​
Extracting Structured Data from Cooking Videos: A Hands-On Guide

​
Introduction

Video content contains a wealth of information that, when properly structured, can power advanced search, analytics, and insights. In this guide, we’ll walk through extracting structured data from cooking videos using Cloudglue’s entity extraction capabilities.

By the end, you’ll learn how to:

  • Create an entity collection for cooking videos
  • Define a structured schema for recipe information
  • Extract detailed data from YouTube cooking videos
  • Analyze and visualize patterns across multiple videos

​
Prerequisites

# Install the Cloudglue Python SDK
pip install cloudglue

# Import required libraries
import os
from cloudglue import CloudGlue
import pandas as pd
import matplotlib.pyplot as plt
import numpy as np
from datetime import timedelta

# Initialize the Cloudglue client
client = CloudGlue(api_key=os.environ.get("CLOUDGLUE_API_KEY"))

​
Creating a Recipe Entity Collection

First, we’ll create a collection specifically for cooking videos with a schema that captures recipe details, equipment, actions, ingredients, and cooking phases.

# Define our extraction schema - simplified to include only what we need for analysis
schema = {
  "recipe": {
    "chef_name": "string",
    "dish_name": "string",
    "cuisine_type": "Italian|Mexican|Asian|French|American|Mediterranean|Indian|Thai|Chinese|Japanese|Other",
    "meal_category": "breakfast|lunch|dinner|snack|dessert|appetizer|side_dish"
  },
  "equipment_mentioned": ["string"],
  "cooking_actions": ["string"],
  "ingredients": ["string"],
  "cooking_phase": "prep|active_cooking|plating|tasting|explanation|cleanup|intro|outro"
}

# Define our extraction prompt - with precise field mappings
prompt = """
Extract cooking information from this recipe video transcript using these exact field names:

1. RECIPE (populate "recipe" object):
   - chef_name: Identify the chef's name
   - dish_name: Name of the dish being prepared
   - cuisine_type: Choose one from: Italian, Mexican, Asian, French, American, Mediterranean, Indian, Thai, Chinese, Japanese, Other
   - meal_category: Choose one from: breakfast, lunch, dinner, snack, dessert, appetizer, side_dish

2. EQUIPMENT_MENTIONED (populate "equipment_mentioned" array)

3. COOKING_ACTIONS (populate "cooking_actions" array):
   - Example specific cooking action (e.g., chopping, stirring, mixing, baking, frying, etc.)

4. INGREDIENTS (populate "ingredients" array)

5. COOKING_PHASE (populate "cooking_phase" field):
   - Classify the current segment as one of: prep, active_cooking, plating, tasting, explanation, cleanup, intro, outro

Focus on extracting information exactly as spoken in the transcript.
"""

# Create a collection for recipe videos
collection = client.collections.create(
    name="Cooking Videos Analysis",
    collection_type="entities",
    description="Collection of cooking videos for recipe analysis",
    extract_config={
        "schema": schema,
        "prompt": prompt
    }
)

​
Adding YouTube Cooking Videos

Now, let’s add 10 different cooking videos to our collection. We’ll choose a variety of cuisines and meal types.

# List of YouTube video URLs for different cuisines
youtube_urls = [
    # Replace with urls to youtube videos with different recipe types (or cloudglue:// uploads)
    "https://www.youtube.com/watch?v=VIDEO_ID1",
    "https://www.youtube.com/watch?v=VIDEO_ID2",
    # etc ...
]

# Add each video to the collection
file_ids = []
for url in youtube_urls:
    try:
        # Add YouTube video to collection
        result = client.collections.add_youtube_video(collection.id, url=url)
        file_ids.append(result.file_id)
        print(f"Added video: {url} with file ID: {result.file_id}")
    except Exception as e:
        print(f"Error adding {url}: {str(e)}")

​
Waiting for Processing and Retrieving Entities

The extraction process runs asynchronously. Let’s monitor and wait for the extraction to complete.

import time

# Function to check if all videos are processed
def all_videos_processed(collection_id, file_ids):
    processed_count = 0
    for file_id in file_ids:
        try:
            file_info = client.collections.get_video(collection_id, file_id)
            if file_info.status == "completed":
                # Verifying entities record exists and is ready for video, otherwise 404
                entities = client.collections.get_video_entities(collection_id, file_id)
                processed_count += 1
        except Exception as e:
            print(f"Error checking status for file {file_id}: {str(e)}")

    return processed_count, len(file_ids)

# Check status every 15 seconds
while True:
    processed, total = all_videos_processed(collection.id, file_ids)
    print(f"Processed {processed} of {total} videos")

    if processed == total:
        print("All videos processed!")
        break

    print("Waiting 15 seconds before checking again...")
    time.sleep(15)

​
Extracting and Aggregating the Data

Now that our videos are processed, let’s extract the entities and organize them for analysis.

