mirror of
https://github.com/run-llama/human_in_the_loop_workflow_demo.git
synced 2026-07-01 21:24:04 -04:00
Laurie Voss
60 lines
2.7 KiB
Python
60 lines
2.7 KiB
Python
from llama_index.core.workflow import (
|
|
Workflow,
|
|
step,
|
|
Event,
|
|
Context
|
|
)
|
|
from llama_index.core.workflow.events import (
|
|
StartEvent,
|
|
StopEvent,
|
|
InputRequiredEvent,
|
|
HumanResponseEvent
|
|
)
|
|
|
|
# some event types to define the workflow:
|
|
|
|
# if the user says the research is not good enough, we retry the workflow
|
|
class RetryEvent(Event):
|
|
pass
|
|
|
|
# if the user says the research is good enough, we generate a report
|
|
class ReportEvent(Event):
|
|
pass
|
|
|
|
# we emit progress events to the frontend so the user knows what's happening
|
|
class ProgressEvent(Event):
|
|
pass
|
|
|
|
# this is a dummy workflow to show how to do human in the loop workflows
|
|
# the purpose of the flow is to research a topic, get human review, and then write a report
|
|
class HITLWorkflow(Workflow):
|
|
|
|
# this does the "research", which might involve searching the web or
|
|
# looking up data in a database or our vector store.
|
|
@step
|
|
async def research_query(self, ctx: Context, ev: StartEvent | RetryEvent) -> InputRequiredEvent:
|
|
ctx.write_event_to_stream(ProgressEvent(msg=f"I am doing some research on the subject of '{ev.query}'"))
|
|
await ctx.set("original_query", ev.query)
|
|
|
|
# once we've done the research, we send what we've found back to the human for review
|
|
# this gets handled by the frontend, and we expect a HumanResponseEvent to be sent back
|
|
return InputRequiredEvent(prefix="", query=ev.query,payload=f"Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed ut purus eget sapien. Nulla facilisi. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.")
|
|
|
|
# this accepts the HumanResponseEvent, which is either approval or rejection
|
|
# if it's approval, we write the report, otherwise we do more research
|
|
@step
|
|
async def human_review(self, ctx: Context, ev: HumanResponseEvent) -> ReportEvent | RetryEvent:
|
|
ctx.write_event_to_stream(ProgressEvent(msg=f"The human has responded: {ev.response}"))
|
|
if (ev.response == "yes"):
|
|
return ReportEvent(result=f"Here is the research on {await ctx.get('original_query')}")
|
|
else:
|
|
ctx.write_event_to_stream(ProgressEvent(msg=f"The human has rejected the research, retrying"))
|
|
return RetryEvent(query=await ctx.get("original_query"))
|
|
|
|
# this write the report, which would be an LLM operation with a bunch of context.
|
|
@step
|
|
async def write_report(self, ctx: Context, ev: ReportEvent) -> StopEvent:
|
|
ctx.write_event_to_stream(ProgressEvent(msg=f"The human has approved the research, generating final report"))
|
|
# generate a report here
|
|
return StopEvent(result=f"This is a report on {await ctx.get('original_query')}")
|