SigMap — the deterministic, verifiable grounding layer for AI code work | SigMap

Skip to content

Appearance

SigMapGrounded context AI can trust. Deterministic. Verifiable.<br>The deterministic, verifiable grounding layer for AI code work. Proof — 86.7% hit@5 · 97.0% token reduction · zero deps, fully offline.<br>Get Started →<br>Benchmark Report →<br>GitHub

💬<br>Fewer prompts to finish the task<br>Latest saved run: 2.84 prompts without SigMap vs 1.46 with SigMap. That is a 48.8% reduction across 90 real coding tasks.<br>Task benchmark →

🎯<br>Right file in context<br>86.7% hit@5 across 21 repos and 90 tasks. Random selection finds the right file only 13.6% of the time.<br>Retrieval benchmark →

⚖️<br>Trust the answer, not just the token count<br>Use ask to build focused context, validate to check coverage, judge to score groundedness, and learn to reinforce the files that helped.<br>Workflow docs →

🌐<br>33 languages, zero native deps<br>TypeScript, Python, Go, Rust, Java, Kotlin, Ruby, PHP, Swift, C#, C++, Dart, Scala, Vue, Svelte, GraphQL, SQL, Terraform, R, GDScript, and more.<br>Language support →

🔌<br>MCP-ready and IDE-friendly<br>Works with Copilot, Claude Code, Cursor, Windsurf, Codex, OpenCode, and Gemini CLI. Use MCP for dynamic query_context lookups on demand.<br>MCP setup →

📈<br>One report for the full story<br>Run the benchmark matrix once and open a self-contained HTML dashboard with token, retrieval, quality, and task metrics together.<br>Benchmark overview →

Release: v8.3.0·New — Python site-packages grounding (the moat, both ecosystems): installed-library grounding now covers Python as well as JS/TS. verify-ai-output and the verify_suggestion MCP tool check AI-suggested Python code against the packages actually installed in the project's venv — reading each dependency's __init__.py/.pyi exports + pinned version — so real library calls stop false-flagging. No Python runtime, zero-dependency, deterministic.<br>Benchmark: sigmap-v8.3-main·87% hit@5 · 97.0% token reduction · 2026-07-04

Who is this for? ​<br>I am…Go toNew to SigMapQuick startUsing it dailyask · validate · judgeSetting up a team / CIConfig · StrategiesUsing open-source agents (OpenCode, Aider, Cline)Open-source agents guideRunning local LLMs (Ollama, llama.cpp, vLLM)Local LLMs guide — zero cost, full privacyIntegrating with MCP, Claude, or CursorMCP setupEvaluating for a monorepoStrategies · GeneralizationComparing against embeddings or RAGCompare alternatives<br>30-second start ​<br>Step 1: Generate context for your project<br>bashnpx sigmap<br>Step 2: Ask for relevant files (query-specific context)<br>bashsigmap ask "explain the auth flow"<br># Outputs: ranked file list + .context/query-context.md (ready to paste)<br>Step 3: Copy context to your AI assistant<br>Open .context/query-context.md<br>Paste the content into Claude, Copilot, ChatGPT, or your IDE's AI chat<br>Ask: "Explain the auth flow"<br>Step 4: Save the AI response<br>bash# Copy the AI's answer into a file<br>echo "Paste AI response here..." > response.txt<br>Step 5: Validate coverage (optional)<br>bashsigmap validate --query "auth login token"<br># Check if coverage is high enough to trust the response<br>Step 6: Judge groundedness<br>bashsigmap judge --response response.txt --context .context/query-context.md<br># Score: shows if the answer is grounded in your code<br>That flow gives you: a compact signature map · a focused query context · a coverage sanity check · a groundedness score for the answer.

The workflow ​<br>SigMap is no longer just "shrink the context file." Every step has a purpose:<br>Generate a compact signature map once<br>Ask for the files that matter to the current task<br>Validate whether coverage is high enough to trust the context<br>Judge whether an answer is grounded in the supplied code<br>Learn from good and bad results locally, inside the repo<br>See the full end-to-end walkthrough to watch this in action on a real repo.

Latest saved benchmark snapshot ​<br>MetricWithout SigMapWith SigMapTask success proxy10%67.8% Prompts per task2.841.46 Retrieval hit@513.6%87% (87% graph-boosted)Overall token reduction—97.0% GPT-4o overflow repos16/210/21 Latest saved benchmark run: 2026-07-04 (v8.3.0) .

Benchmark proof, by question ​<br>If you want to prove...OpenSigMap reduces token load dramaticallyToken benchmarkSigMap finds the right file more oftenRetrieval benchmarkSigMap reduces retries and wrong-context answersTask benchmarkSigMap keeps large repos inside model limitsQuality benchmark<br>Where to go next ​<br>New to the product: Quick start<br>Want the core daily flow: ask, validate, judge, learning<br>Using Claude Code or Cursor: MCP setup<br>Evaluating the launch claims: Benchmark overview<br>🌍 See the community: where SigMap's stargazers are around the world