Before diving into this content, it would be helpful to read the following guides:
The following features are available in the following SDK versions:
  • Python SDK: >=0.1.141
  • JS/TS SDK: >=0.2.5
LangSmith supports uploading binary files (such as images, audio, videos, PDFs, and CSVs) with your traces. This is particularly useful when working with LLM pipelines using multimodal inputs or outputs. In both the Python and TypeScript SDKs, attachments can be added to your traces by specifying the MIME type and binary content of each file. This guide explains how to define and trace attachments using the Attachment type in Python and Uint8Array / ArrayBuffer in TypeScript.
In the Python SDK, you can use the Attachment type to add files to your traces. Each Attachment requires:
  • mime_type (str): The MIME type of the file (e.g., "image/png").
  • data (bytes | Path): The binary content of the file, or the file path.
You can also define an attachment with a tuple tuple of the form (mime_type, data) for convenience.Simply decorate a function with @traceable and include your Attachment instances as arguments. Note that to use the file path instead of the raw bytes, you need to set the dangerously_allow_filesystem flag to True in your traceable decorator.
from langsmith import traceable
from langsmith.schemas import Attachment
from pathlib import Path
import os

# Must set dangerously_allow_filesystem to True if you want to use file paths
@traceable(dangerously_allow_filesystem=True)
def trace_with_attachments(
    val: int,
    text: str,
    image: Attachment,
    audio: Attachment,
    video: Attachment,
    pdf: Attachment,
    csv: Attachment,
):
    return f"Processed: {val}, {text}, {len(image.data)}, {len(audio.data)}, {len(video.data)}, {len(pdf.data), {len(csv.data)}}"

# Helper function to load files as bytes
def load_file(file_path: str) -> bytes:
    with open(file_path, "rb") as f:
        return f.read()

# Load files and create attachments
image_data = load_file("my_image.png")
audio_data = load_file("my_mp3.mp3")
video_data = load_file("my_video.mp4")
pdf_data = load_file("my_document.pdf")

image_attachment = Attachment(mime_type="image/png", data=image_data)
audio_attachment = Attachment(mime_type="audio/mpeg", data=audio_data)
video_attachment = Attachment(mime_type="video/mp4", data=video_data)
pdf_attachment = ("application/pdf", pdf_data) # Can just define as tuple of (mime_type, data)
csv_attachment = Attachment(mime_type="text/csv", data=Path(os.getcwd()) / "my_csv.csv")

# Define other parameters
val = 42
text = "Hello, world!"

# Call the function with traced attachments
result = trace_with_attachments(
    val=val,
    text=text,
    image=image_attachment,
    audio=audio_attachment,
    video=video_attachment,
    pdf=pdf_attachment,
    csv=csv_attachment,
)
Here is how the above would look in the LangSmith UI. You can expand each attachment to view its contents.