AI Models Integration Guide
Last Updated: 2025-12-26
Complete reference of all AI models used across the Material KAI Vision Platform.
AI Models Overview
| Model |
Provider |
Purpose |
Capability |
Cost (per 1M tokens) |
| Text Generation |
|
|
|
|
| Claude Sonnet 4.5 |
Anthropic |
Product discovery, enrichment |
95%+ accuracy |
$3 input / $15 output |
| Claude Haiku 4.5 |
Anthropic |
Fast validation |
Real-time |
$0.80 input / $4 output |
| Claude Opus 4.5 |
Anthropic |
Complex reasoning |
Highest accuracy |
$15 input / $75 output |
| GPT-4o |
OpenAI |
Alternative discovery |
94%+ accuracy |
$2.50 input / $10 output |
| GPT-4o Mini |
OpenAI |
Lightweight tasks |
Fast & cheap |
$0.15 input / $0.60 output |
| Text Embeddings |
|
|
|
|
| voyage-3.5 |
Voyage AI |
PRIMARY Text embeddings |
1024D vectors |
$0.06 input |
| voyage-3 |
Voyage AI |
Alternative text embeddings |
1024D vectors |
$0.06 input |
| voyage-3-lite |
Voyage AI |
Lightweight embeddings |
512D vectors |
$0.02 input |
| text-embedding-3-small |
OpenAI |
LEGACY (CI changelog only, retired from production 2026-04) |
1536D vectors |
$0.02 input |
| Vision Models |
|
|
|
|
| Qwen3-VL-32B-Instruct |
HuggingFace Endpoint |
PRIMARY Vision analysis |
State-of-the-art OCR |
Cloud endpoint |
| Visual Embeddings |
|
|
|
|
| SLIG (SigLIP2) Visual |
HuggingFace Endpoint |
General visual embeddings |
768D vectors |
Cloud endpoint |
| SLIG (SigLIP2) Color |
HuggingFace Endpoint |
Color-guided embeddings |
768D vectors |
Cloud endpoint |
| SLIG (SigLIP2) Texture |
HuggingFace Endpoint |
Texture-guided embeddings |
768D vectors |
Cloud endpoint |
| SLIG (SigLIP2) Style |
HuggingFace Endpoint |
Style-guided embeddings |
768D vectors |
Cloud endpoint |
| SLIG (SigLIP2) Material |
HuggingFace Endpoint |
Material-guided embeddings |
768D vectors |
Cloud endpoint |
Model Details
1. Claude Sonnet 4.5 (Anthropic)
Purpose: Product discovery, enrichment, validation
Capabilities:
- Analyze complex product catalogs
- Extract structured metadata
- Validate product completeness
- Generate product descriptions
Performance:
- Accuracy: 95%+
- Latency: 2-5 seconds
- Cost: $3 per 1M input tokens
When to Use:
- Product discovery
- Complex analysis
- Detailed enrichment
- Quality validation
2. Claude Haiku 4.5 (Anthropic)
Purpose: Fast validation, real-time processing
Capabilities:
- Quick product candidate identification
- Basic metadata extraction
- Real-time validation
Performance:
- Accuracy: 92%+
- Latency: 500ms-1s
- Cost: $0.80 per 1M input tokens
When to Use:
- Fast validation
- Real-time processing
- Cost-sensitive operations
- Batch processing
3. GPT-4o (OpenAI)
Purpose: Alternative product discovery
Capabilities:
- Product identification
- Metadata extraction
- Alternative to Claude
Performance:
- Accuracy: 94%+
- Latency: 3-6 seconds
- Cost: $5 per 1M input tokens
When to Use:
- Alternative discovery model
- Comparison validation
- Fallback option
4. Voyage AI voyage-3.5 (updated 2026-04)
Purpose: Generate text embeddings for semantic search (sole production text embedder)
Capabilities:
- Convert text to 1024D vectors (stored as halfvec in VECS)
- Enable semantic similarity search
- Fast embedding generation, supports
document and query input types
- Dict key:
text_1024 (was previously text_1536 under OpenAI text-embedding-3-small)
Performance:
- Dimension: 1024D
- Latency: 100-300ms
- Cost: $0.06 per 1M tokens
Note: OpenAI text-embedding-3-small (1536D) was retired from the production path in 2026-04. It is only retained for the legacy CI changelog workflow.
