Complete AI Models Inventory

Last Updated: January 18, 2026 Version: 3.0.0 Total Models: 30+ across 5 providers

---

📊 AI Models by Provider

1. Anthropic (Claude)

Claude Sonnet 4.5

• Use Cases:
  • Product discovery from PDFs (Stage 0A)
  • Deep metadata extraction
  • Quality validation
  • Web scraping product discovery
  • XML field mapping
  • Search deduplication analysis
• Context: 200,000 tokens
• Cost: ~$0.015 per 1K tokens
• Performance: Highest accuracy for complex reasoning

Claude Haiku 4.5

• Use Cases:
  • Fast content classification
  • Product boundary detection
  • Saved search similarity analysis
  • Quick validation tasks
• Context: 200,000 tokens
• Cost: ~$0.0025 per 1K tokens
• Performance: 3× faster than Sonnet, 90%+ accuracy

---

2. OpenAI

GPT-4o

• Use Cases:
  • Product discovery (alternative to Claude)
  • Conversational AI agents
  • Complex reasoning tasks
  • Document analysis
• Context: 128,000 tokens
• Cost: ~$0.01 per 1K tokens

GPT-5

• Use Cases:
  • Advanced conversational AI
  • Complex multi-step reasoning
  • Future model support
• Context: TBD
• Cost: TBD

text-embedding-3-small (retired from production 2026-04)

• Status: Retired in 2026-04 from the production path. Still used only for the legacy CI changelog workflow.
• Replacement: Voyage AI voyage-3.5 (1024D) is now the sole production text embedder. See below.
• Dimensions: 1536D (historical)
• Cost: $0.02 per 1M tokens

Voyage AI voyage-3.5 (production text embedder, updated 2026-04)

• Use Cases:
  • Text chunk embeddings
  • Semantic search
  • Understanding embeddings (from Qwen3-VL vision_analysis JSON)
• Dimensions: 1024D (stored as halfvec in VECS)
• Dict key: text_1024
• Cost: $0.06 per 1M tokens

---

3. HuggingFace Endpoint

Qwen3-VL 32B Vision

• Use Cases:
  • Material image analysis
  • Product classification
  • OCR and text extraction from images
  • Image quality scoring
  • Generating structured analysis for understanding embeddings
• Parameters: 32 billion
• Modality: Vision + Text
• Deployment: HuggingFace Dedicated Endpoint (auto-pause enabled)
• Performance:
  • 69.4% MMMU score
  • #1 ranked for OCR tasks
  • 85%+ material recognition accuracy
• Cost: $0.30 per 1M tokens
• Pipeline Stages: Image Analysis (Stage 6, 8)

---

4. Google

SLIG (SigLIP2 via HuggingFace Cloud Endpoint) — Primary visual embedder, updated 2026-04

• Use Cases:
  • Visual embeddings (5 types per image)
  • Image-text similarity via similarity mode
  • Visual search
  • Multi-vector search
• Dimensions: 768D per embedding type (halfvec in VECS)
• Performance: Superior to legacy CLIP 512D and SigLIP-SO400M 1152D (both retired 2026-04)
• Cost: Cloud endpoint (auto-pause enabled)
• Pipeline Stages: Image Embedding Generation (Stage 7)

5 Embedding Types Generated (all 768D → VECS):

1. Visual → image_slig_embeddings (producer key visual_768)
2. Color → image_color_embeddings (producer key color_slig_768)
3. Texture → image_texture_embeddings (producer key texture_slig_768)
4. Style → image_style_embeddings (producer key style_slig_768)
5. Material → image_material_embeddings (producer key material_slig_768)

Plus an Understanding Embedding (1024D Voyage AI from Qwen3-VL vision_analysis JSON) → image_understanding_embeddings for spec-based semantic search.

---

5. Replicate (14 Models for Interior Design)

Text-to-Image Models (7 models)

1. FLUX.1-dev

   • Provider: Replicate
   • Model: black-forest-labs/flux-dev
   • Cost: $0.025 per generation
   • Status: ✅ Working

2. FLUX.1-schnell

   • Provider: Replicate
   • Cost: $0.015 per generation
   • Status: ✅ Working

3. SDXL (Stable Diffusion XL)

   • Provider: Replicate
   • Cost: $0.020 per generation
   • Status: ✅ Working

4. Playground v2.5

   • Provider: Replicate
   • Model: playgroundai/playground-v2.5-1024px-aesthetic
   • Cost: $0.010 per generation
   • Status: ✅ Working

5. Stable Diffusion 3

   • Provider: Replicate
   • Model: stability-ai/stable-diffusion-3
   • Cost: $0.055 per generation
   • Status: ✅ Working

6. Kandinsky 2.2

   • Provider: Replicate
   • Cost: $0.015 per generation
   • Status: ✅ Working

7. Proteus v0.2

   • Provider: Replicate
   • Cost: $0.018 per generation
   • Status: ✅ Working

Image-to-Image Models (7 models)

Production-Ready (3 models):

1. ComfyUI Interior Remodel

   • Provider: Replicate
   • Model: jschoormans/comfyui-interior-remodel
   • Cost: $0.020 per generation
   • Status: ✅ Working

2. Interiorly Gen1 Dev

   • Provider: Replicate
   • Model: julian-at/interiorly-gen1-dev
   • Cost: $0.015 per generation
   • Status: ✅ Working

3. Designer Architecture

   • Provider: Replicate
   • Model: davisbrown/designer-architecture
   • Cost: $0.018 per generation
   • Status: ✅ Working

