Material Search

Overview

The Material KAI Vision Platform provides powerful search capabilities to help you find exactly the materials you need. The search system understands natural language, recognizes visual patterns, and can filter by specific material properties.

How Search Works

Intelligent Multi-Dimensional Search

When you search for materials, the platform analyzes your query in multiple ways simultaneously:

All 7 dimensions are combined intelligently to give you the most relevant results.

Specification-Based Search

The Understanding embedding enables searching by technical specifications:

This works because the platform embeds Qwen3-VL's structured analysis of each image (material type, dimensions, finishes, properties) via Voyage AI.

Natural Language Search

Simply describe what you're looking for in plain language:

The system automatically understands your requirements and finds matching materials.

Search Features

Text-Based Search

Search using descriptions, product names, or material properties:

Image-Based Search

Upload an image to find similar materials:

Property Filters

Refine your search with specific material properties:

Search Results

Relevancy Ranking

Results are ranked by how well they match your search:

Result Information

Each search result includes:

Search Tips

Getting Better Results

Common Search Patterns

Finding Alternatives:

Exploring Collections:

Specific Requirements:

Search Performance

The search system is optimized for speed and accuracy:

Advanced Features

Query Understanding

The system automatically understands complex queries:

Query-Adaptive Weight Profiles

The 7-vector fusion search dynamically adjusts embedding weights based on what the query is about. Instead of using fixed weights for every search, the system analyzes the parsed query fields and selects the optimal weight profile.

How It Works:

  1. GPT-4o-mini parses the natural language query into structured fields (colors, finish, material_type, dimensions, pattern, style, designer, collection)
  2. _select_weight_profile() examines which fields are present and selects the best profile
  3. The selected weights are applied to the 7-vector fusion scoring

Available Profiles:

Profile Trigger Text Visual Understanding Color Texture Style Material
product_name Product name or brand detected 40% 25% 15% 5% 5% 5% 5%
color_finish Colors or finish terms present 10% 20% 15% 30% 5% 15% 5%
specification Dimensions detected (e.g., 60x120cm) 25% 10% 40% 5% 5% 5% 10%
texture_pattern Pattern terms present 10% 25% 15% 5% 30% 10% 5%
style_aesthetic Style or application terms 10% 25% 15% 10% 10% 25% 5%
material_search Explicit material type 15% 15% 25% 5% 10% 5% 25%
balanced No specific signal (default) 15% 15% 20% 12.5% 12.5% 12.5% 12.5%

Selection Priority: dimensions → colors/finish → pattern → material → style/application → balanced

Examples:

Monitoring:

Implementation: unified_search_service.pyWEIGHT_PROFILES dict + _select_weight_profile() method

Result Diversity

Search results include variety while maintaining relevance:

Explainable Search Spec

When the KAI agent calls material_search, it now provides a structured search specification that decomposes the user's query across all 7 embedding dimensions. This spec is displayed as a collapsible SearchSpecCard above the product results in the chat.

Fields in the search spec:

The spec is generated by the LLM as part of its tool call (no extra API call) and emitted via onChunk with type search_spec for real-time display in the frontend.

Why it matters: Architects and specifiers need to justify material choices to clients. The search spec shows why each result was selected — turning a black-box search into an explainable recommendation.


Design Inspiration URL Search

Users can paste any design URL (Houzz, Pinterest, Dezeen, ArchDaily, manufacturer sites) and the analyze_inspiration_url tool will:

  1. Scrape the page using Firecrawl API
  2. Extract design tokens via Claude Haiku: colors (with hex codes), materials, textures, styles, room type
  3. Search the catalog for matching products using the 7-vector fusion search
  4. Display results as an InspirationCard (color swatches, material/style tags, source thumbnail) followed by matched products

Accessible to all users via the Globe icon button in the chat toolbar, which opens a modal for URL input with an optional surface focus selector (floors, walls, countertops, etc.).

Credit cost: 1 credit for the Firecrawl scrape + Haiku token cost for extraction.


Related Features