๐ Multi-Agent AI Orchestration System
A comprehensive multi-agent AI orchestration platform featuring secure Retrieval-Augmented Generation (RAG), advanced research capabilities, voice interactions, MCP integration, and sophisticated memory management using Mastra AI framework.
Overview
A comprehensive multi-agent AI orchestration platform that combines secure Retrieval-Augmented Generation (RAG) with advanced research capabilities, web scraping, content analysis, voice interactions, MCP integration, and sophisticated memory management. Built with enterprise-grade security featuring hierarchical RBAC, document classification, and multi-agent security pipelines to ensure users only access authorized corporate knowledge while enabling powerful AI-driven research and analysis workflows.
Key Features
- Hierarchical RBAC: Roles inherit access (public โ employee โ dept viewer/admin โ admin)
- Document Classification: Public/internal/confidential with tag-based filtering
- Multi-Agent Orchestration: 20+ specialized agents for research, analysis, content creation, and security
- Advanced Research: Web scraping, multi-phase research workflows, and content evaluation
- Voice Integration: Google Gemini Live voice capabilities for conversational AI
- MCP Integration: Model Context Protocol support for enhanced agent capabilities
- Secure RAG: Enterprise-grade retrieval with access controls and audit trails
- Advanced Memory: Persistent memory with semantic recall and working memory templates
- Content Analysis: Learning extraction, evaluation scoring, and quality assessment
- Audit-Ready: Citations, logs, and compliance validation throughout
Architecture
flowchart TD
%% Frontend Layer
subgraph "๐จ Frontend Layer"
A[Next.js App Router<br/>React 19 + TypeScript]
B[Cedar OS Integration<br/>Product Roadmap UI]
C[React Components<br/>Chat, Auth, Indexing]
D[Shared Libraries<br/>JWT, MDX, Hooks]
end
%% Backend Layer
subgraph "โ๏ธ Backend Layer"
E[Mastra Core<br/>Orchestration Engine]
F[AI Agents<br/>27+ Specialized Agents<br/>5 Domains]
G[Workflows<br/>10+ Multi-step Orchestration<br/>Complex Tasks]
H[Tools & Services<br/>50+ Tools<br/>Business Logic]
I[API Routes<br/>Chat & Indexing<br/>Streaming]
end
%% Data Layer
subgraph "๐พ Data Layer"
J[(PostgreSQL + PgVector<br/>Vector Embeddings<br/>Document Storage)]
K[(Content Corpus<br/>Markdown Files<br/>RAG Source)]
end
%% External Services
subgraph "๐ External Services"
L[Google Gemini<br/>AI Models]
M[OpenAI<br/>LLM Services]
N[Data Sources<br/>Alpha Vantage, Finnhub<br/>SerpAPI, Polygon]
O[Academic Sources<br/>ArXiv, Research APIs]
end
%% Agent Categories (detailed breakdown)
subgraph "๐ค Agent Categories"
P1[Governed RAG Agents<br/>identity, retrieve, rerank<br/>answerer, verifier]
P2[Domain Agents<br/>research, copywriter<br/>cryptoAnalysis, compliance]
P3[Specialized Agents<br/>productRoadmap, salesIntelligence<br/>stockAnalysis, voiceAgent]
P4[Orchestration Agents<br/>a2aCoordinator, assistant<br/>editor, evaluation]
end
%% Workflow Categories
subgraph "๐ Workflow Categories"
Q1[Chat Workflows<br/>Multiple Versions<br/>Real-time Conversation]
Q2[Analysis Workflows<br/>Financial Analysis V3<br/>Research Workflow]
Q3[Content Workflows<br/>Content Generation<br/>Report Generation]
Q4[RAG Workflows<br/>Governed RAG Answer<br/>Governed RAG Index]
end
%% Tool Categories
subgraph "๐ ๏ธ Tool Categories"
R1[Data Tools<br/>Alpha Vantage, Finnhub<br/>Polygon, SerpAPI]
R2[Processing Tools<br/>Document Chunking<br/>PDF Conversion, Web Scraper]
R3[Analysis Tools<br/>Competitive Intelligence<br/>Compliance Check, Sales Analysis]
R4[Specialized Tools<br/>JWT Auth, Vector Query<br/>Graph RAG, Roadmap Tool]
end
%% Connections
A --> I
B --> I
C --> I
D --> I
I --> E
E --> F
F --> G
G --> H
H --> J
H --> K
J --> H
K --> H
