Skip to content

ROADMAP.md

Latest commit

 

History

History
93 lines (63 loc) · 7.13 KB

ROADMAP.md

File metadata and controls

93 lines (63 loc) · 7.13 KB
Raw

Plan: From proof-of-concept → useful product

Strategic bet: stop competing as a generic "preprocessor" (commoditized — it's a thin wrapper over WebLLM + regex). Reposition around the one job with validated demand and a real client-side gap: redact sensitive data locally → send safely to any cloud LLM → restore the answer. Everything below serves that bet.

✅ Recently shipped

  • Local server / proxy / CLI (redactkit) — runs headroom-style as a local reverse proxy any app can point at, plus redact/restore/clean commands. Node entry at redactkit/node. (src/server/proxy.js, bin/redactkit.js)
  • Tiered detection pipeline — regex (Tier 1) → local NER (Tier 2) → LLM (Tier 3) behind a Detector/router abstraction, selected with --tier. Default is zero-dep regex so it runs on any hardware. (src/detect/)
  • Local NER tier — Xenova/bert-base-NER (Piiranha/GLiNER optional) via transformers.js, fully offline after first download; preferred over the heavier LLM tier.
  • Reversible local map store with TTL (src/server/store.js); SSE streaming restore in the proxy.
  • Rule-based compression — content-type-aware token reduction (JSON minify, HTML strip, whitespace/dedupe), safe & dependency-free; compress CLI/pipeline step + proxy --compress. (src/compress/)
  • Image sanitization — blur sensitive data in images: OCR text-PII (tesseract.js) + faces (transformers.js), blurred with sharp; blur CLI + proxy --blur-images. One-way (not reversible). (src/image/)
  • Multi-provider proxy — request-body redaction for OpenAI Chat Completions, Anthropic Messages, OpenAI Responses API, and Gemini, selectable via --format. Shields coding agents (Claude Code via ANTHROPIC_BASE_URL, Codex via OPENAI_BASE_URL). (src/server/proxy.js)
  • Browser extension (experimental) — MV3 MAIN-world fetch interceptor for claude.ai / chatgpt.com / gemini; Tier-1 redaction in the browser. (extension/)

Next up: evaluation harness (precision/recall on ai4privacy/presidio-research), compliance profiles (--profile hipaa|gdpr|pci), file/document redaction (PDF/DOCX), and a browser (canvas-based) image path.

Phase 0 — Fix the known issues (≈1 hour, do first)

Clean the slate before building anything new. All low-risk.

# Issue File Fix
1 chunk() infinite-loop when overlap ≥ size chunk.js:21 Clamp step: const step = Math.max(1, size - overlap)
2 Error classes defined but unused (dead code) errors.js, index.js, engine.js Wire ModelNotLoadedError/InferenceError into _ensureLoaded + engine, or delete the file. Recommend wiring in.
3 README Quick Start calls chunk() with clean options README.md:70 Change to clean(); fix the result comment
4 Broken badges (USERNAME placeholder, unpublished npm pkg) README.md:3 Point at real repo pras-ops/redactkit; remove/replace npm badge until published
5 Outdated browser table (WebGPU shipped everywhere Nov 2025) README.md:139 Update Firefox/Safari to ✅ — it's now a selling point
6 Unused getLogger imports clean.js, extract.js Remove
7 Missing edge-case test tests/unit/chunk.test.js Add overlap ≥ size test so #1 can't regress

Exit criteria: 67+ tests green, build clean, README truthful.

Phase 1 — The pivot: client-side PII redaction (the core value)

This is what turns it from "an afternoon's worth of regex" into a product. New module src/preprocess/redact.js with a reversible API:

const { redacted, map } = p.redact(text);   // local, instant
//  → send `redacted` to OpenAI/Claude/etc.
const answer = p.restore(response, map);     // placeholders → originals, locally

Detection in two tiers:

  • Tier 1 — rule-based, no model, instant: email, phone, SSN, credit card (Luhn-checked), IP, common API-key/secret patterns. This is the reliable ~80% and needs zero download — the real differentiator vs. the heavyweight LLM story.
  • Tier 2 — optional LLM-assisted: names, addresses, org names (the fuzzy entities regex can't catch). This is where the existing WebLLM path finally earns its 1 GB — NER suits a small local model far better than the full extraction it does today.

Reuse what exists: validation.js verification logic, clean/chunk become supporting utilities, the logger gives an audit trail (valuable for compliance users).

Exit criteria: redact→restore round-trips losslessly on a realistic sample; rule-based path works with no model loaded; demo shows it.

Phase 2 — Make it adoptable (turn it into something people can find and trust)

  1. Reposition the README around the PII workflow, not "preprocessing." Lead with the redact→send→restore story.
  2. Live demo — rewrite examples/basic-demo.html as redact → (mock API call) → restore, runnable with no model download for the rule-based path. This is your shop window.
  3. Fill the test gap — the LLM path is currently mocked out and untested. Add at least one real browser-based test (Playwright/headless Chrome) that loads a model and exercises Tier 2.
  4. Reproducible benchmarks — replace the unsourced perf table with a script anyone can run.
  5. Publish to npm under a name that says what it does (e.g. redactkit). Discoverability is most of "useful in real life."

Phase 3 — Long-run differentiation & stickiness

  • Drop-in fetch wrapper / proxy shim that auto-redacts outgoing LLM calls and auto-restores responses — one-line integration. This is the moat: once it's in someone's request path, it stays.
  • Format-preserving redaction (keep the shape of a phone/card so the model still reasons correctly).
  • Custom entity patterns + allow/deny lists for enterprise.
  • Framework adapters (LangChain.js, Vercel AI SDK) — meet developers where they already are.
  • Compliance angle — the logger already produces an audit trail; lean into "provable, on-device redaction" for healthcare/legal/finance, where the real budget is.

What "useful" looks like (success signals)

  • A developer can npm i and redact PII before an OpenAI call in under 10 lines, with no model download.
  • The demo works offline in any modern browser.
  • The README claim "data never leaves the device" is actually true for the rule-based path.
  • Real adoption signal: GitHub issues/stars from people using it for the PII job, not the generic one.

Suggested sequencing

Phase 0 (now) → Phase 1 (the pivot — this is the make-or-break) → Phase 2 (ship it) → Phase 3 (grow it).

If Phase 1 proves the redact/restore round-trip is clean and fast, you have a real project. If it doesn't, you've spent a day finding that out cheaply — before investing in Phases 2–3.