyword signals pushed performance to 65%.

Benchmarking Against Keyword Search

When tested against the massive zed-industries/zed repository, which contains over 23,000 issues and 1,653 linked duplicates, the tool demonstrated significant gains over traditional methods. While GitHub's native BM25 keyword search achieved a 38% top-three accuracy, the optimized EmbeddingGemma recipe hit 65%. The tool correctly identified the original issue in the top three suggestions for 1,074 of the 1,653 known duplicates in the Zed repo. This suggests that for repos with complex, template-heavy issue structures, semantic understanding is vastly superior to exact string matching.

Privacy and Local Inference

A standout feature of Already Filed is its commitment to privacy and decentralization. Unlike many AI-powered GitHub bots that send issue text to external APIs, this tool runs entirely in the user's browser tab. It utilizes transformers.js to run the 197 MB EmbeddingGemma model via WebGPU or WASM. This means private repository issues never leave the local machine, a crucial consideration for companies managing sensitive codebases. The model downloads once, and subsequent operations are local, eliminating API costs and data residency concerns.

Key Takeaways

  • EmbeddingGemma, when properly prompted and scoped, outperforms BM25 keyword search by 27 percentage points on duplicate detection.
  • Restricting search scope to recent issues significantly reduces false positives and improves relevance.
  • The tool runs locally in the browser, ensuring data privacy for both public and private repositories.
  • A smaller model (bge-small) achieved comparable results to EmbeddingGemma, suggesting the prompt and search strategy matter more than model size.

The Bottom Line

This project proves that the biggest wins in applied LLMs often come from rigorous data engineering and prompt optimization, not just model scale. It is a masterclass in building practical, privacy-first AI tools for developer workflows.