Intro

Frontier vision infrastructure

Docex routes images and PDFs through secure OCR, multimodal models, field evidence, fallback plans, structured output, and usage billing for products that need reliable file intelligence.

Continue with GitHub
Obvious choice

Why Docex is the obvious choice.

One run API for model choice, OCR, field evidence, fallbacks, billing, and traces.

  • Frontier modelsbest-fit routing
  • Many providersno hard lock-in
  • Vision + OCRsame interface
  • Evidence firstfield sources included
  • Cost awarespeed stays in view
  • Your workflowschema or prompt
  • Agent-firstapproval to code
  • Live in secondssetup, key, run
Engine

Innovating upstream of the application layer

Docex turns raw, unstructured files into model-ready evidence, typed JSON, traceable field confidence, and cost-visible provider execution.

01

Evaluation

Workflow-specific benchmarks, fixture runs, confidence scoring, and human verification loops.

02

Verifiers

Post-processing for OCR and vision outputs on high-stakes onboarding, finance, and operations tasks.

03

Reasoning trace data

Field sources, route decisions, provider fallback, timing, estimates, and run metadata captured in one envelope.

04

Provider post-routing

Capability matching across direct vision, OCR-first, fallback, and budget-aware strategies.

05

Security and privacy

Server-side keys, GitHub approval, wallet controls, hashed credentials, and provider-mode guardrails.

Production-grade routing on any image or PDF, at scale. 
await docex.run({
  file: uploadedFile,
  prompt: "extract onboarding risk signals",
  workflow: "kyc-review",
  outputFormat: "json"
})

// completed in 2.4s, charged $0.031
Evidence

Internal routing stays visible without becoming a marketplace.

Docex records the provider, model, route, cost, trace, and field evidence for every run.

OpenRouter Multimodal routing
Mistral OCR OCR-first PDFs
Mock Deterministic tests
Vision OCR PDF Structured JSON Boxes/layout Fallback
Routing

One run envelope across changing model terrain.

Provider adapters, schema defaults, upload storage, wallet billing, and trace output are unified behind a single server-side call.

PrimaryDirect multimodal

Use a frontier vision model when the file and prompt can be solved in one pass.

FallbackOCR then analyze

Extract text first when scans, PDFs, or dense forms need deterministic recovery.

BudgetCost-visible routing

Estimate before queueing and true up after provider usage is known.

RegistrySchema defaults

Use common extraction presets or provide a schema at request time.

CIMock provider

Run tests without external keys, network dependency, or real provider spend.

AnyYour workflow

Describe the task. Docex plans the route and returns structured output.

Deploy

Docex is ready for agent-assisted integration

Paste the setup prompt into your coding agent or run the CLI scaffold. It will detect your stack, request human approval, wire server-side env vars, scaffold an analyze route, and run a smoke test.

prompt.txt 
Install and wire Docex into this project as the vision analysis layer for
[describe the use case, for example "reading trade licenses during onboarding"].

Use the package docexdev. Run:
npx docexdev setup --use-case "[same use case]" --framework auto --top-up 5 --base-url https://api.docex.dev --scaffold --json

Show me the approval URL, wait for GitHub and wallet approval, store
DOCEX_API_KEY and DOCEX_BASE_URL server-side only, scaffold POST /api/analyze
for file + prompt, call createDocex().run(), preserve field evidence in the
response, and run a smoke test end-to-end.
Transform

Messy files go in. Clean JSON comes out.

A dense paper form crosses the parsing line and resolves into normalized fields, source evidence, confidence, and trace-ready structure.

Raw paper Formatted JSON