Log attachments

Variables on spans aren’t limited to plain text. You can attach images, PDFs, and large text summaries as variable values, making them available in the Playground, in datasets, and for continuous evaluations.

URL vs base64

Attachments support two modes — URL and base64 — and the distinction matters:

Mode	What happens	Usable in Playground & datasets
URL	Adaline stores the link. The file must remain publicly accessible.	Only while the URL is live.
base64	Adaline receives the raw data and hosts it in its own storage.	Yes — fully persisted and self-contained.

Images

Attach images to spans as variables with modality: "image". The detail field controls resolution processing (auto, low, medium, high).

Base64

TypeScript
Python
API

import { readFileSync } from "fs";

const imageBase64 = readFileSync("./product-screenshot.png").toString("base64");

const span = trace.logSpan({ name: "analyze-image" });

span.update({
  status: "success",
  content: {
    type: "Model",
    provider: "openai",
    model: "gpt-4o",
    variables: {
      product_image: {
        name: "product_image",
        value: {
          modality: "image",
          detail: "auto",
          value: { type: "base64", base64: imageBase64, mediaType: "png" },
        },
      },
    },
    input: JSON.stringify(params),
    output: JSON.stringify(response),
  },
});

import base64
from adaline_api.models.log_span_content import LogSpanContent
from adaline_api.models.log_span_model_content import LogSpanModelContent

with open("./product-screenshot.png", "rb") as f:
    image_base64 = base64.b64encode(f.read()).decode("utf-8")

span = trace.log_span(name="analyze-image")

span.update({
    "status": "success",
    "content": LogSpanContent(
        actual_instance=LogSpanModelContent(
            type="Model",
            provider="openai",
            model="gpt-4o",
            variables={
                "product_image": {
                    "modality": "image",
                    "detail": "auto",
                    "value": {"type": "base64", "base64": image_base64, "mediaType": "png"},
                },
            },
            input=json.dumps(params),
            output=json.dumps(response),
        )
    ),
})

curl -X POST https://api.adaline.ai/v2/logs/span \
  -H "Authorization: Bearer $ADALINE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "projectId": "your-project-id",
    "traceReferenceId": "your-trace-ref",
    "span": {
      "name": "analyze-image",
      "status": "success",
      "referenceId": "span-img-001",
      "startedAt": 1700000000000,
      "endedAt": 1700000002000,
      "content": { "type": "Other", "input": "{}", "output": "{}" },
      "variables": {
        "product_image": {
          "modality": "image",
          "detail": "auto",
          "value": { "type": "base64", "base64": "<base64-encoded-data>", "mediaType": "png" }
        }
      }
    }
  }'

Supported media types: png, jpeg, webp, gif.

URL

When the image is already hosted publicly, you can pass a URL. Adaline stores the link but does not download or host the file.

TypeScript
Python

const span = trace.logSpan({ name: "analyze-image" });

span.update({
  status: "success",
  content: {
    type: "Model",
    provider: "openai",
    model: "gpt-4o",
    variables: {
      product_image: {
        name: "product_image",
        value: {
          modality: "image",
          detail: "auto",
          value: { type: "url", url: "https://cdn.example.com/product.png" },
        },
      },
    },
    input: JSON.stringify(params),
    output: JSON.stringify(response),
  },
});

span = trace.log_span(name="analyze-image")

span.update({
    "status": "success",
    "content": LogSpanContent(
        actual_instance=LogSpanModelContent(
            type="Model",
            provider="openai",
            model="gpt-4o",
            variables={
                "product_image": {
                    "modality": "image",
                    "detail": "auto",
                    "value": {"type": "url", "url": "https://cdn.example.com/product.png"},
                },
            },
            input=json.dumps(params),
            output=json.dumps(response),
        )
    ),
})

PDFs

Attach PDF documents with modality: "pdf". PDFs include a file object with metadata (name, id, and optionally size).

Base64

TypeScript
Python
API

import { readFileSync } from "fs";

const pdfBase64 = readFileSync("./invoice.pdf").toString("base64");

const span = trace.logSpan({ name: "process-document" });

span.update({
  status: "success",
  content: {
    type: "Model",
    provider: "openai",
    model: "gpt-4o",
    variables: {
      invoice: {
        name: "invoice",
        value: {
          modality: "pdf",
          value: { type: "base64", base64: pdfBase64 },
          file: { name: "invoice.pdf", id: "doc-001" },
        },
      },
    },
    input: JSON.stringify(params),
    output: JSON.stringify(response),
  },
});

import base64

with open("./invoice.pdf", "rb") as f:
    pdf_base64 = base64.b64encode(f.read()).decode("utf-8")

span = trace.log_span(name="process-document")

span.update({
    "status": "success",
    "content": LogSpanContent(
        actual_instance=LogSpanModelContent(
            type="Model",
            provider="openai",
            model="gpt-4o",
            variables={
                "invoice": {
                    "modality": "pdf",
                    "value": {"type": "base64", "base64": pdf_base64},
                    "file": {"name": "invoice.pdf", "id": "doc-001"},
                },
            },
            input=json.dumps(params),
            output=json.dumps(response),
        )
    ),
})

curl -X POST https://api.adaline.ai/v2/logs/span \
  -H "Authorization: Bearer $ADALINE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "projectId": "your-project-id",
    "traceReferenceId": "your-trace-ref",
    "span": {
      "name": "process-document",
      "status": "success",
      "referenceId": "span-pdf-001",
      "startedAt": 1700000000000,
      "endedAt": 1700000003000,
      "content": { "type": "Other", "input": "{}", "output": "{}" },
      "variables": {
        "invoice": {
          "modality": "pdf",
          "value": { "type": "base64", "base64": "<base64-encoded-data>" },
          "file": { "name": "invoice.pdf", "id": "doc-001" }
        }
      }
    }
  }'

