Solace-AI/project-intent-and-objectives.txt at main · Rayyan9477/Solace-AI · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
  Solace-AI System Architecture

  Project Overview

  Solace-AI is an advanced AI-powered mental health companion backend platform that provides:
  - Evidence-based therapeutic interventions (CBT, DBT, ACT, Mindfulness, MI)
  - Dynamic mental health diagnosis using multi-agent systems inspired by Google DeepMind's AMIE
  - Personality-aware communication using Big Five (OCEAN) traits
  - Persistent therapeutic context across sessions via sophisticated memory management

  ---
  Core Architecture Components (5 Modules)

  | Module                 | Purpose                       | Key Features                                                                        |
  |------------------------|-------------------------------|-------------------------------------------------------------------------------------|
  | 00-System Integration  | Platform orchestration        | Multi-agent LangGraph, Safety-first, HIPAA/SOC2 compliance, Event-driven + Sync API |
  | 01-Diagnosis & Insight | Mental health assessment      | AMIE-inspired 4-agent system, Chain-of-Reasoning, Anti-sycophancy, DSM-5-TR + HiTOP |
  | 02-Therapy             | Evidence-based interventions  | Hybrid Rules+LLM, 5-component framework, Stepped Care (4 levels), CBT/DBT/ACT/MI    |
  | 03-Personality         | Communication personalization | Big Five OCEAN detection, MoEL empathy, Style adaptation (r>0.85)                   |
  | 04-Memory & Context    | Persistent context            | 5-tier cognitive hierarchy, Agentic RAG, Ebbinghaus decay with safety override      |

  ---
  Key Architectural Decisions

  1. Multi-Agent Orchestration (LangGraph)

  Priority 0: Safety Agent (Always monitors, can override any agent)
  Priority 1: Supervisor Agent (Routes, coordinates, quality control)
  Priority 2: Clinical Agents (Diagnosis, Therapy, Assessment)
  Priority 3: Support Agents (Personality, Emotion, Chat)

  2. Diagnosis Module - 4-Step Chain-of-Reasoning

  1. ANALYZE - Extract symptoms, timeline, context
  2. HYPOTHESIZE - Generate differential diagnosis (DSM-5-TR + HiTOP)
  3. CHALLENGE - Devil's Advocate agent prevents confirmation bias
  4. SYNTHESIZE - Calibrate confidence, generate response

  3. Therapy Module - Hybrid Architecture

  - Rules-based: Treatment protocols, session structure, safety guardrails
  - LLM-generated: Empathic responses, personalized examples, Socratic questions
  - Stepped Care: PHQ-9 severity determines intensity (0-4 levels)
  - Techniques: CBT, DBT (4 modules), ACT (6 processes), MI (OARS), Mindfulness

  4. Personality Module - Big Five Detection

  - OCEAN: Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism
  - Detection: Ensemble (Fine-tuned RoBERTa + LLM zero-shot + LIWC)
  - Multimodal: Text + Voice + Behavioral signals (Late Fusion)
  - Empathy: MoEL (Mixture of Empathetic Listeners) with 32 emotion classes

  5. Memory Module - 5-Tier Cognitive Hierarchy

  | Tier | Name            | Storage             | Access Time |
  |------|-----------------|---------------------|-------------|
  | 1    | Input Buffer    | In-memory           | <1ms        |
  | 2    | Working Memory  | Redis               | <10ms       |
  | 3    | Session Memory  | Redis+PostgreSQL    | <50ms       |
  | 4    | Episodic Memory | PostgreSQL+Weaviate | <200ms      |
  | 5    | Semantic Memory | Weaviate+PostgreSQL | Permanent   |

  ---
  Safety Architecture - 4 Layers

  1. Layer 1: Input Gate (<10ms) - Crisis keyword detection
  2. Layer 2: Processing Guard (<100ms) - Contraindication checking
  3. Layer 3: Output Filter (<500ms) - Response safety validation
  4. Layer 4: Continuous Monitor - Session trajectory tracking

  Crisis Levels (C-SSRS Mapping):
  - 🟢 Level 0-1: Continue/Check-in
  - 🟡 Level 2-3: Safety assessment/planning
  - 🔴 Level 4-5: Crisis protocol (988 Lifeline, stay engaged, alert clinician)

  ---
  Technology Stack

  | Layer         | Technology                                |
  |---------------|-------------------------------------------|
  | Orchestration | LangGraph (Multi-agent)                   |
  | Backend       | Python/FastAPI                            |
  | LLM           | Anthropic Claude/OpenAI GPT/Google Gemini |
  | Vector DB     | Weaviate (Hybrid BM25 + Semantic)         |
  | Relational DB | PostgreSQL                                |
  | Cache         | Redis Cluster                             |
  | Events        | Kafka                                     |
  | Container     | Kubernetes                                |
  | Service Mesh  | Istio (mTLS)                              |
  | Compliance    | HIPAA, SOC2, Zero Trust                   |

  ---
  Data Flow Summary

  User Input → Safety Gate → Orchestrator → [Memory Context Load]
                                         → [Parallel: Diagnosis + Therapy + Personality]
                                         → Response Aggregation
                                         → Personality Style Application
                                         → Safety Post-check
                                         → [Memory Update]
                                         → Final Response

  ---
  Key Design Patterns Used

  1. Anti-Sycophancy Framework - Devil's Advocate agent challenges every diagnosis
  2. Sample Consistency Confidence - Multiple LLM samples to calibrate confidence
  3. Agentic Corrective RAG - Self-correcting retrieval with document grading
  4. Ebbinghaus Decay + Safety Override - Natural memory decay, but crisis info NEVER decays
  5. Hybrid Content Generation - Rules for clinical fidelity + LLM for conversational warmth
  6. Stepped Care Routing - Severity-appropriate treatment intensity
  7. MoEL Empathy - 32 specialized emotion listeners with soft combination