HoneyHive’s tracing capabilities extend beyond text-based data, allowing you to capture and analyze multi-modal information in your AI applications. This guide focuses on instrumenting functions that handle multi-modal data, particularly those that return S3 URLs pointing to images, audio, or other non-text assets.

​
Why Multi-modal Tracing?

Multi-modal tracing is crucial for applications that process various types of data, such as:
  • Image generation or analysis
  • Audio processing
  • Video content creation or analysis
  • Document processing with embedded media
By tracing these functions, you can gain insights into how your application handles different data types and how they impact your AI pipeline’s performance and accuracy.

​
Using the trace Decorator for Multi-modal Data

To instrument functions that return S3 URLs for multi-modal data, you’ll use the same trace decorator as with text-based functions. Here’s how to set it up:
  1. First, ensure you’ve initialized the HoneyHiveTracer:
Python
from honeyhive import HoneyHiveTracer

HoneyHiveTracer.init(
    api_key=MY_HONEYHIVE_API_KEY,
    project=MY_HONEYHIVE_PROJECT_NAME,
    source=MY_SOURCE,  # e.g., "prod", "dev", etc.
    session_name=MY_SESSION_NAME,
)
  1. Import and use the trace decorator:
Python
from honeyhive import trace

@trace
def process_image(image_path):
    # Image processing logic here
    # ...
    return "s3://my-bucket/processed-images/image123.jpg"

​
Adding Context to Multi-modal Traces

To make your traces more informative, you can add metadata about the multi-modal data:
Python
@trace(
    metadata={
        "data_type": "image",
        "format": "jpg",
        "resolution": "1024x768",
        "processing_steps": ["resize", "enhance", "annotate"]
    }
)
def process_image(image_path):
    # Image processing logic here
    # ...
    return "s3://my-bucket/processed-images/image123.jpg"

​
Handling Different Multi-modal Types

Here are examples of tracing different types of multi-modal data:

​
Audio Processing

Python
@trace(
    metadata={
        "data_type": "audio",
        "format": "wav",
        "duration_seconds": 120,
        "sample_rate": 44100
    }
)
def transcribe_audio(audio_file):
    # Audio transcription logic
    # ...
    return "s3://my-bucket/transcriptions/audio123.txt"

​
Video Analysis

Python
@trace(
    metadata={
        "data_type": "video",
        "format": "mp4",
        "duration_seconds": 300,
        "resolution": "1920x1080",
        "fps": 30
    }
)
def analyze_video(video_file):
    # Video analysis logic
    # ...
    return "s3://my-bucket/video-analysis/video123.json"

​
Best Practices for Multi-modal Tracing

  1. Include relevant metadata: Add information about the data type, format, size, and any processing steps to provide context.
  2. Use consistent naming conventions: For S3 URLs, use a consistent structure to make it easier to analyze and group related assets.
  3. Consider privacy and data protection: Ensure that your S3 URLs and metadata don’t contain sensitive information.
  4. Link related traces: If a multi-modal process involves multiple steps, use consistent identifiers in your metadata to link related traces.