Automate browser based workflows with AI
Python aiapiautomationbrowserbrowser-automationcomputergptllmplaywrightpowerautomatepuppeteerpythonrpaseleniumvisionworkflow
Fix: retry empty LLM responses in extraction task generation (#SKY-8264) (#5103)
Shuchang Zheng54b73c0ec01a
diff --git a/skyvern/services/task_v2_service.py b/skyvern/services/task_v2_service.pyindex d644ab1f6..0c5659d67 100644--- a/skyvern/services/task_v2_service.py+++ b/skyvern/services/task_v2_service.py@@ -18,6 +18,7 @@ from skyvern.exceptions import (from skyvern.forge import appfrom skyvern.forge.prompts import prompt_enginefrom skyvern.forge.sdk.api.llm.api_handler_factory import LLMAPIHandlerFactory+from skyvern.forge.sdk.api.llm.exceptions import EmptyLLMResponseError, InvalidLLMResponseFormatfrom skyvern.forge.sdk.artifact.models import ArtifactTypefrom skyvern.forge.sdk.core import skyvern_contextfrom skyvern.forge.sdk.core.hashing import generate_url_hash@@ -881,17 +882,26 @@ async def run_task_v2_helper(breakif task_type == "extract":- block, block_yaml_list, parameter_yaml_list = await _generate_extraction_task(- task_v2=task_v2,- workflow_id=workflow_id,- workflow_permanent_id=workflow.workflow_permanent_id,- workflow_run_id=workflow_run_id,- current_url=current_url,- scraped_page=scraped_page,- data_extraction_goal=plan,- task_history=task_history,- )- task_history_record = {"type": task_type, "task": plan}+ try:+ block, block_yaml_list, parameter_yaml_list = await _generate_extraction_task(+ task_v2=task_v2,+ workflow_id=workflow_id,+ workflow_permanent_id=workflow.workflow_permanent_id,+ workflow_run_id=workflow_run_id,+ current_url=current_url,+ scraped_page=scraped_page,+ data_extraction_goal=plan,+ task_history=task_history,+ )+ task_history_record = {"type": task_type, "task": plan}+ except Exception:+ LOG.exception("Failed to generate extraction task")+ task_v2 = await mark_task_v2_as_failed(+ task_v2_id=task_v2_id,+ workflow_run_id=workflow_run_id,+ failure_reason="Failed to generate the extraction task.",+ )+ breakelif task_type == "navigate":original_url = url if i == 0 else Nonenavigation_goal = MINI_GOAL_TEMPLATE.format(main_goal=user_prompt, mini_goal=plan)@@ -1482,12 +1492,30 @@ async def _generate_extraction_task(local_datetime=datetime.now(context.tz_info).isoformat(),)- generate_extraction_task_response = await app.LLM_API_HANDLER(- generate_extraction_task_prompt,- task_v2=task_v2,- prompt_name="task_v2_generate_extraction_task",- organization_id=task_v2.organization_id,- )+ max_retries = 3+ last_exc: Exception | None = None+ generate_extraction_task_response: dict[str, Any] | None = None+ for attempt in range(max_retries):+ try:+ generate_extraction_task_response = await app.LLM_API_HANDLER(+ generate_extraction_task_prompt,+ task_v2=task_v2,+ prompt_name="task_v2_generate_extraction_task",+ organization_id=task_v2.organization_id,+ )+ break+ except (InvalidLLMResponseFormat, EmptyLLMResponseError) as e:+ LOG.warning(+ "Empty or invalid LLM response during extraction task generation, retrying",+ attempt=attempt + 1,+ max_attempts=max_retries,+ error=str(e),+ )+ last_exc = e+ else:+ raise last_exc # type: ignore[misc]++ assert generate_extraction_task_response is not NoneLOG.info("Data extraction response", data_extraction_response=generate_extraction_task_response)# create OutputParameter for the data_extraction block