F --> L
F --> M
G --> L
G --> M
H --> N
H --> O
F --> P1
F --> P2
F --> P3
F --> P4
G --> Q1
G --> Q2
G --> Q3
G --> Q4
H --> R1
H --> R2
H --> R3
H --> R4
%% Styling for Dark Mode
classDef darkMode fill:#1e1e1e,stroke:#ffffff,stroke-width:2px,color:#ffffff
class A,B,C,D,E,F,G,H,I,J,K,L,M,N,O,P,Q,R darkMode
class P1,P2,P3,P4 darkMode
class Q1,Q2,Q3,Q4 darkMode
class R1,R2,R3,R4 darkMode
class sys,web,api,db darkMode
class sys,web,api,db,auth,sec darkMode
class auth,policy,retrieve,rerank,answerer,verifier,research,report,copywriter,evaluation,learning,product,editor,self,assistant,starter,voice,mcp,template,network,rag,content darkMode
Research & Analysis Pipeline:
flowchart TD
%% Research Input
A[Research Query] --> B[Research Agent<br/>Query Decomposition<br/>Multi-phase Planning]
%% Initial Research Phase
B --> C[Web Scraping Tools<br/>Site Mapping, Content Extraction<br/>Batch Processing]
%% Analysis Phase
C --> D[Content Analysis<br/>Learning Extraction<br/>Evaluation Scoring]
%% Follow-up Research
D --> E[Iterative Research<br/>Follow-up Queries<br/>Depth Analysis]
%% Synthesis Phase
E --> F[Content Synthesis<br/>Report Generation<br/>Validation]
%% Output
F --> G[Research Report<br/>With Citations & Sources]
%% Supporting Tools
subgraph H["Research Tools"]
I[Web Scraper<br/>Content Extraction]
J[Evaluator<br/>Quality Assessment]
K[Learning Extractor<br/>Insight Mining]
L[Graph RAG<br/>Knowledge Graph]
end
%% Tool Integration
B -.-> I
C -.-> J
D -.-> K
E -.-> L
%% Styling for Dark Mode
classDef darkMode fill:#1e1e1e,stroke:#ffffff,stroke-width:2px,color:#ffffff
class A,B,C,D,E,F,G,H darkMode
Secure RAG Pipeline:
flowchart TD
%% Input Layer
A[User Query + JWT] --> B[Identity Agent<br/>JWT Validation & User Context]
%% Security Layer
B --> C[Policy Agent<br/>Role-Based Access Control<br/>Document Classification Filtering]
%% Retrieval Layer
C --> D[Retrieve Agent<br/>PgVector Similarity Search<br/>Security-Filtered Results]
%% Processing Layer
D --> E[Rerank Agent<br/>Relevance Scoring<br/>Context Ranking]
%% Generation Layer
E --> F[Answerer Agent<br/>Secure Response Generation<br/>Citation Assembly]
%% Verification Layer
F --> G[Verifier Agent<br/>Compliance Validation<br/>PII Detection<br/>Policy Enforcement]
%% Output Layer
G --> H[Secure Response<br/>With Citations & Metadata]
%% Multi-Agent Enhancement
I[Research Content Network<br/>Multi-Source Analysis] -.-> F
J[Governed RAG Network<br/>Cross-Agent Coordination] -.-> G
%% Supporting Components
subgraph K["Available Tools (12 total)"]
L[Vector Query Tool<br/>Secure Similarity Search]
M[Document Chunking Tool<br/>Intelligent Text Splitting]
N[Web Scraper Tool<br/>Content Extraction]
O[JWT Auth Tool<br/>Token Validation]
P[Content Tools<br/>Copywriter, Editor, Evaluation]
Q[Data Tools<br/>Weather, Roadmap, File Manager]
end
%% Tool Usage
D -.-> L
D -.-> M
F -.-> P
I -.-> N
I -.-> Q
B -.-> O
%% Styling for Dark Mode
classDef darkMode fill:#1e1e1e,stroke:#ffffff,stroke-width:2px,color:#ffffff
class A,B,C,D,E,F,G,H,I,J,K,L,M,N,O,P,Q darkMode
Getting Started
Prerequisites
- Node.js >=20.9.0
- Docker and Docker Compose
- Git
- OpenAI API key
Quick Setup
-
Clone and install:
git clone https://github.com/ssdeanx/governed-rag-ai.git cd mastra-governed-rag npm install -
Configure environment:
cp .env.example .env # Edit .env with your API keys and other settings: # - OPENAI_API_KEY or GOOGLE_GENERATIVE_AI_API_KEY (required for AI models) # - SERPAPI_API_KEY (optional, for web search tools - get from https://serpapi.com/manage-api-key) # - SUPABASE credentials (required for authentication) # - DATABASE_URL (required for PostgreSQL) -
Set up authentication (Supabase):
This application uses Supabase for user authentication. You need to:
a. Create a Supabase project at https://supabase.com/dashboard
