LLM-Driven Extraction of Unstructured Data — Built for API Deployments & ETL Pipeline Workflows
Python ai-agentsdata-engineeringdocument-aigenerative-aiidpjson-extractionllmmcp-serverocrpdf-extractionprompt-engineeringstructured-output
from typing import Any
from workflow_manager.workflow_v2.enums import ExecutionStatus
MESSAGE_KEY = "message"
def flatten_error(data: dict[str, Any]) -> str:
"""Join the details without their `attr`, as the published Python client
already renders a refused request, so the text callers see is unchanged.
"""
errors = data.get("errors")
if isinstance(errors, list):
return "; ".join(
str(err["detail"])
for err in errors
if isinstance(err, dict) and err.get("detail")
)
if errors is not None:
return str(errors)
return str(data.get("error", ""))
def add_error_message(data: Any) -> Any:
"""Mirror an error body into the `message.error` slot the execute response
uses, so a client reads one key for every outcome.
Strictly additive: published clients parse the existing keys, so they are
left exactly as the exception handler produced them.
"""
if not isinstance(data, dict) or MESSAGE_KEY in data:
return data
data[MESSAGE_KEY] = {
"execution_status": ExecutionStatus.ERROR.value,
"error": flatten_error(data),
}
return data