SGPS-QA Architecture Mapping: Layer-by-Layer Analysis
Trent Carter
8/1/2025
Architecture Flow Table
768D Input
↓
384D Compression (compress_1)
↓
384D Semantic GPS (semantic_gps)
↓
384D PROJECTION HEAD (dimension_partitioning)
├── Core: 256D (semantic meaning)
├── Spatial: 64D (GPS coordinates)
└── Sequential: 64D (A→B→C→D→E tracking)
↓
512D Compression (compress_2)
↓
192D Nuclear Bottleneck + Attention
↓
512D Expansion (nuclear_expand)
↓
384D Expansion (expand_2)
↓
384D PROJECTION DECODE (dimension_reconstruction)
↓
768D Teacher Alignment (teacher_align)
// Compression Path:
768 → 384 (compress_1)
384 → GPS positioning (semantic_gps)
384 → PROJECTION HEAD (dimension_partitioning: 256+64+64)
384 → 512 (compress_2)
512 → 192 (nuclear_compress)
192 → attention
192 → 512 (nuclear_expand)
// Expansion Path:
512 → 384 (expand_2)
384 → PROJECTION DECODE (dimension_reconstruction)
384 → 768 (teacher_align)
Detailed JSON Architecture Map
{
"sgps_qa_architecture": {
"model_name": "Semantic GPS Question Answering System",
"version": "1.0",
"architecture_type": "pyramid_compression_with_spatial_navigation",
"layer_flow": {
"input_layer": {
"dimensions": 768,
"source": "gtr-t5-base embeddings",
"active_components": ["WHAT"]
},
"compression_1": {
"input_dims": 768,
"output_dims": 384,
"transformation": "linear_compression",
"active_components": ["WHAT", "ATTENTION"]
},
"semantic_gps_hub": {
"dimensions": 384,
"description": "Semantic Intelligence Hub - The Trinity Layer",
"component_breakdown": {
"WHAT": {
"dimensions": 256,
"range": "0-255",
"type": "concept_meaning",
"description": "Core semantic embeddings"
},
"WHERE": {
"dimensions": 64,
"range": "256-319",
"type": "spatial_gps",
"description": "Coordinate positioning in semantic space"
},
"WHEN": {
"dimensions": 64,
"range": "320-383",
"type": "sequential_temporal",
"description": "Context-sensitive relational positioning"
}
},
"active_components": ["WHAT", "WHERE", "WHEN", "ATTENTION"]
},
"compression_2": {
"input_dims": 384,
"output_dims": 256,
"transformation": "linear_compression",
"component_preservation": {
"WHAT": "256→171 dims",
"WHERE": "64→43 dims",
"WHEN": "64→43 dims"
},
"active_components": ["WHAT", "WHERE", "WHEN", "ATTENTION"]
},
"compression_3": {
"input_dims": 256,
"output_dims": 192,
"transformation": "linear_compression",
"component_preservation": {
"WHAT": "171→128 dims",
"WHERE": "43→32 dims",
"WHEN": "43→32 dims"
},
"active_components": ["WHAT", "WHERE", "WHEN", "ATTENTION"]
},
"semantic_bottleneck": {
"dimensions": 192,
"description": "Maximum semantic density - All intelligence compressed",
"component_breakdown": {
"WHAT": "128 dims (66.7%)",
"WHERE": "32 dims (16.7%)",
"WHEN": "32 dims (16.7%)"
},
"active_components": ["WHAT", "WHERE", "WHEN", "ATTENTION", "PREDICTION"],
"prediction_head_connection": "192D → 512D → 10M concepts"
},
"prediction_head": {
"input_dims": 192,
"hidden_dims": [384, 512],
"output_dims": 768,
"description": "Concept vector prediction for vec2text → cloud lookup",
"architecture": [
{
"layer": "linear_1",
"input": 192,
"output": 384,
"activation": "ReLU"
},
{
"layer": "linear_2",
"input": 384,
"output": 512,
"activation": "ReLU"
},
{
"layer": "dropout",
"rate": 0.1
},
{
"layer": "linear_3",
"input": 512,
"output": 768,
"activation": "none"
}
],
"parameters": "664K params (~2.6MB)",
"downstream_flow": "768D → vec2text → cloud_concept_lookup",
"active_components": ["PREDICTION"]
},
"expansion_1": {
"input_dims": 192,
"output_dims": 256,
"transformation": "linear_expansion",
"component_restoration": {
"WHAT": "128→171 dims",
"WHERE": "32→43 dims",
"WHEN": "32→43 dims"
},
"active_components": ["WHAT", "WHERE", "WHEN", "ATTENTION"]
},
"expansion_2": {
"input_dims": 256,
"output_dims": 384,
"transformation": "linear_expansion",
"component_restoration": {
"WHAT": "171→256 dims",
"WHERE": "43→64 dims",
"WHEN": "43→64 dims"
},
"active_components": ["WHAT", "WHERE", "WHEN", "ATTENTION"]
},
"expansion_3": {
"input_dims": 384,
"output_dims": 768,
"transformation": "linear_expansion",
"component_restoration": {
"WHAT": "256→512 dims",
"WHERE": "64→128 dims",
"WHEN": "64→128 dims"
},
"active_components": ["WHAT", "WHERE", "WHEN", "ATTENTION"]