When to Use:
- Text embedding generation
- Semantic search
- Similarity comparison
5. Qwen3-VL-32B-Instruct (HuggingFace Endpoint) - PRIMARY VISION MODEL
Purpose: State-of-the-art vision-language model for image analysis, OCR, material recognition
Endpoint Configuration:
- URL:
https://gbz6krk3i2is85b0.us-east-1.aws.endpoints.huggingface.cloud
- Service Name:
mh-qwen332binstruct
- Namespace:
basiliskan
- Model: Locked to 32B only (8B removed)
Capabilities:
- Advanced image understanding
- Optical Character Recognition (OCR)
- Material property extraction
- Color and texture analysis
- Image quality scoring
- Multi-image reasoning
Performance:
- Accuracy: State-of-the-art vision understanding
- Latency: 2-5 seconds per image
- Cost: Cloud endpoint (auto-pause enabled)
When to Use:
- Image analysis
- OCR on material specs
- Material property extraction
- Quality scoring
6-10. SLIG (SigLIP2) Specialized Embeddings (updated 2026-04)
Purpose: Multi-modal visual embeddings via HuggingFace Cloud Endpoint. Replaced legacy OpenAI CLIP (256/512/1536D) and SigLIP-SO400M (1152D) in 2026-04 — those columns were dropped from the database.
5 Embedding Types (all 768D halfvec, written directly to VECS):
Visual Embeddings (768D) → image_slig_embeddings
- Overall visual appearance
- Enables visual similarity search
- Producer key:
visual_768
Color Embeddings (768D) → image_color_embeddings
- Text-guided color palette analysis
- Producer key:
color_slig_768
Texture Embeddings (768D) → image_texture_embeddings
- Text-guided surface texture analysis
- Producer key:
texture_slig_768
Style Embeddings (768D) → image_style_embeddings
- Text-guided design aesthetic
- Producer key:
style_slig_768
Material Embeddings (768D) → image_material_embeddings
- Text-guided material classification
- Producer key:
material_slig_768
Additionally, an Understanding Embedding (1024D, Voyage AI from Qwen3-VL vision_analysis JSON) → image_understanding_embeddings is generated inline for spec-based semantic search.
Performance:
- Dimension: 768D (specialized) / 1024D (understanding)
- Latency: 150-400ms per image
- Cost: Cloud endpoint, auto-pause enabled
When to Use:
- Visual similarity search
- Color-based discovery
- Texture, style, and material analysis
- Spec-based search via understanding embedding
11. Anthropic Semantic Chunking
Purpose: Intelligent text segmentation
Capabilities:
- Split text at semantic boundaries
- Preserve context
- Quality scoring
- Multiple strategies
Strategies:
- Semantic: Paragraph/sentence boundaries
- Fixed Size: Fixed character count
- Hybrid: Combination of both
- Layout-Aware: Respect document layout
Performance:
- Chunk quality: 0.85-0.95
- Latency: 1-2 seconds per 10K chars
- Cost: Included in Anthropic API
12. Direct Vector DB RAG System (Claude 4.5)
Purpose: Retrieval-Augmented Generation with Multi-Vector Search
Capabilities:
- Direct vector database queries (no intermediate indexing)
- 6-way parallel multi-vector search
- Claude 4.5 synthesis (200K context)
- Intelligent embedding fusion
Performance:
- Retrieval latency: 300-500ms (parallel execution)
- Accuracy: 90%+ (multi-vector fusion)
- Scalability: 1M+ documents
📊 Model Usage by Pipeline Stage
| Stage |
Primary Model |
Secondary Model |
Purpose |
| 0 |
Claude Sonnet 4.5 |
GPT-4o |
Product discovery |
| 2 |
Anthropic Chunking |
- |
Text segmentation |
| 4 |
voyage-3.5 (Voyage AI) |
- |
Text embeddings (1024D, updated 2026-04) |
| 6 |
Qwen3-VL 17B |
- |
Image analysis |
| 7-10 |
SLIG (SigLIP2, 5 types) |
- |
Visual embeddings (768D) |
| 11 |
Claude Haiku 4.5 |
Claude Sonnet 4.5 |
Product validation |
| 13 |
Claude Sonnet 4.5 |
- |
Quality enhancement |
💰 Cost Optimization
Strategies:
- Use Haiku for fast validation (10x cheaper)
- Batch embeddings to reduce API calls
- Cache results for repeated queries
- Use focused extraction to reduce image analysis
Example Cost per PDF:
- Small PDF (10 pages): $0.50-$1.00
- Medium PDF (50 pages): $2.00-$4.00
- Large PDF (200 pages): $8.00-$15.00
🔐 API Keys & Configuration
Required Environment Variables:
ANTHROPIC_API_KEY — Anthropic Claude API keyOPENAI_API_KEY — OpenAI API keyVOYAGE_API_KEY — Voyage AI API keyQWEN_ENDPOINT_URL — HuggingFace Qwen endpoint URLQWEN_ENDPOINT_TOKEN — HuggingFace endpoint tokenSLIG_ENDPOINT_URL — HuggingFace SLIG endpoint URLSLIG_ENDPOINT_TOKEN — HuggingFace SLIG endpoint token
The model configuration maps each task to its designated model: discovery uses claude-sonnet-4-5, validation uses claude-haiku-4-5, text_embeddings uses voyage-3.5, vision uses Qwen/Qwen3-VL-32B-Instruct, and visual_embeddings uses SLIG.
