The deterministic, verifiable grounding layer for AI code work. Proof — 78.6% hit@5 · 96.6% token reduction · zero deps, fully offline.
CLAUDE.md; --report --json promised a CI exit-1 gate that never fired; learn decayed weights toward 0 instead of neutral; --impact printed paths that climbed out of the repo on every macOS checkout; compare crashed after a minute outside this repo; validate reported 218% coverage. Plus two drift gates that derive the CLI's documented surface from the dispatch chain itself, so this class of rot fails CI instead of shipping. 35 languages, zero dependencies, offline, deterministic.| I am… | Go to |
|---|---|
| New to SigMap | Quick start |
| Using it daily | ask · validate · judge |
| Setting up a team / CI | Config · Strategies |
| Using open-source agents (OpenCode, Aider, Cline) | Open-source agents guide |
| Running local LLMs (Ollama, llama.cpp, vLLM) | Local LLMs guide — zero cost, full privacy |
| Integrating with MCP, Claude, or Cursor | MCP setup |
| Evaluating for a monorepo | Strategies · Generalization |
| Comparing against embeddings or RAG | Compare alternatives |
Step 1: Generate context for your project
npx sigmapStep 2: Ask for relevant files (query-specific context)
sigmap ask "explain the auth flow"
# Outputs: ranked file list + .context/query-context.md (ready to paste)Step 3: Copy context to your AI assistant
.context/query-context.mdStep 4: Save the AI response
# Copy the AI's answer into a file
echo "Paste AI response here..." > response.txtStep 5: Validate coverage (optional)
sigmap validate --query "auth login token"
# Check if coverage is high enough to trust the responseStep 6: Judge groundedness
sigmap judge --response response.txt --context .context/query-context.md
# Score: shows if the answer is grounded in your codeThat flow gives you: a compact signature map · a focused query context · a coverage sanity check · a groundedness score for the answer.
SigMap is no longer just "shrink the context file." Every step has a purpose:
See the full end-to-end walkthrough to watch this in action on a real repo.
| Metric | Without SigMap | With SigMap |
|---|---|---|
| Task success proxy | — (proxy, modeled from retrieval tiers) | 61.9% |
| Prompts per task | 2.84 | 1.6 |
| Retrieval hit@5 | 44.0% (single-shot grep) | 78.6% (1.73× honest lift) |
| Overall token reduction | — | 96.6% |
| GPT-4o overflow repos | 16/21 | 0/21 |
Latest saved benchmark run: 2026-09-15 (v8.49.2).
| If you want to prove... | Open |
|---|---|
| SigMap reduces token load dramatically | Token benchmark |
| SigMap finds the right file more often | Retrieval benchmark |
| SigMap reduces retries and wrong-context answers | Task benchmark |
| SigMap keeps large repos inside model limits | Quality benchmark |