SigMap is a zero-dependency code retrieval tool from developer Manoj Mallick that uses TF-IDF instead of embeddings. It hits 81.1% accuracy finding the correct file in the top 5 results, compared to a 13.6% random baseline. Token usage drops by an average of 96.9% across 18 real repositories, according to benchmarks covering 90 coding tasks. This efficiency is on par with a new Claude skill, Caveman, which cuts tokens by 75%.
TF-IDF is decades old. Pure term frequency math, no neural networks involved. It works because code has structure that keyword matching captures well, similar to how structured folders act as a knowledge graph. SigMap skips the vector database entirely. It extracts function and class signatures from 29 programming languages, scores them against natural language queries, and returns deterministic results. The same query always returns the same ranking. No data drift from stale embeddings, and everything runs locally.
Task success rates climbed from 10% to 52.2% with proper context delivery. Average prompts per task fell from 2.84 to 1.68. GPT-4o context overflow scenarios, which occurred in 13 out of 18 test repositories without context management, dropped to zero.
SigMap integrates with GitHub Copilot, Claude, Cursor, Windsurf, OpenAI, and Gemini by writing adapter files to standard locations. It's available as an npm package with no dependencies, standalone binaries, and IDE extensions for VS Code, JetBrains, and Neovim. Setup is one command: npx sigmap.