sigmap

# SigMap

> The deterministic, verifiable grounding layer for AI code work.
> A reproducible signature-and-evidence map that agents, CI, and reviewers can trust and audit. No embeddings, no vector DB, fully offline.

SigMap is the deterministic, verifiable grounding layer for AI code work. It
extracts function and class signatures from a codebase and builds a byte-stable
signature-and-evidence map that agents, CI, and reviewers can trust and audit —
proving which files and symbols are real before acting. Deterministic TF-IDF
ranking keeps the relevant context in scope (cutting tokens ~97% as a side
effect), with no LLM calls, embeddings, or vector database. Works with Claude,
Cursor, GitHub Copilot, Aider, Windsurf, local LLMs, and MCP.

# Version: 8.51.6 | Benchmark: sigmap-v8.51-main (2026-09-28)
# Source: auto-generated from package.json, version.json, benchmarks/latest.json, src/mcp/tools.js, src/config/defaults.js
# Regenerate: npm run generate:llms   |   Validate: npm run validate:llms

## What SigMap solves

- AI agents hallucinate functions, files, and imports that don't exist — `verify-ai-output` flags fabricated references before you trust them.
- Agents load 60–80K tokens of raw source just to orient — `ask` ranks the codebase and sends ~2K tokens of the relevant signatures.
- Cold start every session — `note` + the `read_memory` MCP tool carry decisions and focus across sessions.
- No blast-radius awareness before editing a hub file — `--impact` shows every file a change touches.
- Pasted stack traces, CI logs, and JSON bloat the prompt — `squeeze` minimizes them and enriches the top frame from the symbol index.

## Core metrics (benchmark: sigmap-v8.51-main, 2026-09-28)

- hit@5 retrieval: 78.6% (retrieval corpus)
- honest vs grep: 86.4% vs 40.8% single-shot grep (2.12× lift, 125 tasks / 19 repos)
- Token reduction: 96.1% average across benchmark repos
- Task-success proxy: 61.0% (modeled from retrieval tiers, not measured LLM sessions)
- Prompts per task: 1.61 vs 2.84 baseline (43.4% fewer, modeled)
- Languages: 36 supported · MCP tools: 22
- Dependencies: zero npm runtime dependencies · fully offline

## Quick start
```bash
npx sigmap                          # generate compact signature context for the repo
npx sigmap ask "<your query>"       # rank the files relevant to a task
npx sigmap verify-ai-output ans.md  # flag fabricated files/imports/symbols in an AI answer
npx sigmap --mcp                    # start the MCP server over stdio
```

## Docs
- [Full documentation](https://sigmap.io/)
- [CLI reference](https://sigmap.io/guide/cli)
- [MCP tools](https://sigmap.io/guide/mcp)
- [Benchmark methodology](https://sigmap.io/guide/benchmark)
- [Configuration guide](https://sigmap.io/guide/config)
- [Changelog](https://github.com/manojmallick/sigmap/blob/main/CHANGELOG.md)

## Optional
- [GitHub repository](https://github.com/manojmallick/sigmap)
- [npm package](https://www.npmjs.com/package/sigmap)
- [Benchmark dataset (Zenodo)](https://doi.org/10.5281/zenodo.19898842)
- [Full LLM reference](https://sigmap.io/llms-full.txt)

SigMap — the deterministic, verifiable grounding layer for AI code work. The
reproducible signature-and-evidence map agents, CI, and reviewers can audit,
which agentic grep cannot produce.