# Get all the entity data
video_data = []
segment_data = []

for file_id in file_ids:
    # try:
        # Get entities for this video
    entities = client.collections.get_video_entities(collection.id, file_id)

    # Get file info to get the video title
    file_info = client.collections.get_video(collection.id, file_id)

    # Extract video-level entities
    video_info = {
        'file_id': file_id,
        'filename': file_info.file.filename,
        'chef_name': entities.entities.get('recipe', {}).get('chef_name', 'Unknown'),
        'dish_name': entities.entities.get('recipe', {}).get('dish_name', 'Unknown'),
        'cuisine_type': entities.entities.get('recipe', {}).get('cuisine_type', 'Unknown'),
        'meal_category': entities.entities.get('recipe', {}).get('meal_category', 'Unknown'),
        'ingredient_count': len(entities.entities.get('ingredients', [])),
        'equipment_count': len(entities.entities.get('equipment_mentioned', [])),
    }
    video_data.append(video_info)

    # Extract segment-level entities
    for segment in entities.segment_entities:
        segment_info = {
            'file_id': file_id,
            'filename': file_info.file.filename,
            'start_time': segment.start_time,
            'end_time': segment.end_time,
            'duration': segment.end_time - segment.start_time,
            'cooking_phase': segment.entities.get('cooking_phase', 'Unknown'),
            'action_count': len(segment.entities.get('cooking_actions', [])),
            'ingredients_mentioned': len(segment.entities.get('ingredients', [])),
            'equipment_used': len(segment.entities.get('equipment_mentioned', []))
        }
        segment_data.append(segment_info)

    print(f"Extracted data for {file_info.file.filename}")

# Convert to pandas DataFrames
videos_df = pd.DataFrame(video_data)
segments_df = pd.DataFrame(segment_data)

​
Analyzing the Data

​
1. Ingredient Count Comparison by Cuisine Type

Let’s compare the average number of ingredients used in each cuisine type.

# Group by cuisine_type and calculate average ingredient count
cuisine_ingredients = videos_df.groupby('cuisine_type').agg({
    'ingredient_count': 'mean',
    'file_id': 'count'
}).reset_index()

# Rename columns for clarity
cuisine_ingredients.columns = ['Cuisine Type', 'Avg Ingredients', 'Videos']

# Print table
print("Ingredient Count by Cuisine Type:")
print(cuisine_ingredients.to_string(index=False, formatters={
    'Avg Ingredients': '{:.1f}'.format
}))

# Create a bar chart
plt.figure(figsize=(10, 6))
plt.bar(cuisine_ingredients['Cuisine Type'], cuisine_ingredients['Avg Ingredients'], color='skyblue')
plt.title('Average Number of Ingredients by Cuisine Type')
plt.xlabel('Cuisine Type')
plt.ylabel('Average Ingredient Count')
plt.xticks(rotation=45)

# Add video count as text above each bar - using reliable column access
for i, (_, row) in enumerate(cuisine_ingredients.iterrows()):
    plt.text(i, row['Avg Ingredients'] + 0.3, f"{row['Videos']} videos", ha='center')

plt.tight_layout()
plt.savefig('ingredient_comparison.png')
plt.show()

# Print insight
if len(cuisine_ingredients) >= 2:
    max_cuisine = cuisine_ingredients.loc[cuisine_ingredients['Avg Ingredients'].idxmax()]
    min_cuisine = cuisine_ingredients.loc[cuisine_ingredients['Avg Ingredients'].idxmin()]

    percentage_diff = ((max_cuisine['Avg Ingredients'] - min_cuisine['Avg Ingredients']) /
                      min_cuisine['Avg Ingredients']) * 100

    print(f"\nINSIGHT: {max_cuisine['Cuisine Type']} recipes in our sample need {percentage_diff:.0f}% more ingredients "
          f"than {min_cuisine['Cuisine Type']} ones")

Ingredient Count by Cuisine Type:
Cuisine Type Avg Ingredients  Videos
       Asian            16.0       1
     Italian            16.0       2
    Japanese            19.0       2
     Mexican            22.0       3
       Other            26.0       1
        Thai            23.0       1

INSIGHT: Other recipes in our sample need 62% more ingredients than Asian ones

​
2. Phase Duration Analysis

Let’s analyze how much time each video spends in different cooking phases.

