Skip to content

Classify and route

classify reads the document and sends it down one edge per category; only the extraction on that edge runs, with the schema that fits. Below the confidence threshold the document takes the unclassified edge instead of being forced into a category.

The documents are synthetic (invented names, amounts and accounts), generated by scripts/docs_examples.py:

The graph

parse → classify
classify → (invoice) extract_invoice
classify → (receipt) extract_receipt
classify → (unclassified) extract_other

The full version body (graph and schemas), as the API takes it:

{
  "graph": {
    "nodes": [
      {
        "id": "parse",
        "type": "parse"
      },
      {
        "id": "classify",
        "type": "classify",
        "config": {
          "categories": [
            {
              "id": "invoice",
              "name": "Invoice",
              "description": "A bill with an invoice number"
            },
            {
              "id": "receipt",
              "name": "Receipt",
              "description": "A till receipt"
            }
          ],
          "threshold": 0.7,
          "unclassified_branch": true
        }
      },
      {
        "id": "extract_invoice",
        "type": "extract",
        "config": {
          "schema_id": "invoice"
        }
      },
      {
        "id": "extract_receipt",
        "type": "extract",
        "config": {
          "schema_id": "receipt"
        }
      },
      {
        "id": "extract_other",
        "type": "extract",
        "config": {
          "schema_id": "other"
        }
      }
    ],
    "edges": [
      {
        "source": "parse",
        "target": "classify"
      },
      {
        "source": "classify",
        "target": "extract_invoice",
        "branch": "invoice"
      },
      {
        "source": "classify",
        "target": "extract_receipt",
        "branch": "receipt"
      },
      {
        "source": "classify",
        "target": "extract_other",
        "branch": "unclassified"
      }
    ]
  },
  "schemas": {
    "invoice": {
      "type": "object",
      "properties": {
        "invoice_number": {
          "type": "string"
        },
        "total": {
          "type": "number"
        }
      }
    },
    "receipt": {
      "type": "object",
      "properties": {
        "merchant_name": {
          "type": "string"
        },
        "total": {
          "type": "number"
        }
      }
    },
    "other": {
      "type": "object",
      "properties": {
        "title": {
          "type": "string"
        }
      }
    }
  }
}

Create it and run it

Set DX_BASE_URL and DX_API_KEY (see the quickstart), cd to docs/examples/gallery/classify-route/ and run any of the three. They do the same.

Python

# SPDX-License-Identifier: Apache-2.0
# expect: "status": "succeeded"
"""Classify and route: create the workflow, run it on the document and print what comes back."""

import json
import os
import secrets
from pathlib import Path

import httpx

api = httpx.Client(
    base_url=os.environ["DX_BASE_URL"],
    headers={"Authorization": f"Bearer {os.environ['DX_API_KEY']}"},
    timeout=120,
)
name = f"classify-route-{secrets.token_hex(3)}"  # unique, so the example can run again

# 1. Create the workflow, add the version (graph + schemas), publish it, point `prod` at it.
workflow = api.post("/v1/workflows", json={"name": name}).raise_for_status().json()
version = (
    api.post(
        f"/v1/workflows/{workflow['id']}/versions",
        json=json.loads(Path("workflow.json").read_text()),
    )
    .raise_for_status()
    .json()
)
api.post(f"/v1/workflow-versions/{version['id']}:publish", json={}).raise_for_status()
api.put(
    f"/v1/workflows/{workflow['id']}/aliases/prod", json={"version_id": version["id"]}
).raise_for_status()

# 2. Run it on the document and wait for the result.
run = (
    api.post(
        "/v1/runs",
        data={"workflow": f"{name}@prod"},
        files=[("files", ("document.pdf", Path("document.pdf").read_bytes(), "application/pdf"))],
    )
    .raise_for_status()
    .json()
)
run = api.get(f"/v1/runs/{run['id']}", params={"wait": 60}).raise_for_status().json()
result = api.get(f"/v1/runs/{run['id']}/result").raise_for_status().json()

label = result["outputs"]["classifications"]["classify"]
view = {
    "status": run["status"],
    "label": label["label"],
    "confidence": label["confidence"]["band"],
    "ran": {node: e["data"] for node, e in result["outputs"]["extractions"].items()},
}
print(json.dumps(view, indent=2))