b. Get your credentials from Project Settings:
SUPABASE_URL- Your project URL (e.g.,https://xxxxx.supabase.co)SUPABASE_ANON_KEY- Your anon/public key
c. Update your
.envfile with Supabase credentials:SUPABASE_URL="https://your-project.supabase.co" SUPABASE_ANON_KEY="your_anon_key_here" NEXT_PUBLIC_SUPABASE_URL="https://your-project.supabase.co" NEXT_PUBLIC_SUPABASE_ANON_KEY="your_anon_key_here"d. Configure GitHub OAuth (optional):
- In Supabase Dashboard: Authentication > Providers > GitHub
- Add your GitHub OAuth app credentials
- Callback URL:
https://your-project.supabase.co/auth/v1/callback - Update
.envwithGITHUB_CLIENT_IDandGITHUB_CLIENT_SECRET
๐ See docs/AUTH_ARCHITECTURE.md for detailed authentication documentation.
-
Start database (optional):
# Only needed if using local PostgreSQL docker-compose up -d db -
Index documents:
npm run cli index -
Start development:
npm run dev # http://localhost:3000
AI & Research Capabilities
- Multi-Agent Orchestration: 20+ specialized agents for research, analysis, content creation, and security
- Advanced Research Workflows: Multi-phase research with web scraping, content evaluation, and learning extraction
- Secure Vector Search: Filtered retrieval based on user permissions with PgVector
- Content Analysis: Automated learning extraction, quality evaluation, and insight mining
- Web Search & Data Collection: SerpAPI-powered Google Search, News, Trends, Shopping (Amazon, Walmart, eBay, Home Depot), Academic (Google Scholar), Finance, and Local Business (Yelp) searches
- Web Scraping Tools: Comprehensive web content extraction and processing
- Contextual Reranking: Relevance scoring with security constraints
- Citation Generation: Source attribution with access verification
- Voice Integration: Google Gemini Live voice capabilities for conversational AI
- MCP Integration: Model Context Protocol support for enhanced agent tools
- Advanced Memory: Persistent memory with semantic recall and working memory templates
- Quality Assurance: Automated evaluation scoring and compliance validation
- Streaming Responses: Real-time answer generation with SSE
AI Agents
This system employs a comprehensive multi-agent architecture with 20+ specialized agents:
Core RAG Agents
- Identity Agent - Validates user authentication and permissions
- Policy Agent - Enforces access control and security policies
- Retrieve Agent - Performs intelligent document retrieval with security filtering
- Rerank Agent - Ranks and scores retrieved documents for relevance
- Answerer Agent - Generates secure responses with citations
- Verifier Agent - Validates responses for compliance and accuracy
Specialized Agents
- Research Agent - Conducts in-depth research and analysis
- Report Agent - Generates structured reports and summaries
- Copywriter Agent - Creates marketing and communication content
- Evaluation Agent - Assesses content quality and relevance
- Learning Extraction Agent - Identifies and extracts key learnings
- Product Roadmap Agent - Analyzes product strategy and roadmaps
- Editor Agent - Reviews and improves content quality
- Self-Referencing Agent - Maintains context across conversations
- Assistant Agent - Provides general AI assistance
- Starter Agent - Handles initial query processing
- Voice Agent - Manages voice-based interactions and audio processing
- MCP Agent - Integrates with Model Context Protocol tools
- Template Reviewer Agent - Reviews and validates templates and workflows
Network Agents
- Research Content Network - Orchestrates multi-agent research workflows
- Governed RAG Network - Manages secure RAG operations across agents
Each agent follows a single-tool-call policy, ensuring predictable and auditable AI behavior while maintaining security governance throughout the entire pipeline.