# 1. Calculate duration by phase for each video
phase_duration = segments_df.groupby(['file_id', 'filename', 'cooking_phase'])['duration'].sum().reset_index()

# 2. Calculate total duration for each video
video_total_duration = segments_df.groupby(['file_id', 'filename'])['duration'].sum().reset_index()

# 3. Merge and calculate percentages
phase_percentage = pd.merge(phase_duration, video_total_duration, on=['file_id', 'filename'])
phase_percentage['percentage'] = (phase_percentage['duration_x'] / phase_percentage['duration_y']) * 100

# 4. Create pivot table
phase_pivot = phase_percentage.pivot_table(
    index=['file_id', 'filename'],
    columns='cooking_phase',
    values='percentage',
    fill_value=0
).reset_index()

# 5. Prepare for visualization
main_phases = ['prep', 'active_cooking', 'plating', 'tasting', 'explanation']
available_phases = [col for col in phase_pivot.columns if col in main_phases]

# Create a stacked bar chart
plt.figure(figsize=(14, 8))

# Sample 3 videos for clearer visualization
if len(phase_pivot) > 3:
    sample_videos = phase_pivot.sample(n=3)
else:
    sample_videos = phase_pivot

# Create shorter names for display
sample_videos['short_name'] = sample_videos['filename'].str.slice(0, 30) + '...'

# Plot stacked bars
bottom = np.zeros(len(sample_videos))
for phase in available_phases:
    plt.barh(sample_videos['short_name'], sample_videos[phase], left=bottom, label=phase)
    bottom += sample_videos[phase]

# Add percentage text inside bars
for i, (_, row) in enumerate(sample_videos.iterrows()):
    x_pos = 0
    for phase in available_phases:
        if row[phase] > 5:  # Only show if > 5%
            width = row[phase]
            plt.text(x_pos + width/2, i, f"{int(width)}%",
                     ha='center', va='center', color='white', fontweight='bold')
            x_pos += width

plt.title('Cooking Phase Distribution by Video (Sample)')
plt.xlabel('Percentage of Video Duration')
plt.legend(title='Cooking Phase')
plt.tight_layout()
plt.savefig('phase_duration.png')
plt.show()

This visualization reveals how different cooking videos distribute their time across various cooking phases. We can observe that:

  1. Active cooking takes up a significant portion (17-38%) of the videos
  2. Videos vary considerably in how they balance explanation (16-21%) and prep (17-33%)
  3. Some videos allocate a small portion (around 8%) to tasting phase
  4. The distribution of phases can indicate the style and target audience of a cooking video:
    • Videos with more prep time may be better for beginners
    • Videos with more active cooking focus on the technical aspects
    • Videos with more explanation time provide more context and background

This type of analysis helps content creators understand the structure of successful cooking videos and allows viewers to find videos that match their preferred learning style.

​
3. Action Complexity Timeline for a Single Video

Let’s select one representative video and chart how the complexity (measured by actions per segment) changes over time.

# Select one video for timeline analysis (pick the one with most actions)
video_action_counts = segments_df.groupby('file_id')['action_count'].sum().reset_index()
representative_file_id = video_action_counts.sort_values('action_count', ascending=False).iloc[0]['file_id']

# Get the representative video's name
representative_name = videos_df[videos_df['file_id'] == representative_file_id]['filename'].iloc[0]

# Filter segments for the selected video
video_segments = segments_df[segments_df['file_id'] == representative_file_id].sort_values('start_time')

# Create a timeline plot with larger figure size for better spacing
plt.figure(figsize=(14, 8))
plt.plot(video_segments['start_time']/60, video_segments['action_count'], 'o-', color='blue', label='Actions')
plt.title(f'Action Complexity Timeline: {representative_name}')
plt.xlabel('Time (minutes)')
plt.ylabel('Number of Cooking Actions')
plt.grid(True, linestyle='--', alpha=0.7)

# Add cooking phase as background colors
phase_colors = {
    'intro': 'lightgray',
    'prep': 'lightyellow',
    'active_cooking': 'lightcoral',
    'plating': 'lightgreen',
    'tasting': 'lightblue',
    'explanation': 'lavender',
    'outro': 'lightgray'
}

# Add colored backgrounds for phases
for i in range(len(video_segments)-1):
    phase = video_segments.iloc[i]['cooking_phase']
    start = video_segments.iloc[i]['start_time']/60
    end = video_segments.iloc[i]['end_time']/60
    if phase in phase_colors:
        plt.axvspan(start, end, alpha=0.3, color=phase_colors[phase])

