Skip to main content

​
Same document, three URLs

OkraPDF’s redaction engine runs server-side at the edge. PII is removed before the response leaves the Worker — it never reaches the browser. 
/s/{admin-token}/fw9.md   → full text
/s/{viewer-token}/fw9.md  → SSN: ***-**-****, [EMAIL], [PHONE]
/s/{public-token}/fw9.md  → allowlisted sections only
Each URL is an HMAC-signed capability token. No API keys, no sessions, no cookies. The token IS the auth.

​
How it works

  1. Parse your PDF with any vendor (LlamaParse, Docling, Unstructured, Azure Doc Intel)
  2. Deploy with @okrapdf/edge-kit — pass pages + redaction config
  3. Get back 3 URLs — admin, viewer, public
import { deploy } from '@okrapdf/edge-kit';

const pii = {
  preset: 'hipaa',
  patterns: ['SSN', 'EMAIL', 'PHONE_US', 'TAX_ID_US'],
  includeNames: true,
  includeAddresses: true,
};

const result = await deploy({
  pages,  // from any PDF parser
  meta: { title: 'W-9', filename: 'fw9.pdf' },
  redact: {
    pii,
    publicFieldAllowlist: ['Form W-9', 'General Instructions'],
  },
  apiKey: process.env.OKRA_API_KEY!,
});

result.urls.admin   // full text
result.urls.viewer  // PII redacted
result.urls.public  // allowlist + redacted

​
PII detection with OpenRedaction

Pass a pii config object and the SDK uses OpenRedaction under the hood — compliance presets, name/address detection, context-aware matching, and 400+ pattern types out of the box. No pii field uses OpenRedaction defaults (all patterns enabled). 
const pii = {
  preset: 'hipaa',                                   // or 'gdpr', 'ccpa'
  patterns: ['SSN', 'EMAIL', 'PHONE_US', 'TAX_ID_US'],
  includeNames: true,
  includeAddresses: true,
};

​
Custom patterns

For domain-specific patterns, pass customPatterns — raw regex alongside presets: 
const pii = {
  preset: 'hipaa',
  customPatterns: [
    { type: 'DEAL_VALUE', regex: /\$[\d,]+\.\d{2}/g, priority: 10, placeholder: '[AMOUNT_{n}]', severity: 'high' },
    { type: 'INTERNAL_REF', regex: /REF-[A-Z]{3}-\d{4}/g, priority: 5, placeholder: '[REF_{n}]', severity: 'medium' },
  ],
};

​
Config-per-document

Each document gets its own redaction config. No global settings to manage. 
// Tax forms: HIPAA preset, names + addresses
await deploy({
  pages: w9Pages,
  redact: {
    pii: { preset: 'hipaa', includeNames: true, includeAddresses: true },
    publicFieldAllowlist: ['Form W-9', 'Part I'],
  },
  apiKey,
});

// Contracts: custom patterns for deal values
await deploy({
  pages: contractPages,
  redact: {
    pii: {
      customPatterns: [
        { type: 'DEAL_VALUE', regex: /\$[\d,]+\.\d{2}/g, priority: 10, placeholder: '[AMOUNT_{n}]', severity: 'high' },
      ],
    },
    publicFieldAllowlist: ['Terms', 'Parties'],
  },
  apiKey,
});

​
Vendor-agnostic

The PageInput format works with any parser: 
interface PageInput {
  pageNum: number;
  text: string;
  items?: Array<{ text: string; bbox?: { x: number; y: number; w: number; h: number } }>;
}
No vendor lock-in. Parse with LlamaParse today, switch to Docling tomorrow — redaction works the same.

​
Every access path, not just URLs

Redaction isn’t just for static markdown URLs. The same lens applies to:
  • Completions endpoint — agent tool results are redacted before the LLM sees them. The model can’t leak PII it never received.
  • Agent SQL queries — query_sql results pass through the lens. A SELECT * FROM nodes returns [REDACTED] for PII fields.
  • Text search — search results are filtered through the active role. Searching for a raw SSN against a viewer token returns zero matches.
See the Redact & Deploy cookbook for implementation details.

​
Architecture

  • Redaction runs on Cloudflare Workers — sub-5ms, no cold starts
  • Pages stored in R2 — zero egress fees
  • HMAC-signed tokens — no database lookup needed to verify
  • Markdown output — Content-Type: text/markdown, no HTML rendering overhead
See the Redact & Deploy cookbook for a full working example.