JSON Output

Enforcing structured JSON schema output is important for handling LLM outputs downstream with other systems and APIs in your applications.

For an in-depth technical deep dive of how we implemented this feature, see our blog post.

Python SDKREST API

You can enforce JSON schema via the Lamini class is the base class for all runners. Lamini wraps our REST API endpoint.

First, return a string:

from lamini import Lamini

llm = Lamini(model_name="meta-llama/Meta-Llama-3-8B-Instruct")
output = llm.generate(
    "How are you?",
    output_type={"answer": "str"}
)

First, get a basic string output out:

curl --location "https://api.lamini.ai/v1/completions" \
--header "Authorization: Bearer $LAMINI_API_KEY" \
--header "Content-Type: application/json" \
--data '{
    "model_name": "meta-llama/Meta-Llama-3-8B-Instruct",
    "prompt": "How are you?",
    "output_type": {
        "answer": "str"
    }
}'

Expected Output

{
    "answer":"I'm doing well, thanks for asking! How about you"
}

Values other than strings

You can change the output type to be a different type, e.g. int or float. This typing is strictly enforced.

Please let us know if there are specific types you'd like to see supported.

Python SDKREST API

llm.generate(
    "How old are you?",
    output_type={"age": "int"}
)

curl --location "https://api.lamini.ai/v1/completions" \
--header "Authorization: Bearer $LAMINI_API_KEY" \
--header "Content-Type: application/json" \
--data '{
    "model_name": "meta-llama/Meta-Llama-3-8B-Instruct",
    "prompt": "How old are you?",
    "output_type": {
        "age": "int"
    }
}'

Expected Output

{
    "age": 25
}

Multiple outputs in JSON schema

You can also add multiple output types in one call. The output is a JSON schema that is also strictly enforced.

Python SDKREST API

llm.generate(
    "How old are you?",
    output_type={"age": "int", "units": "str"}
)

curl --location "https://api.lamini.ai/v1/completions" \
--header "Authorization: Bearer $LAMINI_API_KEY" \
--header "Content-Type: application/json" \
--data '{
    "model_name": "meta-llama/Meta-Llama-3-8B-Instruct",
    "prompt": "How old are you?",
    "output_type": {
        "age": "int",
        "units": "str"
    }
}'

Expected Output

{
    "age": 25,
    "units": "years"
}

Great! You've successfully run an LLM with structured JSON schema outputs.

Known issue: Long JSON output times out

To ensure that requests complete in a reasonable amount of time, there is a time limit on all requests including json requests. If your requests exceeds the time limit, try guiding the model to generate a shorter json object, e.g. write a description in 3 sentences or less. Timed out requests may result in failed, incomplete, or missing output.

Workaround

• Reduce the size of the output by limiting the number of fields or the prompt.
• Break down the JSON output into separate smaller requests.
• Contact us to discuss alternative solutions or workarounds for your use case.

Future support

We are evaluating the feasibility of improving our system to handle large JSON output in the future. If we decide to support this feature, we will update our documentation and notify users.

Feel free to contact us with any questions or concerns.