📈 Performance Benchmarks
Accuracy:
- Product detection: 95%+
- Material recognition: 90%+
- Metafield extraction: 88%+
- Search relevance: 85%+
Speed:
- Product discovery: 3-5 seconds
- Image analysis: 2-4 seconds per image
- Embedding generation: 100-200ms
- Search query: 200-800ms
Cost:
- Per PDF: $0.50-$15.00
- Per search: $0.001-$0.01
- Per image: $0.05-$0.20
New Models (2025-12-26)
Voyage-3 (Voyage AI) - PRIMARY TEXT EMBEDDINGS
Purpose: High-quality text embeddings for semantic search and retrieval
Capabilities:
- Generate 1024-dimensional text embeddings
- Superior semantic understanding vs OpenAI
- Optimized for retrieval tasks
- Better performance on domain-specific content
Performance:
- Dimensions: 1024D
- Latency: 100-300ms
- Cost: $0.06 per 1M tokens
- Quality: Superior to text-embedding-3-small
When to Use:
- PRIMARY choice for all text embeddings
- Product descriptions
- Material specifications
- Document chunks
- Semantic search
Migration: Replaced text-embedding-3-small in production 2026-04 (1536D → 1024D, dict key text_1536 → text_1024)
SLIG (SigLIP2) - VISUAL EMBEDDINGS (HuggingFace Endpoint)
Purpose: Cloud-based visual embeddings for image similarity search
Endpoint Configuration:
- URL:
https://xxxxxxxx.us-east-1.aws.endpoints.huggingface.cloud
- Service Name:
mh-siglip2
- Namespace:
basiliskan
- Dimensions: 768D (all embeddings)
Capabilities:
- General visual embeddings (image_embedding mode)
- Text-guided visual embeddings (text_embedding mode)
- Color-guided embeddings
- Texture-guided embeddings
- Material-guided embeddings
- Style-guided embeddings
- Zero-shot classification (zero_shot mode)
- Image-text similarity (similarity mode)
Performance:
- Dimensions: 768D (all embeddings)
- Latency: 500ms-2s
- Cost: Cloud endpoint (auto-pause enabled)
- Quality: Superior to CLIP, optimized for materials
When to Use:
- PRIMARY choice for visual embeddings
- Image similarity search
- Text-guided visual search (color, texture, material, style)
- Multimodal fusion (text + visual)
Voyage-3-Lite (Voyage AI) - LIGHTWEIGHT EMBEDDINGS
Purpose: Fast, cost-effective embeddings for simple tasks
Capabilities:
- Generate 512-dimensional embeddings
- Faster than Voyage-3
- Lower cost
Performance:
- Dimensions: 512D
- Latency: 50-150ms
- Cost: $0.02 per 1M tokens
- Quality: Good for simple tasks
When to Use:
- Simple semantic search
- Fast lookups
- Cost-sensitive operations
- Non-critical embeddings
Model Selection Guide
Text Embeddings
- Voyage-3 (PRIMARY) - All production text embeddings
- Voyage-3-Lite - Simple/fast tasks only
- text-embedding-3-small - Retired 2026-04 (CI changelog workflow only)
Vision Analysis
- Qwen3-VL-32B-Instruct (PRIMARY) - All production vision tasks (HuggingFace endpoint)
- Claude Sonnet 4.5 - Validation for low-confidence results
Visual Embeddings
- SLIG (SigLIP2) (PRIMARY) - All visual embeddings (768D, HuggingFace endpoint)
- General visual (image_embedding mode)
- Text-guided (color, texture, material, style) (text_embedding mode)
Text Generation
- Claude Sonnet 4.5 (PRIMARY) - Complex reasoning
- Claude Haiku 4.5 - Fast validation
- GPT-4o - Alternative/fallback
Last Updated: December 26, 2025
Version: 2.0.0
Status: Production