Research & Analysis Features
Beyond secure RAG, this system provides comprehensive research and analysis capabilities:
Multi-Phase Research Workflows
- Query Decomposition: Breaking down complex research questions into focused search queries
- Web Content Extraction: Advanced web scraping with site mapping and content cleaning
- Content Evaluation: Automated assessment of information quality and relevance
- Learning Extraction: AI-powered identification and extraction of key insights
- Iterative Research: Follow-up research based on initial findings
- Report Synthesis: Structured report generation with citations and sources
Advanced Analysis Tools
- Graph RAG Queries: Complex knowledge graph-based retrieval and analysis
- Content Quality Scoring: Automated evaluation using LLM-based scorers
- Semantic Memory: Persistent context with working memory templates
- Compliance Validation: PII detection and policy enforcement
- Multi-Source Synthesis: Combining information from diverse sources
Research Agent Example
// Research agent with multi-phase workflow
const researchAgent = new Agent({
name: 'Research Agent',
instructions: `
PHASE 1: Deconstruct topic into focused search queries
PHASE 2: Web scraping and content extraction
PHASE 3: Content evaluation and learning extraction
PHASE 4: Follow-up research and synthesis
`,
tools: {
webScraperTool, evaluateResultTool, extractLearningsTool,
graphRagQueryTool, mdocumentChunker
},
memory: pgMemory,
scorers: {
relevancy: createAnswerRelevancyScorer(),
safety: createToxicityScorer()
}
})
AI Tools
The system includes 12 specialized tools that agents can invoke to perform specific operations:
Core RAG Tools
- Vector Query Tool - Performs secure vector similarity searches with access filtering
- JWT Auth Tool - Validates and processes JWT tokens for user authentication
- Document Chunking Tool - Intelligently splits documents into manageable chunks for indexing with metadata extraction
- Graph RAG Query Tool - Executes complex graph-based retrieval augmented generation queries
Content & Analysis Tools
- Copywriter Agent Tool - Generates marketing and communication content
- Editor Agent Tool - Reviews and improves content quality and clarity
- Evaluate Result Tool - Assesses the quality and relevance of generated content
- Extract Learnings Tool - Identifies and extracts key insights and learnings
- Starter Agent Tool - Handles initial query processing and routing
Research & Data Tools
- Web Scraper Tool - Extracts and processes web content for research
- Weather Tool - Provides weather data and forecasting capabilities
- Roadmap Tool - Analyzes product strategy and roadmap information
- Data File Manager - Manages file operations and data processing tasks
Core RAG Workflows
- Governed RAG Answer Workflow - Main workflow for secure question answering with citations
- Governed RAG Index Workflow - Handles document indexing with classification and security tagging
- Chat Workflow - Manages conversational interactions with context preservation
Specialized Workflows
- Research Workflow - Conducts comprehensive research operations across multiple sources
- Content Generation Workflow - Creates various types of content using multiple agents
- Generate Report Workflow - Produces structured reports and analytical summaries
- Template Reviewer Workflow - Reviews and validates templates and workflows with claim extraction, planning, and scoring
Supporting Workflows
- Chat Workflow Types - Type definitions and utilities for chat operations
- Chat Workflow Shared Types - Common types and interfaces for workflow communication
All workflows implement comprehensive error handling, tracing, and security validation at each step, ensuring reliable and auditable AI operations.
