Image Embeddings (Gemini Embedding 2)

Note: The CLIP/Jina pipeline was removed in early 2026. This document now describes the replacement: Gemini Embedding 2 multimodal embeddings.

Overview

Images are embedded using Gemini Embedding 2 (gemini-embedding-2-preview), a native multimodal model that encodes images and text into a shared 1536-dimensional vector space. This enables cross-modal search: describe what you want in words and find matching images.

Features

1. Visual Semantic Search

Find images using natural language descriptions (text → image cross-modal search).

2. Duplicate Detection

Identify duplicate or near-duplicate images using perceptual hash (pHash) fingerprinting stored in digital_assets.fingerprint.

3. Product Image Search

Find images visually similar to a product by querying the unified embedding space.

Architecture

ImageFullAnalysisWorker
        │
        ├─ Step 1: pHash fingerprint (local, no API)
        │          → stored in digital_assets.fingerprint
        │
        ├─ Step 2: Gemini Embedding 2 (image + metadata text)
        │          → stored in content_embeddings (content_type='unified',
        │            embedding_model='gemini-embedding-2-preview',
        │            unified_embedding vector(1536))
        │
        └─ Step 3: Gemini Flash vision analysis (optional, for CRM metadata)
                   → stored in digital_assets.ai_visual_description

Nightly Backfill

ImageEmbeddingPopulationWorker (Sidekiq::IterableJob) runs nightly at 2:30 AM CT. It queues up to 5,000 images per run, prioritising product primary images first. Uses cursor-based checkpointing so a deploy mid-run resumes from the last record.

Database Schema

The content_embeddings table (partitioned; image partition: content_embeddings_images):

Column	Type	Description
content_type	string	'unified' for Gemini image embeddings
embedding_model	string	'gemini-embedding-2-preview'
unified_embedding	vector	1536-dimensional vector
embedding_dimensions	integer	1536

Fingerprints live on the parent record:

Column	Type	Description
digital_assets.fingerprint	bigint	pHash perceptual hash

Configuration

The Gemini API key is stored in Rails credentials at google.gemini.api_key.

Rate limiting is handled by Embedding::Gemini via a Redis sliding-window limiter (default: 300 requests/minute, configurable via GEMINI_EMBED_REQUESTS_PER_MINUTE).

Usage

Queue full analysis for an image

ImageFullAnalysisWorker.perform_async(image.id)

Semantic image search

# Find images matching a text description
ContentEmbedding::ImageEmbedding.semantic_search("bathroom with heated floors", limit: 10)

# Via the top-level service
ContentEmbedding.unified_visual_search("snow melting driveway", limit: 10)

Find similar images

image = Image.find(123)
image.image_embeddings.unified_content
     .nearest_neighbors(:unified_embedding, image_vector, distance: :cosine)
     .limit(10)

Rake Tasks

# Check backfill progress
rake embeddings:progress

# Check detailed image essentials (fingerprints + Gemini coverage)
rake embeddings:essentials_stats

# Trigger nightly backfill worker manually (all active images)
# Enqueues ImageEmbeddingPopulationWorker via Sidekiq
rake embeddings:queue_all_image_full

# Product primary images only
rake embeddings:queue_all_product_full

# Incremental batches (resumable)
rake embeddings:populate_image_full        # full pipeline, batch: 50
rake embeddings:populate_image_vision      # vision analysis only, batch: 100

# Fingerprints
rake embeddings:populate_fingerprints      # incremental, batch: 100
rake embeddings:queue_all_fingerprints     # all at once

# Duplicate detection (pHash)
rake embeddings:find_phash_duplicates
rake embeddings:find_phash_duplicates[10]  # Hamming distance <= 10

# Test Gemini API connectivity
rake embeddings:test_gemini_embed

# Embedding statistics
rake embeddings:stats

Performance

• Embedding latency: ~1–3s per image (includes image download + Gemini API call)
• Rate limit: 300 requests/minute (shared across all workers)
• Vector dimensions: 1536 (Matryoshka truncation from 3072 full quality)
• Index type: HNSW with cosine distance, per-partition (~1–5ms queries)
• Storage: ~6KB per embedding (1536 floats × 4 bytes)