Approach to Saving an Agent Output to the File System

[pl-shernandez]

I'm having fun learning how to use agents more effectively -- but struggling with how to save their structured response to the file system by the top level orchestrating prompt that receives the response or let the called agent itself save the response json data generated itself. I just can't get it to save...

I am trying to make something inspired by this but instead as a fake companies product engineering organization.

My agent is defined like this
I see it is interacting with the file system and I have other agents that read successfully from a markdown file in that same folder that acts as a "company wiki".

I'm using the documentation for System Prompts as a reference.

Is defFileOutput the correct approach? Are agents allowed to perform saves? Should defOutputProcessor( be used instead?

[pl-shernandez]

Ok after tinkering this seems like a reliable approach I'll adopt for what I'm working on. I'm curious still the best approach to instruct the LLM to reliably save a file with just 'then save a file to...' instructions in the text of a prompt using runPrompt( It seems defFileOutput is supposed to hint it where to put it.

script({
  title: 'Knowledge Base Buddy',
  model: 'openai:gpt-4o-mini',
})
let message = await host.input('Message:')
while (message.trim() !== '/exit') {
  await runPrompt(
    (ctx) => {
      ctx.$`
        Task:
        Create a knowledge entry for the given message.
        
        Message:
        <MESSAGE>
      `
      ctx.def('MESSAGE', message)
      ctx.defOutputProcessor((out) => {
        const filename = Math.random().toString(36).slice(-6)
        workspace.writeText(`${filename}.json`, out.text)
      })
    },
    {
      responseType: 'json',
      responseSchema: {
        title: 'Knowledge',
        description: 'A single piece of knowledge stored in the knowledge base.',
        type: 'object',
        properties: {
          id: {
            type: 'string',
            description: 'A unique identifier for this knowledge item.',
          },
          title: {
            type: 'string',
            description: 'A concise, descriptive title.',
          },
          body: {
            type: 'string',
            description: 'The main text or description of the knowledge.',
          },
          tags: {
            type: 'array',
            items: { type: 'string' },
            description: 'A list of tags or categories for easy organization.',
          },
          createdAt: {
            type: 'string',
            description: 'When this knowledge item was created.',
          },
          updatedAt: {
            type: 'string',
            description: 'When this knowledge item was last updated.',
          },
          source: {
            type: 'string',
            description: 'Where the information came from (optional).',
          },
        },
        required: ['id', 'title', 'body', 'createdAt'],
        additionalProperties: false,
      },
    }
  )
  message = await host.input('Message:')
}

[pelikhan]

The idea of "defFileOutput" is to define "routes" where it is acceptable for genaiscript to edit/update files - without confirmation from the user. I need to think about this request and come back to you.