AI Services
The system includes 13 specialized services that handle business logic and integrations:
Core Services
- Authentication Service - Manages user authentication and session handling
- Role Service - Handles role-based access control and permissions
- Validation Service - Provides data validation and sanitization
- Rate Limiting Service - Implements API rate limiting and throttling
Document & Content Services
- Document Indexing Service - Manages document indexing and metadata
- Document Processor Service - Processes and transforms documents
- Chunking Service - Handles intelligent text chunking strategies
- Embedding Service - Manages vector embeddings generation
Data & Storage Services
- Vector Query Service - Provides vector similarity search capabilities
- Vector Storage Service - Manages vector database operations
- Tier Management Service - Handles subscription tiers and feature access
Workflow Services
- Workflow Decorators - Provides decorators for workflow enhancement
Data Schemas
The system uses Zod schemas for comprehensive data validation:
- Agent Schemas - Validation schemas for all agent inputs and outputs
- Workflow Schemas - Type-safe workflow definitions and contracts
- API Schemas - Request/response validation for all endpoints
Access Control Policy
Security policies are defined in src/mastra/policy/acl.yaml with hierarchical role-based access:
Role-Based Access Control
admin (100) > dept_admin (80) > dept_viewer (60) > employee (40) > public (10)
Document Access Control
docs:
- path: './corpus/finance-policy.md'
allow:
roles: ['finance.viewer', 'finance.admin']
tiers: ['pro', 'enterprise']
tenant: 'acme'
classification: 'internal'
- path: './corpus/hr-confidential.md'
allow:
roles: ['hr.admin', 'admin']
tiers: ['enterprise']
tenant: 'acme'
classification: 'confidential'
Feature Tiers
-
Free: Basic RAG capabilities, public docs access
-
Pro: Internal docs, advanced analytics, custom integrations
-
Enterprise: Confidential docs, white-label, on-premise deployment
-
Type-Safe Development: Full TypeScript with Zod schema validation
-
Hot Reload: Concurrent development for frontend and backend
-
Comprehensive Testing: Vitest framework with service and workflow tests
-
Docker Integration: Containerized development environment
-
Extensive Documentation: Complete API references and guides
Project Structure
mastra-governed-rag/
โโโ app/ # Next.js routes and API endpoints
โโโ components/ # React UI components
โโโ cedar/ # Cedar OS interactive components
โโโ corpus/ # Sample documents for indexing
โโโ docs/ # Comprehensive documentation
โโโ src/
โ โโโ mastra/ # Core Mastra implementation
โ โ โโโ agents/ # AI agents (20+ specialized agents for research, analysis, security)
โ โ โโโ workflows/ # Orchestrated agent workflows (10 workflows for research, RAG, content)
โ โ โโโ tools/ # Reusable agent tools (12 tools for web scraping, analysis, search)
โ โ โโโ services/ # Business logic and integrations (13 services)
โ โ โโโ networks/ # Multi-agent orchestration networks
โ โ โโโ schemas/ # Data validation schemas (Zod)
โ โ โโโ config/ # Configuration and external services
โ โ โโโ policy/ # Access control policies and ACL
โ โโโ cli/ # Command-line interface
โโโ lib/ # Shared utilities and client libraries
โโโ docker-compose.yml # Development services
Usage Examples
Research Query with Multi-Phase Analysis
// Advanced research with web scraping and analysis
const researchRequest = {
question: "What are the latest developments in AI agent orchestration frameworks?",
researchDepth: 3, // Multi-phase research
includeWebSources: true,
evaluationRequired: true
}
// Response includes:
// - Multi-source web research
// - Content evaluation scores
// - Extracted learnings and insights
// - Citations and source validation
// - Structured synthesis report
Secure RAG Query with Authentication
// Generate JWT for a finance viewer
const jwt = generateDemoJWT('finance')
// Query the system
const response = await fetch('/api/chat', {
method: 'POST',
headers: { 'Content-Type': 'application/json' },
body: JSON.stringify({
jwt,
question: 'What is the expense reimbursement policy?',
}),
})
const result = await response.json()
// Returns secure answer with citations
Role-Based Access Demonstration
# Test different user roles
npm run cli query "$(npm run jwt:finance)" "What are expense approval thresholds?"
