Image Search & Recognition

Overview

The Material KAI Vision Platform uses advanced AI to understand and search through material images. When you upload a PDF catalog or search with an image, the system automatically analyzes every image to help you find exactly what you're looking for.

What It Does

Intelligent Image Understanding

Every image in your catalogs is analyzed using state-of-the-art AI to understand:

• Visual Appearance - Overall look and style of the material
• Colors - Dominant colors and color palettes
• Textures - Surface patterns and textures
• Material Type - What kind of material it is (fabric, tile, wood, etc.)
• Application - Where and how the material is typically used

This multi-dimensional understanding allows the platform to find materials that match your needs in different ways.

Smart Image Search

Upload any image to find similar materials in your catalog:

1. Visual Similarity - Find materials that look similar overall
2. Color Matching - Find materials with similar color schemes
3. Texture Matching - Find materials with similar surface patterns
4. Material Type - Find materials of the same type
5. Application - Find materials suitable for similar uses

Automatic Product Linking

The system automatically connects images to the products they represent:

• Product Images - Images showing the actual product
• Detail Shots - Close-up images of product features
• Application Examples - Images showing the product in use
• Variants - Images showing different colors or patterns

This ensures that when you search or browse, you see all relevant images for each product.

How It Works

AI-Powered Analysis

The platform uses two advanced AI systems working together:

Primary Analysis - Fast, accurate material detection using Qwen3-VL Vision AI

• Identifies material types, colors, and textures
• Provides quality scores for each analysis
• Processes images quickly and efficiently

Quality Validation - Claude AI validates uncertain results

• Reviews images that need additional analysis
• Provides enhanced descriptions and corrections
• Ensures high-quality results

Multi-Vector Search Architecture

The platform uses a sophisticated 7-embedding fusion system that combines multiple AI models in parallel for maximum search accuracy:

Embedding Types & Default Weights (Balanced Profile):

• Text Embedding (15%) - Semantic understanding from product names, descriptions, and metadata
• Visual Embedding (15%) - General visual similarity using SigLIP 768D embeddings
• Understanding Embedding (20%) - Spec-based search from Qwen3-VL analysis via Voyage AI (1024D)
• Color Embedding (12.5%) - Specialized color palette matching
• Texture Embedding (12.5%) - Surface pattern and texture recognition
• Style Embedding (12.5%) - Design aesthetic and style matching
• Material Embedding (12.5%) - Material type and category classification

Note: These are the default "balanced" weights. The system dynamically adjusts weights per-query using Query-Adaptive Weight Profiles — see below.

How It Works:

1. Query Understanding - GPT-4o-mini parses the query into structured fields (colors, finish, dimensions, pattern, style, etc.)
2. Weight Profile Selection - _select_weight_profile() picks optimal weights based on detected fields (e.g., color queries upweight color embedding to 30%)
3. Query Processing - Your search query is converted into visual and understanding embeddings
4. Parallel Search - All 6 embedding collections are searched simultaneously using async processing
5. Text Scoring - Keyword matching is performed on product metadata in parallel
6. Score Fusion - Results from all 7 embeddings are combined using the selected weight profile
7. Metadata Filtering - Your filters are applied as soft boosts to improve relevance
8. Final Ranking - Products are sorted by combined score and returned

Query-Adaptive Weight Profiles:

The system selects from 7 weight profiles based on query intent:

Profile	When Selected	Key Emphasis
product_name	Brand or product name detected	Text 40%
color_finish	Color or finish terms present	Color 30%
specification	Dimensions detected (e.g., 60x120cm)	Understanding 40%
texture_pattern	Pattern terms present	Texture 30%
style_aesthetic	Style or application terms	Style 25%
material_search	Explicit material type	Material 25%
balanced	No specific signal (default)	Even distribution

Profile selection is tracked in search_query_tracking for analytics.

Performance:

• Typical search time: 300-500ms
• All searches run in parallel using asyncio.gather() and thread pools
• Handles thousands of products efficiently

This multi-dimensional approach ensures you get the most relevant results by considering all aspects of material similarity simultaneously.

