LogSpanModelContent

Content type for standard LLM inference spans. All fields are optional.

Overview

LogSpanModelContent captures the input, output, and metadata of an LLM model call. It is wrapped in a LogSpanContent union via the actual_instance pattern. For the best observability experience, pass the raw provider request as input and the full provider response as output — Adaline will automatically extract cost, token usage, and model metadata.

from adaline_api.models.log_span_model_content import LogSpanModelContent

Fields

type

str | None

Must be "Model" when provided.

provider

str | None

The provider name (e.g., "openai", "anthropic"). 1–512 characters.

model

str | None

The model identifier (e.g., "gpt-4o", "claude-sonnet-4-20250514"). 1–512 characters.

input

str | None

The input payload as a JSON string. Must be valid, parseable JSON (the result of json.dumps()).

output

str | None

The output payload as a JSON string. Must be valid, parseable JSON (the result of json.dumps()).

variables

LogSpanVariable | None

Variable associated with this span for evaluation tracking. See LogSpanVariable.

cost

float | None

Cost of inference in USD. Overrides the automatic cost calculated by Adaline. Minimum: 0.

Construction Pattern

All span content is wrapped in LogSpanContent using the actual_instance parameter:

from adaline_api.models.log_span_content import LogSpanContent
from adaline_api.models.log_span_model_content import LogSpanModelContent

content = LogSpanContent(
    actual_instance=LogSpanModelContent(
        type="Model",
        provider="openai",
        model="gpt-4o",
        input=json.dumps(params),
        output=json.dumps(response.model_dump()),
    )
)

`input` and `output` Best Practices

Both input and output must be valid JSON strings. For Model spans, you get the most out of Adaline when you pass the exact request payload as input and the full provider response as output. This enables Adaline to automatically:

Calculate cost from token counts and model pricing
Extract token usage (prompt, completion, and total tokens)
Surface metadata such as stop reason and tool calls
Power continuous evaluations with structured I/O

For a deeper walkthrough of this pattern across providers, see Span content: input and output.

Examples

OpenAI
Anthropic

import json
from openai import OpenAI
from adaline_api.models.log_span_content import LogSpanContent
from adaline_api.models.log_span_model_content import LogSpanModelContent

openai = OpenAI()

params = {
    "model": "gpt-4o",
    "messages": [
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Explain quantum computing simply."},
    ],
    "temperature": 0.7,
}

response = openai.chat.completions.create(**params)

span.update({
    "status": "success",
    "content": LogSpanContent(
        actual_instance=LogSpanModelContent(
            type="Model",
            provider="openai",
            model="gpt-4o",
            input=json.dumps(params),
            output=json.dumps(response.model_dump()),
        )
    ),
})

import json
import anthropic
from adaline_api.models.log_span_content import LogSpanContent
from adaline_api.models.log_span_model_content import LogSpanModelContent

client = anthropic.Anthropic()

params = {
    "model": "claude-sonnet-4-20250514",
    "max_tokens": 1024,
    "messages": [
        {"role": "user", "content": "Explain quantum computing simply."},
    ],
}

response = client.messages.create(**params)

span.update({
    "status": "success",
    "content": LogSpanContent(
        actual_instance=LogSpanModelContent(
            type="Model",
            provider="anthropic",
            model="claude-sonnet-4-20250514",
            input=json.dumps(params),
            output=json.dumps(response.model_dump()),
        )
    ),
})

With Variables

import json
from adaline_api.models.log_span_content import LogSpanContent
from adaline_api.models.log_span_model_content import LogSpanModelContent
from adaline_api.models.log_span_variable import LogSpanVariable
from adaline_api.models.text_content import TextContent

content = LogSpanContent(
    actual_instance=LogSpanModelContent(
        type="Model",
        provider="openai",
        model="gpt-4o",
        input=json.dumps(params),
        output=json.dumps(response.model_dump()),
        variables=LogSpanVariable(
            name="user_question",
            value=TextContent(modality="text", value="What is quantum computing?")
        ),
    )
)

With Explicit Cost

content = LogSpanContent(
    actual_instance=LogSpanModelContent(
        type="Model",
        provider="openai",
        model="gpt-4o",
        input=json.dumps(params),
        output=json.dumps(response.model_dump()),
        cost=0.0032,
    )
)

Avoid cherry-picking or reshaping the request/response before serializing. Pass the raw objects — Adaline’s automatic parsing depends on seeing the provider’s native schema.

Serialization

from adaline_api.models.log_span_model_content import LogSpanModelContent

model_content = LogSpanModelContent(
    type="Model",
    provider="openai",
    model="gpt-4o",
    input='{"model":"gpt-4o","messages":[]}',
    output='{"choices":[]}',
)

d = model_content.to_dict()
j = model_content.to_json()

restored = LogSpanModelContent.from_dict(d)
restored = LogSpanModelContent.from_json(j)

Get started

API

SDKs

Proxy

LogSpanModelContent

LogSpanModelContent

Overview

Fields

Construction Pattern

`input` and `output` Best Practices

Examples

With Variables

With Explicit Cost

Serialization

​LogSpanModelContent

​Overview

​Fields

​Construction Pattern

​input and output Best Practices

​Examples

​With Variables

​With Explicit Cost

​Serialization

LogSpanModelContent

Overview

Fields

Construction Pattern

`input` and `output` Best Practices

Examples

With Variables

With Explicit Cost

Serialization