# Add annotations for peak complexity - position adjusted to avoid overlap
peak_idx = video_segments['action_count'].idxmax()
peak_time = video_segments.loc[peak_idx, 'start_time']/60
peak_actions = video_segments.loc[peak_idx, 'action_count']
peak_phase = video_segments.loc[peak_idx, 'cooking_phase']

# Position the annotation to the left and below the peak to avoid title overlap
plt.annotate(f'Peak: {peak_actions} actions\nPhase: {peak_phase}',
             xy=(peak_time, peak_actions),
             xytext=(peak_time-2.0, peak_actions-1),
             arrowprops=dict(arrowstyle='->'))

# Add a legend for phases - position in the upper right to avoid overlap
phase_patches = [plt.Rectangle((0,0),1,1, color=color, alpha=0.3) for color in phase_colors.values()]
plt.legend(phase_patches, phase_colors.keys(), loc='upper right', title='Cooking Phases')

# Add some padding to the top of the plot for the title
plt.ylim(top=plt.ylim()[1] * 1.1)

plt.tight_layout()
plt.savefig('action_complexity.png')
plt.show()

# Print insight
peak_minute = int(peak_time)
print(f"\nINSIGHT: For '{representative_name}', peak complexity happens at minute {peak_minute} during {peak_phase} phase")
print(f"INSIGHT: The busiest segment has {peak_actions} distinct cooking actions")

INSIGHT: For 'Sauces | Basics with Babish', peak complexity happens at minute 3 during prep phase
INSIGHT: The busiest segment has 7 distinct cooking actions

This visualization shows how cooking action complexity changes throughout a single video. By mapping the number of distinct cooking actions performed in each segment, we can identify:

  1. The peak complexity moments in the video (occurring at specific times with higher numbers of distinct actions)
  2. How cooking phases relate to action complexity (with notable activity in both prep and active cooking phases)
  3. The rhythm of the recipe - showing multiple spikes of 4-5 actions throughout the video interspersed with less complex segments
  4. Potential points where viewers might need to pause or rewatch to follow along

In the example above, we can see that this video has several periods of moderate complexity with 3-4 actions, with the peak complexity reaching 7 distinct cooking actions during the prep phase. This kind of insight can help content creators design more balanced cooking videos or add additional explanations at high-complexity points.

​
Exploring Advanced Queries

Besides the visualizations above, you can perform more targeted analyses with simple pandas queries:

# Find segments with the most ingredients mentioned
ingredients_by_segment = segments_df.sort_values('ingredients_mentioned', ascending=False)
print("\nSegments with Most Ingredients Mentioned:")
print(ingredients_by_segment[['filename', 'start_time', 'end_time', 'ingredients_mentioned']].head(3))

# Find distribution of cooking phases across all videos
phase_distribution = segments_df['cooking_phase'].value_counts(normalize=True) * 100
print("\nDistribution of Cooking Phases:")
for phase, percentage in phase_distribution.items():
    print(f"{phase}: {percentage:.1f}%")

Segments with Most Ingredients Mentioned:
                                               filename  start_time  end_time  \
192  Easy JAPANESE CURRY RICE » Made with Golden Curry          40        60
174     The Easiest Ramen To Make At Home - Miso Ramen         280       300
222                 Easy Carnitas | Basics with Babish         280       300

     ingredients_mentioned
192                      9
174                      7
222                      7

Distribution of Cooking Phases:
prep: 28.0%
active_cooking: 27.6%
explanation: 18.2%
tasting: 7.0%
intro: 5.9%
N/A: 5.2%
plating: 4.5%
outro: 3.1%
cleanup: 0.3%

​
Conclusion

In this guide, we’ve demonstrated how to:

  1. Create a structured extraction schema for cooking videos
  2. Process multiple YouTube videos with Cloudglue’s entity extraction
  3. Analyze the extracted data to uncover insights about ingredients, cooking phases, and action complexity
  4. Visualize the results using matplotlib

The structured data extraction capabilities of Cloudglue make it possible to transform unstructured video content into actionable insights. This approach can be extended to any domain where videos contain valuable information that needs to be structured for analysis.

​
Next Steps

  • Try different extraction schemas for other video domains (e.g., sports analysis, product reviews)
  • Implement real-time dashboards using extracted data
  • Build a search interface that lets users find specific segments across videos
  • Create a recommendation system based on extracted entities