Reflection Workflow for Structured Outputs

This notebook walks through setting up a Workflow to provide reliable structured outputs through retries and reflection on mistakes.

This notebook works best with an open-source LLM, so we will use Ollama. If you don’t already have Ollama running, visit https://ollama.com to get started and download the model you want to use. (In this case, we did ollama pull llama3.1 before running this notebook).

!pip install -U llama-index llama-index-llms-ollama

Since workflows are async first, this all runs fine in a notebook. If you were running in your own code, you would want to use asyncio.run() to start an async event loop if one isn’t already running.

async def main():
    <async code>

if __name__ == "__main__":
    import asyncio
    asyncio.run(main())

Designing the Workflow

To validate the structured output of an LLM, we need only two steps:

1. Generate the structured output
2. Validate that the output is proper JSON

The key thing here is that, if the output is invalid, we loop until it is, giving error feedback to the next generation.

The Workflow Events

To handle these steps, we need to define a few events:

1. An event to pass on the generated extraction
2. An event to give feedback when the extraction is invalid

The other steps will use the built-in StartEvent and StopEvent events.

from llama_index.core.workflow import Event


class ExtractionDone(Event):
    output: str
    passage: str


class ValidationErrorEvent(Event):
    error: str
    wrong_output: str
    passage: str

Item to Extract

To prompt our model, lets define a pydantic model we want to extract.

from pydantic import BaseModel


class Car(BaseModel):
    brand: str
    model: str
    power: int


class CarCollection(BaseModel):
    cars: list[Car]

The Workflow Itself

With our events defined, we can construct our workflow and steps.

Note that the workflow automatically validates itself using type annotations, so the type annotations on our steps are very helpful!

import json

from llama_index.core.workflow import (
    Workflow,
    StartEvent,
    StopEvent,
    Context,
    step,
)
from llama_index.llms.ollama import Ollama

EXTRACTION_PROMPT = """
Context information is below:
---------------------
{passage}
---------------------

Given the context information and not prior knowledge, create a JSON object from the information in the context.
The JSON object must follow the JSON schema:
{schema}

"""

REFLECTION_PROMPT = """
You already created this output previously:
---------------------
{wrong_answer}
---------------------

This caused the JSON decode error: {error}

Try again, the response must contain only valid JSON code. Do not add any sentence before or after the JSON object.
Do not repeat the schema.
"""


class ReflectionWorkflow(Workflow):
    max_retries: int = 3

    @step
    async def extract(
        self, ctx: Context, ev: StartEvent | ValidationErrorEvent
    ) -> StopEvent | ExtractionDone:
        current_retries = await ctx.store.get("retries", default=0)
        if current_retries >= self.max_retries:
            return StopEvent(result="Max retries reached")
        else:
            await ctx.store.set("retries", current_retries + 1)

        if isinstance(ev, StartEvent):
            passage = ev.get("passage")
            if not passage:
                return StopEvent(result="Please provide some text in input")
            reflection_prompt = ""
        elif isinstance(ev, ValidationErrorEvent):
            passage = ev.passage
            reflection_prompt = REFLECTION_PROMPT.format(
                wrong_answer=ev.wrong_output, error=ev.error
            )

        llm = Ollama(
            model="llama3",
            request_timeout=30,
            # Manually set the context window to limit memory usage
            context_window=8000,
        )
        prompt = EXTRACTION_PROMPT.format(
            passage=passage, schema=CarCollection.schema_json()
        )
        if reflection_prompt:
            prompt += reflection_prompt

        output = await llm.acomplete(prompt)

        return ExtractionDone(output=str(output), passage=passage)

    @step
    async def validate(
        self, ev: ExtractionDone
    ) -> StopEvent | ValidationErrorEvent:
        try:
            CarCollection.model_validate_json(ev.output)
        except Exception as e:
            print("Validation failed, retrying...")
            return ValidationErrorEvent(
                error=str(e), wrong_output=ev.output, passage=ev.passage
            )

        return StopEvent(result=ev.output)

And thats it! Let’s explore the workflow we wrote a bit.

• We have one entry point, extract (the steps that accept StartEvent)
• When extract finishes, it emits a ExtractionDone event
• validate runs and confirms the extraction:
  • If its ok, it emits StopEvent and halts the workflow
  • If nots not, it returns a ValidationErrorEvent with information about the error
• Any ValidationErrorEvent emitted will trigger the loop, and extract runs again!
• This continues until the structured output is validated

Run the Workflow!

NOTE: With loops, we need to be mindful of runtime. Here, we set a timeout of 120s.

w = ReflectionWorkflow(timeout=120, verbose=True)

# Run the workflow
ret = await w.run(
    passage="I own two cars: a Fiat Panda with 45Hp and a Honda Civic with 330Hp."
)

Running step extract
Step extract produced event ExtractionDone
Running step validate
Validation failed, retrying...
Step validate produced event ValidationErrorEvent
Running step extract
Step extract produced event ExtractionDone
Running step validate
Step validate produced event StopEvent

print(ret)

{ "cars": [ { "brand": "Fiat", "model": "Panda", "power": 45 }, { "brand": "Honda", "model": "Civic", "power": 330 } ] }