CV
Nested data from a CV: a list of small objects (skills) and a list of objects (experience). Fields marked x-dx.pii are tagged as personal data in the schema, so downstream redaction and retention rules can find them.
The documents are synthetic (invented names, amounts and accounts), generated by
scripts/docs_examples.py:
The graph
parse → extract
The full version body (graph and schemas), as the API takes it:
{
"graph": {
"nodes": [
{
"id": "parse",
"type": "parse"
},
{
"id": "extract",
"type": "extract",
"config": {
"schema_id": "cv"
}
}
],
"edges": [
{
"source": "parse",
"target": "extract"
}
]
},
"schemas": {
"cv": {
"type": "object",
"x-dx": {
"document_type": "cv"
},
"properties": {
"full_name": {
"type": "string"
},
"email": {
"type": "string",
"x-dx": {
"pii": "personal"
}
},
"skills": {
"type": "array",
"items": {
"type": "object",
"properties": {
"name": {
"type": "string"
}
}
}
},
"experience": {
"type": "array",
"items": {
"type": "object",
"properties": {
"company": {
"type": "string"
},
"role": {
"type": "string"
},
"start_year": {
"type": "integer"
},
"end_year": {
"type": "integer"
}
}
}
}
}
}
}
}
Create it and run it
Set DX_BASE_URL and DX_API_KEY (see the quickstart), cd to
docs/examples/gallery/cv/ and run any of the three. They do the same.
Python
# SPDX-License-Identifier: Apache-2.0
# expect: "status": "succeeded"
"""CV: 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"cv-{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", ("cv.pdf", Path("cv.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()
node = result["outputs"]["extractions"]["extract"]
view = {
"status": run["status"],
"data": node["data"],
"citations": [
{"path": f["path"], "page": f["evidence"][0]["page"], "quote": f["evidence"][0]["quote"]}
for f in node["fields"]
if f["evidence"] and "[" not in f["path"]
],
}
print(json.dumps(view, indent=2))
TypeScript
// SPDX-License-Identifier: Apache-2.0
// expect: "status": "succeeded"
// CV: 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 ${process.env.DX_API_KEY}` };
const name = `cv-${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", `${name}@prod`);
form.append("files", new Blob([readFileSync("cv.pdf")], { type: "application/pdf" }), "cv.pdf");
const started = await call("/v1/runs", { method: "POST", body: form });
const run = await call(`/v1/runs/${started.id}?wait=60`);
const result = await call(`/v1/runs/${run.id}/result`);
const node = result.outputs.extractions["extract"];
const view = {
status: run.status,
data: node.data,
citations: node.fields
.filter((f: any) => f.evidence.length > 0 && !f.path.includes("["))
.map((f: any) => ({ path: f.path, page: f.evidence[0].page, quote: f.evidence[0].quote })),
};
console.log(JSON.stringify(view, null, 2));
curl
#!/usr/bin/env bash
# SPDX-License-Identifier: Apache-2.0
# expect: "status": "succeeded"
# CV: 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="cv-$RANDOM$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 "files=@cv.pdf" | 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, data: .outputs.extractions["extract"].data, citations: [.outputs.extractions["extract"].fields[] | select((.evidence | length) > 0 and (.path | contains("[") | not)) | {path, page: .evidence[0].page, quote: .evidence[0].quote}]}'
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",
"data": {
"email": "lucia@example.com",
"experience": [
{
"company": "Initech Labs",
"end_year": 2024,
"role": "Data engineer",
"start_year": 2021
},
{
"company": "Hooli Systems",
"end_year": 2021,
"role": "Analyst",
"start_year": 2018
}
],
"full_name": "Lucia Example",
"skills": [
{
"name": "Python"
},
{
"name": "SQL"
},
{
"name": "Docker"
}
]
},
"citations": [
{
"path": "email",
"page": 1,
"quote": "lucia@example.com"
},
{
"path": "full_name",
"page": 1,
"quote": "Lucia Example"
}
]
}