Offloading Extraction

Offloading extraction

POST /v1/extract is the single submit endpoint. It returns 202 Accepted with job_ids (extraction) and crawl_job_ids (URL crawls) — never the extraction result inline. Pair every submit with either a poll loop (tracking-cloud-jobs skill) or a webhook.

When to reach for this

• File is on a remote URL.
• File is on disk but the local kreuzberg CLI is not installed.
• You want server-side parallelism for a batch.
• The user wants webhook-delivered results to skip blocking.
• File is larger than ~50 MB → use presigned-uploads instead — the base64 JSON body is too big.

Endpoint

POST https://api.kreuzberg.dev/v1/extract
Authorization: Bearer $KREUZBERG_API_KEY
Content-Type: application/json | multipart/form-data

Returns 202 Accepted with ExtractResponse.

Three submission shapes

1. Base64 JSON (small files, <5 MB recommended)

curl -X POST https://api.kreuzberg.dev/v1/extract \
  -H "Authorization: Bearer $KREUZBERG_API_KEY" \
  -H "Content-Type: application/json" \
  -d @- <<JSON
{
  "documents": [
    {
      "filename": "invoice.pdf",
      "mime_type": "application/pdf",
      "data": "$(base64 -w0 invoice.pdf)"
    }
  ],
  "options": {
    "extraction_config": {
      "output_format": "markdown",
      "ocr": { "backend": "tesseract", "language": "eng" }
    }
  }
}
JSON

2. Multipart (binary, recommended for anything over ~1 MB)

curl -X POST https://api.kreuzberg.dev/v1/extract \
  -H "Authorization: Bearer $KREUZBERG_API_KEY" \
  -F "file=@invoice.pdf;type=application/pdf" \
  -F 'options={"extraction_config":{"output_format":"markdown"}};type=application/json'

Add a webhook part as a JSON string:

  -F 'webhook={"url":"https://hooks.example.com/x","secret":"shh"};type=application/json'

3. URL crawl

curl -X POST https://api.kreuzberg.dev/v1/extract \
  -H "Authorization: Bearer $KREUZBERG_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "urls": [{"url": "https://example.com/docs"}],
    "crawl_config": {"max_depth": 2, "max_pages": 50, "stay_on_domain": true},
    "webhook": {"url": "https://hooks.example.com/x"}
  }'

URL crawls return crawl_job_ids instead of (or alongside) job_ids.

Response (202)

{
  "job_ids": ["550e8400-e29b-41d4-a716-446655440000"],
  "crawl_job_ids": [],
  "status": "pending"
}

status is always pending at submit time; the per-job status is retrieved via GET /v1/jobs/{id}.

The options block

Shape mirrors the local ExtractionConfig:

{
  "extraction_config": {
    "output_format": "markdown",
    "ocr": { "backend": "tesseract", "language": "eng+deu" },
    "extract_tables": true,
    "extract_images": false,
    "chunking": { "max_chars": 4000, "overlap": 200 }
  }
}

Supported output_format values: markdown, text, json, djot, html. Default is markdown.

The webhook block

{
  "url": "https://hooks.example.com/x",
  "secret": "shared-secret-32-bytes-min",
  "metadata": { "request_id": "abc123", "user_id": "u_42" }
}

secret is the HMAC key used to sign the webhook payload — see tracking-cloud-jobs for verification. metadata is echoed back in the delivered payload, useful for correlating server-side requests.

TypeScript SDK

import { KreuzbergCloud } from "@kreuzberg/cloud";
import { readFile } from "node:fs/promises";

const client = new KreuzbergCloud({ apiKey: process.env.KREUZBERG_API_KEY! });

const data = await readFile("invoice.pdf");
const job = await client.extract({
  file: { name: "invoice.pdf", data, mimeType: "application/pdf" },
  options: { extractionConfig: { outputFormat: "markdown" } },
});
console.log(job.id, job.status);

For submit + wait in one call:

const result = await client.extractAndWait({
  file: { name: "invoice.pdf", data },
});
console.log(result.result?.content);

Python SDK

from pathlib import Path
from kreuzberg_cloud import KreuzbergCloud

with KreuzbergCloud(api_key=os.environ["KREUZBERG_API_KEY"]) as client:
    job = client.extract(file=Path("invoice.pdf"))
    print(job.id, job.status)

Submit + wait:

job = client.extract_and_wait(file=Path("invoice.pdf"))
print(job.result.content if job.result else job.status)

Batch submission

JSON: pass multiple entries in documents. Multipart: repeat the file part. SDKs expose extractBatch / extract_batch helpers that fan out correctly per platform (parallel HTTP for the async Python client, sequential for the sync one).

Errors

Next step

After every submit, hand off to the tracking-cloud-jobs skill — cloud extraction is asynchronous and the result is delivered via either polling or webhook callback. Never assume a result is ready immediately after the 202 response.