Structured output from API using responseSchema - need help!

Gemini API
,
1

So I’m trying to translate some names from English into some Indian languages.

I’ve had success earlier with natural language prompting, defining the desired schema in the prompt itself and then parsing the response to get a json object.

But i wanted to experiment with using response schemas in my queries to the Gemini Pro API. Am not having much luck though.

So this is my code:

import google.generativeai as genai

genai.configure(api_key='XXXXX')
model = genai.GenerativeModel("gemini-1.5-pro-latest")

en_name_list = ["John", "Jane", "Doe"]

response_schema = {"type": "object", "properties": {}}
for name in en_name_list:
	response_schema["properties"][name] = {
		"type": "object",
		"properties": {
			"hi": {"type": "string"},
			"mr": {"type": "string"},
			"gu": {"type": "string"},
		},
	}

result = model.generate_content(
	f"Transliterate EACH of the following list of names into Hindi, Marathi and Gujarati: {en_name_list}",
	generation_config=genai.GenerationConfig(
		response_mime_type="application/json",
		response_schema=response_schema,
		temperature=0.1
	),
)

result.text

But i get this as output for result.text:
'{"John": {"gu": "જોન", "hi": "जॉन", "mr": "जॉन"} }\n'

So I only get the translations for one name in my output.

When i queried the Gemini Pro API about the issue, it had this to say:

The issue is likely not with your code directly, but with how the Gemini model is interpreting and responding to the prompt, even with the schema provided. While the schema defines the structure of the expected output, it doesn’t force the model to populate all fields or follow it perfectly. The model might decide to only process a subset of the input, especially if the prompt is ambiguous or if the temperature is very low.

How can i modify my code to achieve better results and translate all the names in the list?

3

Hi @Shijith_Kunhitty , You can try below code:

import typing_extensions as typing

class TranslateName(typing.TypedDict):
    HindiNames: list[str]
    MarathiNames: list[str]
    GujaratiNames: list[str]
     

model = genai.GenerativeModel("gemini-1.5-pro-latest")

en_name_list = ["John", "Jane", "Doe"]

result = model.generate_content(
	f"Transliterate EACH of the following list of names into Hindi, Marathi and Gujarati: {en_name_list}",
	generation_config=genai.GenerationConfig(
		response_mime_type="application/json",
		response_schema=TranslateName,
		temperature=0.1
	),
)

result.text
4

oh that’s ok @GUNAND_MAYANGLAMBAM , i resorted to defining the schema using natural language itself, instead of using responseSchema. Will try that some other time, thanks for the quick response!

This is the code i ended up using:

import google.generativeai as genai
import json

genai.configure(api_key='XXXX')
model = genai.GenerativeModel("gemini-1.5-pro-latest")

en_name_list = ["John", "Jane", "Doe"]

prompt = f"""
Example:
	Input: ['Alice', 'Bob']
	Output: {{'Alice': {{'hi': 'एलिस', 'mr': 'एलिस', 'gu': 'એલિસ'}}, 'Bob': {{'hi': 'बॉब', 'mr': 'बॉब', 'gu': 'બોબ'}}}}

So the desired output is a JSON object where:
- Each English name is a key
- The value for each key is another JSON object with 'hi', 'mr', and 'gu' keys containing the Hindi, Marathi and Gujarati transliterations

Return only a JSON object as output.

Now, transliterate EACH of the following names into Hindi, Marathi, and Gujarati, following the same format as the example: {en_name_list}
"""

result = model.generate_content(
		prompt,
		generation_config=genai.GenerationConfig(
			response_mime_type="application/json",
			temperature=0.1
		),
	)

json_data = json.loads(result.text.strip().replace('\n', ''))