URL

TypeScript
Python

const span = trace.logSpan({ name: "process-document" });

span.update({
  status: "success",
  content: {
    type: "Model",
    provider: "openai",
    model: "gpt-4o",
    variables: {
      invoice: {
        name: "invoice",
        value: {
          modality: "pdf",
          value: { type: "url", url: "https://cdn.example.com/invoice.pdf" },
          file: { name: "invoice.pdf", id: "doc-001" },
        },
      },
    },
    input: JSON.stringify(params),
    output: JSON.stringify(response),
  },
});

span = trace.log_span(name="process-document")

span.update({
    "status": "success",
    "content": LogSpanContent(
        actual_instance=LogSpanModelContent(
            type="Model",
            provider="openai",
            model="gpt-4o",
            variables={
                "invoice": {
                    "modality": "pdf",
                    "value": {"type": "url", "url": "https://cdn.example.com/invoice.pdf"},
                    "file": {"name": "invoice.pdf", "id": "doc-001"},
                },
            },
            input=json.dumps(params),
            output=json.dumps(response),
        )
    ),
})

Text

For large text — retrieved context, summaries, full documents — use text variables or attributes. Variables are the better choice when the text needs to flow into datasets and evaluations. Attributes work well for shorter metadata you want to filter and search by in the Monitor.

Via variables

TypeScript
Python

const span = trace.logSpan({ name: "rag-response" });

span.update({
  status: "success",
  content: {
    type: "Model",
    provider: "openai",
    model: "gpt-4o",
    variables: {
      user_question: { name: "user_question", value: { modality: "text", value: userQuery } },
      retrieved_context: { name: "retrieved_context", value: { modality: "text", value: longContextString } },
      system_prompt: { name: "system_prompt", value: { modality: "text", value: systemPromptText } },
    },
    input: JSON.stringify(params),
    output: JSON.stringify(response),
  },
});

span = trace.log_span(name="rag-response")

span.update({
    "status": "success",
    "content": LogSpanContent(
        actual_instance=LogSpanModelContent(
            type="Model",
            provider="openai",
            model="gpt-4o",
            variables={
                "user_question": {"modality": "text", "value": user_query},
                "retrieved_context": {"modality": "text", "value": long_context_string},
                "system_prompt": {"modality": "text", "value": system_prompt_text},
            },
            input=json.dumps(params),
            output=json.dumps(response),
        )
    ),
})

Via attributes

TypeScript
Python

const span = trace.logSpan({
  name: "rag-response",
  attributes: {
    user_question: userQuery,
    retrieved_context: longContextString,
    summary: documentSummary,
  },
});

span = trace.log_span(
    name="rag-response",
    attributes={
        "user_question": user_query,
        "retrieved_context": long_context_string,
        "summary": document_summary,
    },
)

Proxy headers

When using the Proxy, pass variables (including attachments) via the adaline-span-variables header:

headers["adaline-span-variables"] = json.dumps({
    "product_image": {
        "modality": "image",
        "detail": "auto",
        "value": {"type": "base64", "base64": image_base64, "mediaType": "png"},
    },
    "user_question": {
        "modality": "text",
        "value": "What color options are available?",
    },
})

Limits

Attachment type	Max size
Image (base64)	10 MB
PDF (base64)	10 MB
Request body (total)	32 MB

See Limits for full payload constraints. Requests exceeding these limits receive a 413 Payload Too Large response.

If your files exceed these limits, you can either self-host the media and use URL referencing instead of base64, or contact support@adaline.ai to discuss higher limits for your workspace.

Get started

Instrument

Monitor

Iterate

Evaluate

Deploy

Admin

Others

Log attachments

URL vs base64

Images

Base64

URL

PDFs

Base64

URL

Text

Via variables

Via attributes

Proxy headers

Limits

Next steps

Analyze Log Traces

Build Datasets from Logs

Get started

Instrument

Monitor

Iterate

Evaluate

Deploy

Admin

Others

Documentation Index

​URL vs base64

​Images

​Base64

​URL

​PDFs

​Base64

​URL

​Text

​Via variables

​Via attributes

​Proxy headers

​Limits

​Next steps

Analyze Log Traces

Build Datasets from Logs

URL vs base64

Images

Base64

URL

PDFs

Base64

URL

Text

Via variables

Via attributes

Proxy headers

Limits

Next steps