How does Genkit determine which JSON spec to use?

[ariel-pettyjohn]

So, I really thought this was working as I expected before, but now I can't get enum/union types to validate. Can anyone give me a sanity check?

I have a simple flow that's supposed to identify the type of some text. For example, given the input true I would expect the response boolean.

I get a FAILED_PRECONDITION error though:

FAILED_PRECONDITION: Generation resulted in no candidates matching provided output schema.

Candidate[0] Error: INVALID_ARGUMENT: Schema validation failed. Parse Errors:

- (root): must be string

Provided data:

{
  "type": "string",
  "enum": [
    {
      "_def": {
        "value": "boolean",
        "typeName": "ZodLiteral"
      }
    }
  ]
}

Required JSON schema:

{
  "type": "string",
  "enum": [
    {
      "_def": {
        "value": "string",
        "typeName": "ZodLiteral"
      }
    },
    {
      "_def": {
        "value": "array",
        "typeName": "ZodLiteral"
      }
    },
    {
      "_def": {
        "value": "number",
        "typeName": "ZodLiteral"
      }
    },
    {
      "_def": {
        "value": "boolean",
        "typeName": "ZodLiteral"
      }
    },
    {
      "_def": {
        "value": "object",
        "typeName": "ZodLiteral"
      }
    }
  ],
  "$schema": "http://json-schema.org/draft-07/schema#"
}

It seems to have correctly identified "true" as boolean here, but still fails validation. Is this the expected behavior, because I expected it would validate for any string literal in the enum?

Thanks in advance!

[ariel-pettyjohn]

Ah-ha, I think this is just because I wrapped the string literals as Zod literals in the enum but I'm matching against a plain string.

I'll double-check tomorrow and update in case anyone else slips on this banana in the future.

[ariel-pettyjohn]

Okay, there were a few wrinkles here, but wrapping the string literals as Zod literals was only one layer of the problem:

From the Zod docs, it doesn't look like enums support Zod literals as objects, just string literals, numbers, etc.

Zod unions do require that string literals be wrapped as Zod literals however, which can become a block in case anyone stumbles on this in the future.

I think I see the real issue in my case though: because the model must return valid JSON, and because it seems to be using the older RFC 4627 specification, it can only return either an array or an object at the top level, not a string.

So, you can use a Zod array as your top-level output schema because an array is a valid top-level element of JSON, and I suppose you could use a union of array literals, but you can't use an enum or union of strings, numbers, etc.

I tried to instruct the model to use a different specification that allows strings as top-level elements, like ECMA-404, but it still produced the same results, so I'm guessing that it's using RFC 4627 by design?

Can anyone confirm? @mbleigh

Thanks in advance! 👍