DOC: RetryWithErrorOutputParser usage examples

[austinmw]

Checklist

• I added a very descriptive title to this issue.
• I included a link to the documentation page I am referring to (if applicable).

Issue with current documentation:

I don't see any usage examples for RetryWithErrorOutputParser. I'd like to update the following code that uses chain.with_retry to use retry with errors instead, but it's not clear how from the documentation:

sonnet = ChatBedrock(model_id="anthropic.claude-3-sonnet-20240229-v1:0")

parser = PydanticOutputParser(pydantic_object=GeneratePlanStructAPI)

generate_args_for_gen_plan_struct_prompt_template = """
You are an expert at reading natural language feedback and generating the most relevant API arguments.

Look at the previous arguments and feedback below and generate the most appropriate arguments for the API endpoint.
Only modify arguments that are relevant to the feedback; leave the rest as they are.

Previous arguments:
{previous_generate_args}

User feedback: {feedback}

Important: Only output valid parsable JSON without any descriptions or comments. Follow the formatting instructions below:
{format_instructions}
"""

prompt = PromptTemplate(
    template=generate_args_for_gen_plan_struct_prompt_template,
    input_variables=["previous_generate_args", "feedback"],
    partial_variables={"format_instructions": parser.get_format_instructions()},
)

chain = prompt | sonnet | parser

retryable_gen_plan_chain = chain.with_retry(
    retry_if_exception_type=(ValueError,), # Retry only on ValueError
    wait_exponential_jitter=False,
    stop_after_attempt=5,
)

feedback = "I want to workout more days per week and do more cardio"

previous_generate_args = {
    "workout_prefs": ('strength', 'strength', 'intro'),
    "num_days": 3,
}

create_args = retryable_gen_plan_chain.invoke({"previous_generate_args": previous_generate_args, "feedback": feedback})

Idea or request for content:

No response

[austinmw]

Any help @dosubot? 😅

[dosubot]

To use RetryWithErrorOutputParser effectively in your scenario, you'll need to integrate it with your existing code structure. This involves setting up the parser to work with your PromptTemplate and handling retries based on parsing errors. Here's how you can adjust your code:

from langchain.output_parsers import RetryWithErrorOutputParser, PydanticOutputParser
from langchain.prompts import PromptTemplate
from langchain_openai import OpenAI

# Assuming GeneratePlanStructAPI is your Pydantic model
parser = PydanticOutputParser(pydantic_object=GeneratePlanStructAPI)

# Your existing prompt setup remains unchanged
generate_args_for_gen_plan_struct_prompt_template = """
...Your existing template...
"""

prompt = PromptTemplate(
    template=generate_args_for_gen_plan_struct_prompt_template,
    input_variables=["previous_generate_args", "feedback"],
    partial_variables={"format_instructions": parser.get_format_instructions()},
)

# Initialize RetryWithErrorOutputParser with your parser and an OpenAI instance
retry_parser = RetryWithErrorOutputParser.from_llm(llm=OpenAI(temperature=0), parser=parser)

# Example usage with a bad response and formatted prompt
bad_response = '{"action": "search"}'  # Example of a bad response
prompt_value = prompt.format_prompt(previous_generate_args=previous_generate_args, feedback=feedback)
parsed_output = retry_parser.parse_with_prompt(bad_response, prompt_value)
print(parsed_output)

This setup demonstrates how to incorporate RetryWithErrorOutputParser to handle retries based on parsing errors. The parser is designed to retry the parsing process by incorporating error messages into the retry prompt, potentially leading to better outcomes.

For more detailed examples and documentation, consider checking the official documentation of RetryWithErrorOutputParser. Remember, the effectiveness of this approach relies on the accurate identification and reporting of errors by the initial parser. Each retry involves an additional call to the language model, which may increase operation latency, especially if multiple retries are necessary.

Details

About Dosu
This response is meant to be useful and save you time. It is not meant to be a precise solution, but rather a starting point for your own research.

Help me be more useful! Please leave a 👍 if this is helpful and 👎 if it is irrelevant.

To continue the conversation, mention @dosu.

[keenborder786]

Why don't you create a custom corrector chain? and then use try,exception block to invoke the corrector chain:

try:
# your original chain invoke
except Exception as e:
# your corrector chain invoke with the exception error