Adding output parsers to spaces

[itsyogesh]

There is a lot of work that's happening around llm outputs. Some interesting things we can look at:

https://hwchase17.github.io/langchainjs/docs/modules/prompts/output_parsers/

https://github.com/shreyar/guardrails

https://eyurtsev.github.io/kor/index.html

We can look at specific implementations that are described here or we can build one from scratch.

[dosco]

Thanks this looks cool i'll dig in deeper. There are two built in output parsers in GenerateText both are used based on Prompt metadata for example. the param keyValueResponse: true enables splitting the response using a colon seperator.

1. Key Value Output:
   https://github.com/dosco/minds/blob/50e864f91806c2d723fe6edd46bc3a6a2ecd17d8/src/prompts/extract.ts#L75

2. Tools / Actions but this one is used internally:
   https://github.com/dosco/minds/blob/50e864f91806c2d723fe6edd46bc3a6a2ecd17d8/src/prompts/answers.ts#L46

I tend to add features based on uses-cases so will try and figure out some to extending our current output parsing the zod librasry that langchain uses can be used by us as well.

[dosco]

just finished a big refactor to make the api simpler to use and extend. also added support for zod. we now have output parser support for plaintext, simple key-value or a zod schema built in.

The prompts are now super-simple and mostly extend from AIPrompt which does all the heavy lifting. For example see the
https://github.com/dosco/minds/blob/main/src/prompts/zprompt.ts

const prompt = new ZPrompt(zodSchema, actions);

const CustomerResponse = z.object({
  data: z
    .array(
      z.object({
        name: z.string().describe('product name'),
        units: z.number().describe('units in stock'),
        desc: z.string().max(15).describe('product description'),
      })
    )
    .describe('inventory information'),
  response: z.string().max(50).describe('customer response'),
});

const prompt = new ZPrompt(CustomerResponse, actions);

[Madd0g]

Are ZPrompt or zod (or grammars) supported currently? I couldn't find anything

[dosco]

The whole framework has changed significantly and is now based on the prompt as. a program idea from the Stanford DSP paper. while zod is not supported anymore it's not required you can have typed inputs and outputs on the prompt (prompt signature)

[dosco]

And you can use asserts to integrate with zod if you still need to.