# Documents in. Grounded data out.

import { Card, CardHeader, CardTitle, CardDescription, CardContent } from "zudoku/ui/Card.js";
import { Button } from "zudoku/ui/Button.js";

<div className="flex flex-col items-center text-center gap-5 pt-2 pb-12 border-b border-border">

<p className="text-lg text-muted-foreground max-w-2xl">
DocAI parses PDFs, scans and spreadsheets into markdown, table geometry and
schema-shaped JSON — with page coordinates on every segment, so an answer can
point at the line it came from.
</p>

<div className="flex gap-3 flex-wrap justify-center">
  <Button asChild size="lg"><a href="/quickstart">Start in four calls</a></Button>
  <Button asChild size="lg" variant="outline"><a href="/api">Browse the API</a></Button>
</div>

<p className="text-sm text-muted-foreground">45 operations · 36 endpoints · 128 documented error codes</p>

</div>

<div className="grid md:grid-cols-3 gap-4 py-12">

<Card>
  <CardHeader>
    <CardTitle>Parse</CardTitle>
    <CardDescription>Upload anything, get a bundle back</CardDescription>
  </CardHeader>
  <CardContent>
    Markdown in reading order, per-segment bounding boxes, table cell geometry,
    page renders and extracted figures. Born-digital and scanned take different
    paths automatically.
  </CardContent>
</Card>

<Card>
  <CardHeader>
    <CardTitle>Extract</CardTitle>
    <CardDescription>Your schema, filled in</CardDescription>
  </CardHeader>
  <CardContent>
    Describe the fields you want and get JSON shaped to them, with citations
    back to the exact segment each value came from. Generate a schema from a
    document if you do not have one yet.
  </CardContent>
</Card>

<Card>
  <CardHeader>
    <CardTitle>Search</CardTitle>
    <CardDescription>Evidence, not embeddings</CardDescription>
  </CardHeader>
  <CardContent>
    Full-text search across a knowledge base returns file and segment pointers
    you can cite. It calls no model, so it costs no credits — search first,
    then read only what you need.
  </CardContent>
</Card>

</div>

<div className="py-12 border-t border-border">

## Work from your editor

The MCP server puts 23 DocAI tools inside Claude Code, Cursor and Codex —
upload, parse, extract, split, search, and sync a local folder — without
leaving the file you are in.

```bash
/plugin marketplace add ProvidusAI/docai-plugins
/plugin install docai
```

<div className="flex gap-3 flex-wrap pt-2">
  <Button asChild variant="outline"><a href="/mcp">What the tools do</a></Button>
  <Button asChild variant="outline"><a href="/plugins">Install for Cursor or Codex</a></Button>
</div>

</div>

<div className="grid md:grid-cols-2 gap-4 pb-16">

<Card>
  <CardHeader>
    <CardTitle>Cloud or your own hardware</CardTitle>
  </CardHeader>
  <CardContent>
    The same API, the same tools and the same docs run against
    `api.providus.ai` or an on-prem deployment. Point `DOCAI_BASE_URL` at your
    instance; nothing else changes.
  </CardContent>
</Card>

<Card>
  <CardHeader>
    <CardTitle>Errors you can branch on</CardTitle>
  </CardHeader>
  <CardContent>
    Every failure returns a stable `code` beside its human-readable message.
    All 128 are listed, generated from the API's own registry — so the
    reference cannot fall behind the code.
  </CardContent>
</Card>

</div>

<div className="py-12 border-t border-border">

## What a document becomes

A parse produces a bundle, not a blob:

| Artifact | What it is |
|---|---|
| `result_md` | The document as markdown, in reading order |
| `grounding_json` | Every segment with its page and bounding box |
| `tables_json` | Table structure with cell geometry |
| `result_json` | Schema-shaped extraction output |
| `pages`, `imgs` | Page renders and extracted figures |

Because every segment carries coordinates, an answer can point at the exact
place in the source it came from. That is what "grounded" means here, and it is
the difference between a citation and a guess.

</div>

<div className="py-12 border-t border-border">

## Where to go

Read in this order if you are new. Skip to the page you need if you are not.

- [Quickstart](/quickstart) — start here: keys, the four calls that parse a
  document, and the conventions every endpoint shares.
- [Parsing](/parsing) — for anyone who needs the artifacts a parse produces and
  the coordinates that ground them.
- [Extraction](/extraction) — for pulling schema-shaped JSON, with citations,
  out of a parsed document.
- [Knowledge bases and search](/knowledge-bases) — for organising documents and
  retrieving evidence without spending credits.
- [Split](/split) and [Classify](/classify) — separate combined PDFs and
  label documents; [document workflows](/workflows) covers the queued
  operation model they share, plus edit.
- [MCP server](/mcp) — for what each of the 23 agent tools does and what it
  costs.
- [Plugins and commands](/plugins) — for installing those tools into Claude
  Code, Cursor or Codex.
- [Agent guidance](/skill) — for the four rules that ship with the tools, and
  why they are the safety mechanism.
- [API reference](/api) — for every endpoint, request body and response.
- [Error codes](/errors) — for the stable `code` behind each failure and what
  to do about it.

</div>