Classify the document
PolDex identifies the insurance document family, schema fit, source type, and evidence posture inside the 56-schema release-ready universe.
PolDex is the vertical API for insurance extraction. Developers, operators, and agents use different surfaces, but every route feeds the same classifier, extraction core, credit ledger, export model, and deterministic output contracts.
Commercial P&C was the first proof wedge; the release-ready surface now spans the full 56-schema insurance universe. The workflow is built for the whole insurance universe: classify, extract, verify evidence, export, deliver, and reuse the same API contract.
PolDex identifies the insurance document family, schema fit, source type, and evidence posture inside the 56-schema release-ready universe.
The live schema contracts enforce required fields, evidence expectations, conflicts, confidence, unresolved items, abstention rules, and schema-scoped output shaping.
Results leave as canonical JSON plus CSV, XLSX, ZIP artifacts, signed links, webhooks, and connector events.
Every schema sits on the same 100+ document release gate behind the public 99%+ accuracy posture.
PolDex returns 202 Accepted immediately. No long blocking request cycle.
API, processor, playground, signed links, and email intake all converge on the same extraction core.
Paid production work shows credit estimates before holds are created and work begins.
Canonical JSON is enforced first, scoped to the requested schema, then CSV and XLSX are derived from that same result.
Webhooks are signed. Polling remains available. Delivery retries and DLQ states are explicit.
Developers use the API. Non-developers use the processor review cockpit. Proof and playground stay free for evaluation. Signed links and email intake support operational handoff.
# Direct API
POST /v1/extract
# No-code processor
POST /v1/batches/estimate
POST /v1/batches
# Playground
POST /v1/playground/run
# Agent interfaces
npx -y @poldex/mcp-server@0.0.3
poldex extract policy.pdf
# Signed intake links
POST /v1/intake/links
POST /v1/intake/links/{token}/submit
# Email intake
POST /v1/intake/email/registerPaid production work starts by verifying the API key and estimating credits from source type, page count, and complexity band. The user sees the estimate before work begins.
POST /v1/batches/estimate
{
"estimated_credits": 6,
"items": [
{
"source_name": "contractor-coi.pdf",
"estimated_pages": 4,
"estimated_credits": 2
}
],
"available_credits": 100
}When the user confirms the estimate, PolDex holds credits, creates jobs or batch items, and returns quickly. Long documents run asynchronously by default.
HTTP 202 Accepted
{
"batch_id": "batch_01hx4mz9p3kqa8",
"status": "queued",
"credits_held": 6,
"items": [
{
"item_id": "item_01hx4n0",
"source_name": "contractor-coi.pdf",
"status": "queued"
}
]
}Direct browser uploads stay capped for launch safety. Large documents use URL-based async processing. Workers stream document content in bounded chunks and avoid loading large PDFs as single blobs.
# Internal worker behavior # streamed retrieval -> temp storage # bounded memory profile # cleanup after processing
FastScript is the managed extraction engine across the completed schema families. Underlying model routing and failover stay internal while customer-facing output remains stable.
{
"schema_id": "commercial_gl",
"extractor": "fastscript-engine",
"model_name": "fastscript-engine",
"fallback_reason": null
}Facts are normalized, evidence-linked, truth-scored, and filtered back to the requested schema. Contradictions between forms and endorsements are returned as explicit conflicts.
{
"facts": [{
"field": "policy_number",
"status": "active",
"evidence": [{ "page": 1, "citation": "Declarations" }]
}],
"conflicts": []
}One canonical JSON result is produced, then CSV and XLSX exports are generated from the same source truth. Processor downloads inherit source filenames so bulk users do not lose track of files.
{
"source_name": "contractor-coi.pdf",
"result_links": {
"json": "https://.../contractor-coi.json",
"csv": "https://.../contractor-coi.csv",
"xlsx": "https://.../contractor-coi.xlsx"
}
}Developers receive data through signed webhooks, polling, or signed links. Processor users review facts and evidence, save decisions or field notes, then download JSON, CSV, XLSX, manifest CSV, ZIP, or combined exports.
POST https://your.app/webhook
X-PolDex-Signature: t=1745280000,v1=...
GET /v1/jobs/{job_id}
-> { "status": "complete", "result_links": { ... } }Successful jobs capture held credits. URL jobs reconcile holds after fetch/page estimation. Internal failures release holds. Delivery retries are bounded and terminal failures keep replay metadata.
{
"status": "dlq",
"failure_code": "delivery_failed_max_retries",
"credits_captured": 0
}Best for engineering teams that want full control with POST /v1/extract and webhook orchestration.
Best for AI agents and terminal environments that need MCP, CLI, OpenAPI, discovery files, and safe confirmation before paid extraction.
Best for operators who paste an API key, add files or URLs, approve an estimate, and download named JSON, CSV, or XLSX outputs.
Non-developers can run URL, upload, or pasted text through /playground against the same backend contracts.
Ops teams can receive private intake links with strict expiry and submit documents without handling raw API authentication.
Registered inboxes can forward attachments directly into intake. Jobs are created on the same extraction engine and return signed result links.
PolDex connects to existing workflows without replacing them: API for developers, webhooks for systems, processor review cockpit for ops teams, email intake for inbox workflows, signed links for third-party submissions, and CSV/XLSX exports for legacy systems.
Use the API, idempotency keys, polling, signed webhooks, and stable result contracts to wire PolDex into internal platforms.
Use the processor, email intake, signed links, review notes, decisions, and named exports when a team needs results without opening Postman.
CSV, XLSX, manifest files, and signed result links let teams move extracted data into spreadsheets and older systems.
Prebuilt connectors wrap the same rails for Reducto parse imports, Zapier, Make, n8n, Airtable, Google Sheets, Excel, OneDrive, SharePoint, Slack, and Teams.
AMS, CRM, underwriting, compliance, and procurement connectors are sequenced by customer demand, not guessed upfront.
The API remains the product boundary. Connectors are adapters, not a replacement dashboard or forced workflow suite.
PolDex does not ask users to guess cost. Short/simple documents start at 1 credit, while larger documents and batches are estimated before processing. Invalid, unsafe, or inaccessible inputs are rejected before capture.
Page and complexity bands produce a visible estimate before production work starts.
Confirmed work holds the estimated credits so balances cannot go negative.
Successful completed results capture the held credits from the immutable ledger.
System-side failures release held credits. Invalid user inputs are rejected before capture.
System-of-record payload with schema, evidence, conflicts, and metadata.
Flat operational views generated from the canonical JSON result for quick downstream use.
XLSX workbook output for broker and ops workflows that still run in spreadsheet environments.
Internal extraction failures release held credits and return explicit failure codes.
Delivery retries with backoff. Terminal failures move to DLQ and can be replayed from admin.
Insufficient credits return 402. Abuse controls return 429 with stable, contract-safe errors.
Use live proof, playground, or direct API integration on the live product surface.