TypeScript

// SPDX-License-Identifier: Apache-2.0
// expect: "status": "succeeded"
// Classify and route: create the workflow, run it on the document and print what comes back.
// Node 22+: node --experimental-strip-types create.ts
import { readFileSync } from "node:fs";

const base = process.env.DX_BASE_URL as string;
const auth = { Authorization: `Bearer ${process.env.DX_API_KEY}` };
const name = `classify-route-${Math.random().toString(16).slice(2, 8)}`; // unique, so it can run again

async function call(path: string, init: RequestInit = {}): Promise<any> {
  const response = await fetch(base + path, { ...init, headers: { ...auth, ...init.headers } });
  if (!response.ok) throw new Error(`${response.status} ${path}: ${await response.text()}`);
  return response.json();
}
const json = (body: unknown): RequestInit => ({
  method: "POST",
  headers: { "Content-Type": "application/json" },
  body: JSON.stringify(body),
});

// 1. Create the workflow, add the version (graph + schemas), publish it, point `prod` at it.
const workflow = await call("/v1/workflows", json({ name }));
const version = await call(
  `/v1/workflows/${workflow.id}/versions`,
  json(JSON.parse(readFileSync("workflow.json", "utf8"))),
);
await call(`/v1/workflow-versions/${version.id}:publish`, json({}));
await call(`/v1/workflows/${workflow.id}/aliases/prod`, {
  ...json({ version_id: version.id }),
  method: "PUT",
});

// 2. Run it on the document and wait for the result.
const form = new FormData();
form.append("workflow", `${name}@prod`);
form.append("files", new Blob([readFileSync("document.pdf")], { type: "application/pdf" }), "document.pdf");
const started = await call("/v1/runs", { method: "POST", body: form });
const run = await call(`/v1/runs/${started.id}?wait=60`);
const result = await call(`/v1/runs/${run.id}/result`);

const label = result.outputs.classifications.classify;
const view = {
  status: run.status,
  label: label.label,
  confidence: label.confidence.band,
  ran: Object.fromEntries(
    Object.entries<any>(result.outputs.extractions).map(([node, e]) => [node, e.data]),
  ),
};
console.log(JSON.stringify(view, null, 2));

curl

#!/usr/bin/env bash
# SPDX-License-Identifier: Apache-2.0
# expect: "status": "succeeded"
# Classify and route: create the workflow, run it on the document and print what comes back.
# Needs curl and jq; DX_BASE_URL and DX_API_KEY in the environment.
set -euo pipefail
api() { curl -sS --fail-with-body -H "Authorization: Bearer $DX_API_KEY" "$@"; }
JSON='Content-Type: application/json'
NAME="classify-route-$RANDOM$RANDOM"  # unique, so the example can run again

# 1. Create the workflow, add the version (graph + schemas), publish it, point `prod` at it.
WORKFLOW=$(api -X POST "$DX_BASE_URL/v1/workflows" -H "$JSON" -d "{\"name\": \"$NAME\"}" | jq -r .id)
VERSION=$(api -X POST "$DX_BASE_URL/v1/workflows/$WORKFLOW/versions" -H "$JSON" -d @workflow.json | jq -r .id)
api -X POST "$DX_BASE_URL/v1/workflow-versions/$VERSION:publish" -H "$JSON" -d '{}' > /dev/null
api -X PUT "$DX_BASE_URL/v1/workflows/$WORKFLOW/aliases/prod" -H "$JSON" \
  -d "{\"version_id\": \"$VERSION\"}" > /dev/null

# 2. Run it on the document and wait for the result.
RUN=$(api -X POST "$DX_BASE_URL/v1/runs" -F "workflow=$NAME@prod" -F "[email protected]" | jq -r .id)
STATUS=$(api "$DX_BASE_URL/v1/runs/$RUN?wait=60" | jq -r .status)
api "$DX_BASE_URL/v1/runs/$RUN/result" | jq --arg status "$STATUS" '{status: $status, label: .outputs.classifications.classify.label, confidence: .outputs.classifications.classify.confidence.band, ran: (.outputs.extractions | map_values(.data))}'

Response

Produced by running the script above against a server with the fake model and these documents; make docs runs all the examples again and fails if an answer changes.

{
  "status": "succeeded",
  "label": "receipt",
  "confidence": "medium",
  "ran": {
    "extract_receipt": {
      "merchant_name": "CAFE LUNA",
      "total": 7.5
    }
  }
}