integrate structured output into chains?

[dmlond]

It does not seem like it is possible to run a chain that produces structured output. Am I missing something? It seems like there are a couple of places where structured output would be useful.

1. Within an agent that runs in a chain.
   I had to ensure that the fetch agent returned structured output from a url, and some urls do not. I had to add `respond in YAML format" to the standard fetch agent information input

@fast.agent(
    "gather",
    "Given a URL, provide a complete and comprehensive summary of the required attachments in Section IV Application and Submission Information. respond in YAML format",
    servers=["fetch"], # Name of an MCP Server defined in fastagent.config.yaml
)

Seems like having a way to inject a Pydantic or other supported structured output would be nicer

@fast.agent(
  name="gather",
  instruction="Given a URL....",
  structured=SomeModel
)

2. At the end of a chain, it would be nice to finanlize it into structured output.
   This could be accomplished if the last agent run in a chain was configured with a structured, e.g. 1. Or you could allow something like

async def main():
    async with fast.run() as agent:
        # using chain workflow
        requirements, message = await agent.mychain(args).structured(model=SomeModel)
        ....

[dmlond]

I tried

@fast.agent(
    name="gather",
    instruction="Given a URL, provide a complete and comprehensive summary of the required attachments in Section IV Application and Submission Information",
    request_params=RequestParams(
        response_format=RequiredAttachments
    ),
    servers=["fetch"], # Name of an MCP Server defined in fastagent.config.yaml
)

and got
raise TypeError(
"You tried to pass a BaseModel class to chat.completions.create(); You must use chat.completions.parse() instead"
)

[evalstate]

I think what we want is a Generate -> Structured combo isn't it? (I need this as well so it's a priority). QQ -> Are you ever in a situation where you need a longer chain than the pair?

[ChrisW-priv]

looks like there already is a solution for chains? there exists structured_impl definition

[dmlond]

where is this documented?

[ChrisW-priv]

Xd, touché. Documentation is certainly lacking. But there exists an example:
https://github.com/evalstate/fast-agent/blob/main/examples%2Fotel%2Fagent.py#L31

And it should work with other agent types as well. Tested it in evaluator and it works, so since code impl is the same I can confidently say it will work too. Would you mind creating an example in the examples dir and make a PR? Would be super helpful for others.

[dmlond]

This is different than what I am asking for. I know how to get structured output from a single agent, like this example. I cannot see how to make one agent in a chain structure its output before sending it to the next agent as input,

@fast.agent(
    "url_fetcher",
    "Given a URL, provide a complete and comprehensive summary",
    servers=["fetch"], # Name of an MCP Server defined in fastagent.config.yaml,
    structured=SomeModel
)
@fast.agent(
    "social_media",
    """
    Write a 280 character social media post for any given text.
    Respond only with the post, never use hashtags.
    """,
)
@fast.chain(
    name="post_writer",
    sequence=["url_fetcher", "social_media"],
)

async def main():
    async with fast.run() as agent:
        # using chain workflow
        await agent.post_writer("http://llmindset.co.uk")

and I do not see how to get structured output from a chain.

@fast.agent(
    "url_fetcher",
    "Given a URL, provide a complete and comprehensive summary",
    servers=["fetch"], # Name of an MCP Server defined in fastagent.config.yaml
)
@fast.agent(
    "social_media",
    """
    Write a 280 character social media post for any given text.
    Respond only with the post, never use hashtags.
    """,
)
@fast.chain(
    name="post_writer",
    sequence=["url_fetcher", "social_media"],
)

async def main():
    async with fast.run() as agent:
        # using chain workflow
        await agent.post_writer("http://llmindset.co.uk").structured(model=SomeModel)

[ChrisW-priv]

Interesting... in that case, I do not know if it is possible or not. I guess you'd have to define some custom agent. Tho as a side note, I wonder what the use case would be since the llms would be perfectly fine with processing text as output... but I digress. Good luck with the development. If you figure it out, do not forget to post ;) .