LLM-Driven Extraction of Unstructured Data — Built for API Deployments & ETL Pipeline Workflows
Python ai-agentsdata-engineeringdocument-aigenerative-aiidpjson-extractionllmmcp-serverocrpdf-extractionprompt-engineeringstructured-output
"""E2E: deploy a workflow as an API, POST a document, get structured JSON back.
Dispatches async and polls the status endpoint for the result: a file-bearing
execution only reaches COMPLETED once the callback rejoins, which the sync POST
cannot reliably wait out under CI load.
"""
from __future__ import annotations
import io
import pytest
from tests.e2e.api_deployment.conftest import (
ApiDeployment,
dispatch_async,
poll_delivered,
)
pytestmark = [pytest.mark.e2e, pytest.mark.critical]
@pytest.mark.critical_path("api-deployment-run")
@pytest.mark.critical_path("usage-token-tracking")
def test_api_deployment_returns_mocked_answer(
api_deployment: ApiDeployment, llm_mock_response: str
) -> None:
document = io.BytesIO(b"Hello invoice 123. This is a test document about widgets.")
_, status_url = dispatch_async(
api_deployment, {"files": ("probe.txt", document, "text/plain")}
)
body = poll_delivered(api_deployment, status_url, include_metadata=True)
assert body["status"] == "COMPLETED", body
file_result = body["message"][0]
assert file_result["status"] == "Success", file_result
assert file_result["result"]["output"]["answer"] == llm_mock_response
# 10/20/30 is litellm's fixed usage for one mocked completion. Counts are
# summed per reason, so a multiple of it means the prompt made extra calls
# rather than that tracking is broken.
usage = file_result["result"]["metadata"]["extraction_llm"][0]
counts = (usage["input_tokens"], usage["output_tokens"], usage["total_tokens"])
assert counts == (10, 20, 30), (
f"expected exactly one mocked extraction call, got {counts} — a multiple "
f"means the prompt issued more completions than this test assumes: {usage}"
)