textsift

Personal learning project. Treat as such — no SLA, no roadmap commitment. See the main README for context.

PII detection + redaction running openai/privacy-filter on the user's device. Per-platform GPU backends (Metal on macOS, Vulkan on Linux, Dawn on Windows, WGSL in browsers); Zig + SIMD128 WASM as the no-GPU fallback. The native BPE tokenizer is pure TS, so apps that don't already ship @huggingface/transformers save a multi-MB dep.

npm install textsift

Two entry points so browsers never bundle native code:

// Browser / Node-via-WASM
import { PrivacyFilter } from "textsift/browser";

// Node — auto-picks the platform-native fast path (Metal on macOS,
// Vulkan on Linux, Dawn on Windows) and falls back to WASM if no GPU.
import { PrivacyFilter } from "textsift";

const filter = await PrivacyFilter.create();
const { redactedText } = await filter.redact(
  "Hi John Smith, your email [email protected] is on file.",
);

Why this matters on Linux

If you're on Linux and want GPU-accelerated PII filtering in Node, the realistic options today are bad:

| Path | Setup | T=32 latency on a typical iGPU box | |---|---|---:| | ONNX Runtime Node CPU | npm i onnxruntime-node, write your own inference loop | ~600–800 ms | | transformers.js (Node) | npm i @xenova/transformers, no GPU on Node so WASM | ~80–100 ms | | PyTorch CPU | pip install torch transformers safetensors, write inference | ~150–500 ms | | PyTorch CUDA | NVIDIA GPU + driver + cuda-toolkit + matched torch wheel | n/a (no NVIDIA on most laptops) | | textsift native | npm install textsift | ~28 ms |

On the same Linux box (Intel Iris Xe, Mesa Vulkan), textsift native is 22–28× faster than ORT Node CPU because it talks Vulkan directly with hand-written GLSL→SPIR-V kernels — no CUDA, no driver dance, no model conversion.

End-to-end on a 122-character input with 4 PII spans: redact() returns in ~50–75 ms.

Per-platform fast paths

| Platform | Backend | What it uses | |---|---|---| | macOS arm64/x64 | Metal-direct | Hand-written MSL kernels via Obj-C bridge | | Linux x86_64/arm64 | Vulkan-direct | Hand-written GLSL → SPIR-V via glslangValidator | | Windows x86_64 | Dawn-direct | Tint → D3D12 via statically-linked Google Dawn | | (any platform, no GPU) | WASM fallback | Zig + SIMD128 in WebAssembly |

Each platform's .node binary is built with comptime-gated Zig code so it only contains the relevant backend — Mac binaries don't ship Vulkan code, Windows binaries don't ship Obj-C, etc. npm picks the right optionalDependencies subpackage at install time (textsift-{linux-x64,linux-arm64,darwin-x64,darwin-arm64,windows-x64}).

Linux prereqs (one-time)

For the GPU fast path on Linux, you need a Vulkan loader. Most distros ship one in their default packages:

# Ubuntu/Debian
sudo apt install -y libvulkan1 mesa-vulkan-drivers

# Fedora/RHEL
sudo dnf install -y vulkan-loader mesa-vulkan-drivers

# Arch
sudo pacman -S vulkan-icd-loader vulkan-mesa-layers

If Vulkan isn't available, import { PrivacyFilter } from "textsift" automatically falls back to the WASM CPU path — same API, slower runtime (still faster than ORT Node CPU thanks to Zig SIMD128 kernels).

Why two entry points

Bundlers (Vite/Webpack/esbuild/etc.) resolve textsift/browser and pull in only the WASM/WebGPU code path. The native NAPI binding lives at the bare textsift import, so a Node CLI / server can use it without forcing browser code into anything else's bundle.

Public API

See the API reference. Highlights:

• PrivacyFilter.create({ backend, modelSource, markers, enabledCategories, rules, presets, minConfidence, cacheDir, modelPath, offline })
• filter.detect(text | AsyncIterable<string>) — batch returns a Promise; streaming returns a sync handle with spanStream + result
• filter.redact(text | AsyncIterable<string>) — same shape; streaming surfaces textStream of safe-to-emit pieces
• filter.classifyColumns(rows, { headerRow, sampleSize }) — per-column PII classification for tabular data
• filter.redactTable(rows, { mode }) — "redact" / "synth" / "drop_column" for one-shot CSV cleaning
• presets: ["secrets"] enables JWT, GitHub PAT, AWS, Slack, OpenAI/Anthropic/Google/Stripe keys, and PEM private-key headers (all severity "block")
• markerPresets.faker() — realistic-looking fake values instead of [label] markers (consistent across mentions)
• Custom rules (regex or function) merge with model spans
• SARIF v2.1.0 export at textsift/sarif for GitHub Code Scanning / similar consumers

Other surfaces

Same engine, four surfaces total:

• npx textsift — CLI (redact, detect, table, classify, download, cache).
• Pre-commit hook — block commits containing PII.
• GitHub Action — uses: teamchong/textsift@v1 with PR annotations + Security-tab integration via SARIF.

License

Apache 2.0.