Intelligent Relevancy

Images are automatically linked to products and descriptions based on:

• Page Location - Images on the same page as product descriptions
• Visual Similarity - How well the image matches the product
• AI Confidence - How certain the AI is about the connection

The system assigns relevancy scores to help you find the most important images first.

Using Image Search

Search by Uploading an Image

When you have a material image and want to find similar materials in your catalog:

1. Upload Your Image - Upload any image of a material you're looking for
2. AI Analysis - The system analyzes the image to understand its characteristics
3. Find Matches - Get a list of similar materials from your catalog
4. View Results - See matching products with images and details

The search can find materials based on:

• Overall visual similarity
• Matching colors
• Similar textures
• Same material type
• Similar applications

Search from 3D Visualizations

When working with 3D room visualizations:

1. Generate 3D Scene - Create a 3D visualization with materials
2. Identify Materials - The system automatically identifies materials in the scene
3. Find Alternatives - Get suggestions for similar or alternative materials
4. Compare Options - View different material options in context

This helps you explore material options and find alternatives that work with your design.

Benefits

For Designers and Architects

• Quick Material Discovery - Find materials faster than browsing catalogs
• Visual Search - Search using images instead of keywords
• Explore Alternatives - Discover similar materials you might not have considered
• Confident Selections - See all relevant images and details before deciding

For Material Suppliers

• Better Product Visibility - Your products are found through visual search
• Automatic Organization - Images are automatically linked to products
• Rich Product Pages - All product images are organized and accessible
• Enhanced Search - Customers can find your products in multiple ways

Search Accuracy

The platform uses advanced AI to ensure accurate results:

• Multi-Dimensional Analysis - Considers visual appearance, color, texture, and more
• Quality Validation - AI validates uncertain results for accuracy
• Relevancy Scoring - Results are ranked by how well they match your search
• Continuous Improvement - The system learns and improves over time

Performance

The image search system is designed for speed and accuracy:

• Fast Search - Multi-vector search returns results in 300-500ms
• Parallel Processing - All 6 embedding collections searched simultaneously
• Efficient Storage - Images are optimized and cached for quick loading
• Scalable - Handles thousands of images without slowing down
• Reliable - Built on enterprise-grade infrastructure

Technical Implementation

True Async Parallel Execution

The multi-vector search uses advanced async programming to achieve maximum performance:

Architecture:

Query → Generate Embeddings → Search 6 Collections in Parallel ├─ Visual (SigLIP 768D) ├─ Understanding (Voyage AI 1024D) ├─ Color (SigLIP 768D) ├─ Texture (SigLIP 768D) ├─ Style (SigLIP 768D) └─ Material (SigLIP 768D) ↓ Combine Scores → Apply Filters → Return Results

Key Technologies:

• asyncio.gather() - Executes all searches simultaneously
• asyncio.to_thread() - Runs blocking VECS queries in thread pool
• VECS (pgvector) - Vector similarity search on PostgreSQL
• SigLIP2 - State-of-the-art vision-language model (768D visual embeddings)
• Voyage AI - Text embedding model for understanding embeddings (1024D)

Performance Benefits:

• 3-4x faster than sequential execution
• Non-blocking event loop for concurrent requests
• Efficient thread pool utilization
• Optimized database queries

Specialized Endpoints

In addition to the main multi-vector search, individual embedding searches are available:

• POST /api/search/color - Color palette matching only
• POST /api/search/texture - Texture pattern matching only
• POST /api/search/style - Design style matching only
• POST /api/search/material-type - Material category matching only

These are useful for:

• UI filter controls (e.g., "Show only warm colors")
• Specialized searches when you know exactly what you want
• A/B testing different embedding types
• Debugging and analysis

Search Response Format

The multi-vector search returns detailed scoring information for each result, including combined score, per-embedding scores (text, visual, understanding, color, texture, style, material), filter boost, and total processing time. This transparency allows you to understand why each result was returned and debug search quality.

Related Features

• Material Search - Search for materials using text descriptions
• PDF Processing - Automatic extraction of images from PDF catalogs
• Product Discovery - Intelligent product identification and organization