sammykumar-ComfyUI-SwissArmyKnife/docs/infrastructure/caching/CACHING.md

Gemini Media Description Caching

This document explains the caching functionality implemented for the Gemini custom nodes.

Overview

The caching system automatically stores Gemini API responses to avoid redundant API calls when analyzing the same media with the same description prompt. This provides:

• Fast repeated analyses - Cached results return in milliseconds instead of waiting for API calls
• API cost savings - Avoids duplicate API requests for identical media+prompt combinations
• Transparent operation - Caching is always enabled and requires no user configuration
• Accurate isolation - Separate cache entries for each unique media+prompt+model combination

How It Works

Cache Key Generation

Each cache entry uses a unique key based on:

1. Media identifier - Uniquely identifies the media content
2. Description mode - The selected prompt option (e.g., "Describe without clothing")
3. Model settings - The Gemini model name and type

Media Identification

• Video files: Uses file path + modification time + file size
• Image tensors: Uses content hash of the tensor data
• Uploaded files: Uses the file path in ComfyUI's input directory + metadata

Description Modes

Caching works with all 4 description modes:

• "Describe without clothing"
• "Describe with clothing"
• "Describe without clothing (No bokeh)"
• "Describe with clothing (No bokeh)"

Each mode creates a separate cache entry for the same media file.

Storage

Cache entries are stored as JSON files in the cache/gemini_descriptions/ directory:

• Each file contains the description, metadata, and timestamp
• Cache persists across ComfyUI restarts
• Cache directory is created automatically
• Cache files are excluded from git via .gitignore

Cache Behavior

Cache Hit (Fast Path)

User analyzes media → Check cache → Found → Return cached description (< 1ms)

Cache Miss (API Path)

User analyzes media → Check cache → Not found → Call Gemini API → Store result → Return description

Cache Isolation Examples

Same video file with different prompts:

• video.mp4 + "Describe without clothing" → Cache entry A
• video.mp4 + "Describe with clothing" → Cache entry B
• video.mp4 + "Describe without clothing (No bokeh)" → Cache entry C
• video.mp4 + "Describe with clothing (No bokeh)" → Cache entry D

Same prompt with different models:

• video.mp4 + "Describe without clothing" + gemini-2.5-flash → Cache entry A
• video.mp4 + "Describe without clothing" + gemini-2.5-pro → Cache entry B

Implementation Details

Files Modified

• utils/cache.py - New cache utility module
• utils/nodes.py - Added caching to GeminiVideoDescribe and GeminiImageDescribe classes

Cache Integration Points

1. Before Gemini API call - Check cache for existing result
2. After successful API response - Store result in cache
3. Cache hit handling - Return cached result with appropriate status message

Status Messages

Cache hits are indicated in the Gemini status output:

🤖 Gemini Analysis Status: ✅ Complete (Cached)
• Model: models/gemini-2.5-flash
• Cached: 2025-09-02 10:16:32

Regular API calls show:

🤖 Gemini Analysis Status: ✅ Complete
• Model: models/gemini-2.5-flash
• API Key: ****1234

Performance Benefits

• First analysis: Normal API call timing (~2-5 seconds)
• Subsequent analyses: Cache retrieval (~0.1ms)
• API savings: 100% reduction in duplicate calls
• User experience: Instant results for repeated workflows

Cache Management

Automatic Behavior

• Cache entries are created automatically
• No user configuration required
• Cache persists across sessions
• Old entries remain until manually cleared

Manual Management

The cache can be inspected or cleared programmatically:

from nodes.cache import get_cache

cache = get_cache()

# Get cache statistics
info = cache.get_cache_info()
print(f"Entries: {info['entries']}, Size: {info['total_size_mb']} MB")

# Clear specific entry (not exposed in UI)
# cache._get_cache_file_path(key) and os.remove()

Cache Directory Location

• Default: <comfyui_swissarmyknife>/cache/gemini_descriptions/
• Contains JSON files with SHA256 hash names
• Safe to delete entire directory to clear all cache

User Workflow Impact

The caching is completely transparent to users:

1. First time analyzing media: Works exactly as before (API call)
2. Re-analyzing same media+prompt: Results appear instantly from cache
3. Different prompts on same media: Each creates new cache entry
4. Different media files: Each gets separate cache entries

Users will notice:

• Faster repeated analyses
• Cache hit indicators in status messages
• No additional configuration needed
• Consistent behavior across ComfyUI restarts