Skip to content

System Overview & Architecture

Executive Summary

The Hybrid GraphRAG platform is an enterprise-grade Life Sciences artificial intelligence architecture engineered for Computer System Validation (CSV), Computer Software Assurance (CSA), Quality Management Systems (QMS), and Asset/CMMS Governance.

The system enforces strict compliance with international regulatory standards:

  • United States FDA 21 CFR Part 11 (Electronic Records and Electronic Signatures)
  • European Union EudraLex Volume 4, Annex 11 (Computerised Systems)
  • ISPE GAMP 5 Second Edition (Risk-Based Approach to Compliant GxP Computerized Systems)
  • International Organization for Standardization ISO 13485:2016 (Medical Devices Quality Management Systems)
  • United States FDA Quality Management System Regulation (QMSR, 21 CFR Part 820)
  • ALCOA+ Data Integrity Principles (Attributable, Legible, Contemporaneous, Original, Accurate, Complete, Consistent, Enduring, Available)

Architectural Blueprint

+---------------------------------------------------------------------------------------+
| CLIENT ACCESS & CONSOLIDATED SUITE |
| |
| GxPChat Unified Application (Next.js 16 + React 19 + Tailwind v4 + shadcn/ui) |
| |-- AI Assistant Console (/) |
| |-- Enterprise QMS (/qms) |
| |-- Physical Asset Management (/assets) |
| |-- IT Help Desk & GxP Service Desk (/helpdesk) |
| |-- IT Computer System Validation (/validation) |
| `-- GxP Backup, Disaster Recovery & SOP Qualification (/backup-restore) |
+-------------------------------------------+-------------------------------------------+
|
v
+---------------------------------------------------------------------------------------+
| API GATEWAY & PROTOCOLS |
| |
| FastAPI Backend Service (:8090) |
| |-- FastMCP Server (Server-Sent Events: /mcp/sse) |
| |-- Reverse Proxy Integration from GxPChat |
| `-- REST API Endpoints (/api/agents/*, /api/storage/*, /api/demos/*) |
+-------------------------------------------+-------------------------------------------+
|
v
+---------------------------------------------------------------------------------------+
| MULTI-AGENT ORCHESTRATION LAYER |
| |
| Lead CSV Orchestrator (Virtual Quality Board) |
| |-- 23 Specialized Pydantic AI Agents (Dependency Injection: GxpAgentDeps) |
| |-- Autonomous Squad Missions (Missions 1 through 8) |
| `-- EvidenceGate Pre-Flight Verification & Live Supersession Checking |
+-------------------------------------------+-------------------------------------------+
|
v
+---------------------------------------------------------------------------------------+
| 3-DATABASE DIVISION OF LABOR |
| |
| +--------------------------+ +--------------------------+ +---------------------+ |
| | POSTGRESQL | | QDRANT | | MEMGRAPH | |
| | Canonical Relational SoR | | Semantic Vector Search | | Structural KG C++ | |
| | 21 CFR Part 11 Triggers | | Dense + Sparse Splade | | Multi-Hop Paths | |
| | Immutable Audit Ledger | | Sub-5ms ANN Latency | | Deep Traversal | |
| | Port: 5434 | | Port: 6335 | | Port: 7688/Lab:3006 | |
| +--------------------------+ +--------------------------+ +---------------------+ |
+---------------------------------------------------------------------------------------+
|
v
+---------------------------------------------------------------------------------------+
| STORAGE, OBSERVABILITY & MICROSERVICES |
| |
| +--------------------------+ +--------------------------+ +---------------------+ |
| | RUSTFS S3 | | LANGFUSE V4 STACK | | DOCX SERVICE | |
| | S3-Compatible Storage | | Web + Worker + ClickHouse| | ASP.NET Core API | |
| | File Drop Watcher Engine | | Redis + MinIO + Tracing | | Formal Word Exports | |
| | Port: 9120 / UI: 9121 | | Port: 3001 | | Port: 5055 | |
| +--------------------------+ +--------------------------+ +---------------------+ |
+---------------------------------------------------------------------------------------+

Core Architectural Pillars

1. Three-Database Division of Labor

The platform avoids the compromises of single-database architectures by enforcing a strict division of labor across three specialized database engines:

  • PostgreSQL 16: Functions as the authoritative System of Record (SoR) and regulatory ledger. Enforces 21 CFR Part 11 append-only immutability via database triggers that physically reject UPDATE and DELETE operations on audit trails, signatures, and backup execution tables.
  • Qdrant Vector Database (v1.13.4): Executes dense and sparse vector similarity matching with sub-5 millisecond retrieval latency. Employs prefetching and Reciprocal Rank Fusion (RRF) to retrieve mathematically relevant text chunks across controlled document vaults.
  • Memgraph MAGE (In-Memory Graph Database): Models relational linkages as a Labeled Property Graph (LPG). Executes multi-hop Cypher queries to evaluate ripple-effect impacts across User Requirements, Functional Specifications, Risk Assessments, Assets, and Deviations.

2. Zero-Trust EvidenceGate

Retrieved vector and graph candidates are not injected blindly into large language model context windows. Instead, all candidates pass through the zero-trust EvidenceGate engine:

  • Verifies that referenced chunk hashes match the live PostgreSQL canonical record.
  • Asserts that the document status is currently effective and has not been superseded or withdrawn.
  • Validates the presence of valid 21 CFR Part 11 cryptographic digital signatures.
  • Drops invalid or unverified chunks prior to model prompt assembly.

3. Specialized Pydantic AI Agent Architecture

Rather than relying on a general-purpose conversational agent, the platform deploys 23 purpose-built Pydantic AI agents coordinated by a Lead CSV Orchestrator. Each agent:

  • Is scoped to a single regulatory or technical responsibility (for example, Traceability, Risk Assessment, Deviation Root Cause Analysis, or Disaster Recovery).
  • Receives strongly typed runtime dependencies via RunContext[GxpAgentDeps].
  • Employs strict Pydantic output validation to ensure that all generated deliverables conform to required data contracts.
  • Connects to frontend interfaces through Model Context Protocol (FastMCP) Server-Sent Events.

4. Consolidated User Application (GxPChat)

All operational applications are united within GxPChat, a Next.js 16 web suite offering:

  • Natural language conversational assistant with smart autocomplete triggers (# action commands and @ entity mentions).
  • Enterprise Quality Management System (QMS) with dual Operational and Auditor Inspection modes.
  • Physical Asset and Metrology Management across 202 enterprise instruments and skids.
  • IT Service Desk with dual-track segregation of standard IT tickets from GxP incidents.
  • Computer System Validation hub with live V-Model matrix, CSA risk calculator, and test execution modals.
  • GxP Backup and Disaster Recovery console governing SOP-IT-018 automated cloud backups and SOP-IT-019 manual dual-custody executions.