},
"output_layer": {
"dimensions": 768,
"description": "Reconstructed semantic representation",
"active_components": ["WHAT", "WHERE", "WHEN"]
}
},
"component_definitions": {
"WHAT": {
"full_name": "Concept Meaning",
"type": "semantic",
"description": "Core semantic understanding - what concepts mean",
"primary_layer": "all_layers",
"peak_layer": "semantic_gps_hub"
},
"WHERE": {
"full_name": "Spatial GPS Positioning",
"type": "spatial",
"description": "Coordinate positioning in semantic knowledge space",
"primary_layer": "semantic_gps_hub",
"range": "dimensions_256-319"
},
"WHEN": {
"full_name": "Sequential/Temporal Positioning",
"type": "temporal",
"description": "Context-sensitive relational positioning (A→B→C→D)",
"primary_layer": "semantic_gps_hub",
"range": "dimensions_320-383"
},
"ATTENTION": {
"full_name": "Multi-Head Attention",
"type": "architectural",
"description": "Attention mechanism across all processing layers",
"primary_layer": "all_processing_layers",
"excluded": ["input_layer", "output_layer", "prediction_head"]
},
"PREDICTION": {
"full_name": "Next Concept Prediction",
"type": "output",
"description": "Predicts next concept from 10M vocabulary",
"primary_layer": "prediction_head",
"connection_point": "semantic_bottleneck",
"output_flow": "768D → vec2text → cloud_lookup"
}
},
"key_innovations": {
"semantic_intelligence_hub": {
"layer": "384D",
"description": "First AI layer to integrate WHAT, WHERE, WHEN",
"breakthrough": "Trinity of conceptual understanding"
},
"context_sensitive_gps": {
"component": "WHERE + WHEN",
"description": "Concepts occupy different coordinates based on relational context",
"example": "cat-as-predator vs cat-as-pet occupy different spatial regions"
},
"spatial_navigation": {
"mechanism": "prediction_head",
"description": "Navigate through coordinate space to find answers",
"innovation": "First AI to reason through spatial movement"
},
"massive_vocabulary": {
"scale": "10M concepts",
"efficiency": "Computed only at 192D bottleneck",
"advantage": "Orders of magnitude beyond token-based systems"
}
},
"information_flow": {
"question_processing": [
"768D input (gtr-t5-base)",
"384D GPS enhancement (add WHERE/WHEN)",
"Progressive compression to 192D bottleneck",
"Spatial navigation through semantic coordinates",
"Prediction head: 192D→384D→512D→768D",
"vec2text: 768D vector → natural language",
"Cloud lookup: text → final concept from 1M-1B+ vocabulary"
],
"answer_generation": [
"Concept prediction at bottleneck",
"Vector reconstruction to 768D",
"Text generation via vec2text",
"Cloud concept database lookup",
"Final answer retrieval"
],
"cloud_architecture_benefits": [
"Model stays compact (13.5MB)",
"Unlimited vocabulary scaling",
"Live concept updates",
"Distributed inference possible"
]
},
"performance_characteristics": {
"parameters": {
"core_model": "2.85M params (10.9MB)",
"prediction_head": "0.66M params (~2.6MB)",
"total": "3.51M params (~13.5MB)"
},
"memory_usage": {
"training": "~512MB",
"inference": "~256MB"
},
"computational_complexity": {
"bottleneck_advantage": "All prediction at 192D",
"efficiency_gain": "Spatial filtering reduces search space",
"parallel_navigation": "Multiple paths explored simultaneously"
},
"cloud_architecture": {
"model_deployment": "13.5MB - fits on any device",
"vocabulary_scaling": "1M to 1B+ concepts via cloud lookup",
"inference_flow": "192D→768D→vec2text→cloud_lookup→final_concept",
"advantages": [
"Compact model with unlimited vocabulary",
"Live concept updates without model retraining",
"Distributed concept database",
"Zero vocabulary size constraints"
]
}
}
}
}
Architecture Visualization
Question Input (768D)
↓
Compression Layer 1
↓
🧠 SEMANTIC GPS HUB (384D) 🧠
┌─────────────────────────────┐
│ WHAT (256D): Core Meaning │
│ WHERE (64D): Spatial GPS │
│ WHEN (64D): Temporal Context│
└─────────────────────────────┘
↓
Compression Layer 2 (256D)
↓
🎯 BOTTLENECK (192D) 🎯
├─────────────────────┐
↓ ↓
Expansion Back 🔮 PREDICTION HEAD
(256D→384D→768D) ↓
↓ 192D→384D→512D→768D
Answer Vector ↓
(768D Output) 🌐 CLOUD ARCHITECTURE
↓
vec2text (768D→Text)
↓
Cloud Concept Lookup
(1M to 1B+ concepts)
↓
Final Answer Concept