Experimental (4 models):

4. Interior AI - Status: ⚠️ Experimental
5. Interior V2 - Status: ⚠️ Experimental
6. Adirik Interior Design - Status: ⚠️ Experimental
7. Interior Design SDXL - Status: ⚠️ Experimental

---

📈 Model Usage by Feature

PDF Processing Pipeline

• Stage 0A (Product Discovery): Claude Sonnet 4.5 or GPT-4o
• Stage 0B (Entity Discovery): Claude Sonnet 4.5 or GPT-4o
• Stage 2 (Text Embeddings): Voyage AI voyage-3.5 (1024D, updated 2026-04)
• Stage 6 (Image Classification): Qwen3-VL 17B Vision
• Stage 7 (SLIG Embeddings): SigLIP2 via HuggingFace cloud endpoint (768D, 5 types + understanding 1024D)
• Stage 8 (Image Analysis): Qwen3-VL 17B Vision

Web Scraping Integration

• Product Discovery: Claude Sonnet 4.5 (default), GPT-5, or Claude Haiku 4.5
• Content Analysis: Same models as PDF Stage 0A

XML Import

• Field Mapping: Claude Sonnet 4.5
• Dynamic mapping with AI reasoning

Interior Design Generation

• Text-to-Image: 7 Replicate models (FLUX, SDXL, etc.)
• Image-to-Image: 3 production-ready Replicate models
• Parallel processing: 3 concurrent generations

Saved Searches Deduplication

• Similarity Analysis: Claude Haiku 4.5
• Semantic fingerprinting and merging

Price Monitoring

• Web Content Analysis: Firecrawl + Claude for price extraction

---

💰 Cost Optimization Strategy

High-Volume Operations (Use Cheaper Models)

• Quick Classification: Claude Haiku 4.5 ($0.0025/1K tokens)
• Text Embeddings: Voyage AI voyage-3.5 ($0.06/1M tokens, updated 2026-04)
• Visual Embeddings: SLIG SigLIP2 cloud endpoint (auto-pause enabled)

High-Accuracy Operations (Use Premium Models)

• Product Discovery: Claude Sonnet 4.5 ($0.015/1K tokens)
• Metadata Extraction: Claude Sonnet 4.5
• Spatial Analysis: Claude Sonnet 4.5 Vision

Parallel Processing

• Interior Design: Process 3 models concurrently
• CLIP Embeddings: Batch process 20 images at a time
• Image Downloads: 5 concurrent downloads

---

🎯 Model Selection Guidelines

When to Use Claude Sonnet 4.5

• Complex reasoning required
• High accuracy needed
• Detailed metadata extraction
• Multi-step analysis
• Cost: Higher but worth it for quality

When to Use Claude Haiku 4.5

• Simple classification tasks
• Quick validation
• High-volume operations
• Real-time analysis
• Cost: 6× cheaper than Sonnet

When to Use GPT-4o/GPT-5

• Conversational AI
• Alternative to Claude (load balancing)
• Multimodal capabilities needed
• Cost: Competitive with Claude

When to Use Qwen3-VL

• Image analysis and OCR
• Material recognition
• Visual quality scoring
• Cost: Moderate, excellent value

When to Use SigLIP CLIP

• Visual similarity search
• Multi-vector embeddings
• High-volume image processing
• Cost: Free (self-hosted)

---

📊 Performance Benchmarks

Model	Use Case	Speed	Accuracy	Cost/Operation
Claude Sonnet 4.5	Product Discovery	3-5s	95%+	$0.05-0.15
Claude Haiku 4.5	Classification	0.5-1s	90%+	$0.01-0.03
GPT-4o	Discovery	2-4s	93%+	$0.04-0.12
Qwen3-VL	Image Analysis	2-3s	90%+	$0.02-0.05
SigLIP CLIP	Embeddings	0.1-0.3s	95%+	$0.00
FLUX Dev	Interior Design	5-13s	92%+	$0.025
ComfyUI	Room Transform	8-15s	88%+	$0.020

---

🔄 Model Fallback Strategy

Primary → Secondary → Tertiary

Product Discovery:

1. Claude Sonnet 4.5 (primary)
2. GPT-4o (secondary)
3. Claude Haiku 4.5 (tertiary, lower accuracy)

Image Analysis:

1. Qwen3-VL 17B (primary)
2. Claude Vision (secondary, more expensive)
3. GPT-4 Vision (tertiary)

Visual Embeddings:

1. SigLIP ViT-SO400M (primary)
2. CLIP ViT-B/32 (secondary)
3. Skip if both fail (graceful degradation)

---

🆕 Recently Added Models

January 2026:

• ✨ Added 14 Replicate models for interior design
• ✨ Expanded SigLIP to generate 5 embedding types (was 1)
• ✨ Added GPT-5 support (future-proofing)

December 2025:

• ✨ Upgraded to Claude Sonnet 4.5 (from 3.5)
• ✨ Upgraded to Claude Haiku 4.5 (from 3.5)
• ✨ Added Google SigLIP as primary CLIP model

---

📚 Related Documentation

• PDF Processing Pipeline - Detailed pipeline with model usage
• Interior Design Models - 14 Replicate models
• Web Scraping Integration - AI-powered product discovery
• Async Processing & Limits - Concurrency limits per model

---

Total Investment: 30+ AI models across 5 providers Total Cost Range: $0.00 - $0.055 per operation (varies by model and task) Success Rate: 95%+ across all models Uptime: 99.5%+ (production environment)