npm run cli query "$(npm run jwt:hr)" "What is executive compensation policy?"
npm run cli query "$(npm run jwt:public)" "What is our company mission?"
[!NOTE] Each role sees different results based on their access level. See Demo Roles for complete examples.
API Reference
Core Endpoints
POST /api/chat- Secure RAG queries with streaming responsesPOST /api/index- Document indexing with classificationGET /api/auth/*- Authentication endpoints
Request/Response Format
// Chat request
{
"jwt": "eyJhbGciOiJIUzI1NiIs...",
"question": "What is the company policy on X?"
}
// Streaming response
data: {"content": "According to the policy..."}
data: {"done": true, "citations": [...]}
For complete API documentation, see API Reference.
Development
Local Development
# Start all services
npm run dev # Frontend + Mastra backend
# Individual services
npm run dev:next # Next.js only
npm run dev:mastra # Mastra only
Testing
# Run all tests
npm test
# CLI operations
npm run cli index # Index documents
npm run cli query # Test queries
npm run cli demo # Interactive demo
Code Quality
# Linting and formatting
npm run lint
npm run pretty
Deployment
Docker Deployment
# Build and run
docker-compose up -d
Production Considerations
- Configure production environment variables
- Set up proper JWT secrets
- Enable audit logging
- Configure PostgreSQL + PgVector for production scale
Security Model
Role Hierarchy
admin (100) > dept_admin (80) > dept_viewer (60) > employee (40) > public (10)
Document Classifications
- Public: General information accessible to all
- Internal: Department-specific content for employees
- Confidential: Highly sensitive data requiring elevated access
Access Control
Policies are defined in src/mastra/policy/acl.yaml with role-based and tier-based access control:
docs:
- path: './corpus/finance-policy.md'
allow:
roles: ['finance.viewer', 'finance.admin']
tiers: ['pro', 'enterprise']
tenant: 'acme'
classification: 'internal'
- path: './corpus/hr-confidential.md'
allow:
roles: ['hr.admin', 'admin']
tiers: ['enterprise']
tenant: 'acme'
classification: 'confidential'
Feature Access Tiers
- Free: Basic RAG, public docs
- Pro: Internal docs, advanced analytics, custom integrations
- Enterprise: Confidential docs, white-label, on-premise deployment
Resources
- Full Documentation
- Architecture Guide
- Security Implementation
- API Reference
- Mastra Framework
- PostgreSQL Documentation
Troubleshooting
Common Issues
PostgreSQL Connection Failed
# Check PostgreSQL status
docker-compose ps
# Check logs
docker-compose logs db
# Restart database
docker-compose down && docker-compose up -d
Authentication Errors
# Regenerate JWT tokens
npm run jwt:finance # or other roles
No Search Results
- Verify document indexing:
npm run cli index - Check user role permissions
- Review PgVector collection status
For detailed troubleshooting, see Troubleshooting Guide.
Roadmap
- Multi-tenant support
- Advanced reranking algorithms
- Integration with additional LLM providers
- Enhanced audit and compliance features
- Performance optimizations for large document sets
Built with โค๏ธ using Mastra โข Next.js โข PostgreSQL
Contributing
We welcome contributions! Please see our Contributing Guide for details.
License
This project is licensed under the MIT License - see the LICENSE file for details.
Built with โค๏ธ by Mastra Community. Questions? Open an issue.
