RAG-Cognitive Routing: AI Agent Architecture

An advanced AI agent system that implements cognitive routing, debate defense, and structured social media content generation using semantic similarity, vector search, and LangGraph orchestration.

Overview

This project demonstrates a complete AI agent architecture assignment featuring:

• Semantic Routing: FAISS vector search with sentence-transformers for persona-based content routing
• LangGraph Orchestration: Multi-node workflow for topic selection, research, and content generation
• RAG-Style Defense: Memory retrieval system to maintain persona consistency against prompt injection
• Structured Outputs: Pydantic-validated LLM responses for reliable generation
• Tavily Integration: Real-time web search for current context

Architecture

graph TD
    A[Incoming Post] --> B[Semantic Router]
    B --> C{Threshold Check}
    C -->|Pass| D[Selected Persona]
    C -->|Fail| E[No Match]

    D --> F[LangGraph Workflow]
    F --> G[Decide Topic & Search]
    G --> H[Web Search via Tavily]
    H --> I[Generate Post]
    I --> J[Structured Output]

    J --> K[RAG Defense Check]
    K --> L[Memory Retrieval]
    L --> M[Consistency Validation]
    M --> N[Final Response]

Loading

Core Components

• Vector Store (agent/vector_store.py): FAISS index with normalized embeddings for cosine similarity routing
• Router (agent/router.py): Threshold-based persona matching with rich console output
• LangGraph (agent/langgraph.py): Orchestrated workflow with structured LLM outputs
• RAG Defense (agent/rag_defense.py): Memory-augmented response generation with injection resistance
• Tools (agent/tools.py): Tavily web search integration
• Schemas (agent/schemas.py): Pydantic models for type safety and validation

Features

Phase 1: Semantic Routing

• Sentence-transformers embeddings (all-MiniLM-L6-v2)
• FAISS IndexFlatIP for efficient similarity search
• Configurable routing thresholds
• Multi-persona support (Tech Maximalist, Doomer/Skeptic, Finance Bro)

Phase 2: Content Generation

• LangGraph state management
• Structured topic selection and search query generation
• Real-time web context retrieval
• Persona-consistent post generation (280 char limit)

Phase 3: Debate Defense

• RAG-style memory retrieval for persona reinforcement
• Prompt injection resistance through behavioral rules
• Debate context awareness
• Argumentative response generation

Setup

Prerequisites

• Python 3.8+
• Ollama with Llama3 model
• Tavily API key

Installation

1. Clone the repository:

git clone <repository-url>
cd rag-cognitive-routing

2. Install dependencies:

pip install -r requirements.txt

3. Set up environment variables:

cp .env.example .env
# Edit .env with your API keys

4. Start Ollama:

ollama serve
ollama pull llama3

Configuration

Create a .env file:

MODEL_NAME=llama3
TAVILY_API_KEY=your_tavily_api_key

Usage

Phase 1: Vector Routing Test

python -m agent.vector_store

Phase 2: Content Generation

python -m agent.langgraph

Phase 3: Defense Testing

python -m agent.rag_defense

Router Testing

python -m agent.router

Project Phases

Phase 1 Output

Similarity Results
| Bot ID | Bot Name          | Similarity Score |
|--------|-------------------|------------------|
| bot_a  | Tech Maximalist   | 0.4050           |
| bot_b  | Doomer / Skeptic  | 0.2330           |
| bot_c  | Finance Bro       | 0.1781           |

Routed Bots: bot_a (Threshold: 0.30)

Phase 2 Output

• Topic: AI Automation Trends
• Search Query: latest developments in AI automation 2024
• Generated Post: "AI automation is accelerating across industries. Recent breakthroughs in robotics and ML are creating unprecedented opportunities for efficiency gains..."

Phase 3 Output

• Maintains persona consistency under adversarial prompts
• Uses retrieved memory to reinforce behavioral rules
• Generates contextually appropriate debate responses

Technical Stack

• LLM: Ollama + Llama3
• Vector Search: FAISS + sentence-transformers
• Orchestration: LangGraph
• Web Search: Tavily API
• Validation: Pydantic
• CLI: Rich console

Architecture Decisions

• Structured Outputs: Ensures reliable parsing and type safety
• Normalized Embeddings: Enables cosine similarity with inner product
• Threshold Routing: Balances precision vs. coverage
• RAG Defense: Memory augmentation for consistency without full fine-tuning
• Modular Design: Clean separation of concerns for maintainability

Security Considerations

• Prompt injection defense through memory retrieval
• Input validation via Pydantic schemas
• Error handling for external API failures
• No sensitive data storage or transmission

Performance

• Embedding generation: ~50ms per text
• Vector search: ~1ms for top-3 results
• LLM generation: ~2-5 seconds per response
• Memory usage: ~500MB (embeddings + model)

Development

Testing

# Run individual modules
python -m agent.vector_store
python -m agent.router
python -m agent.langgraph
python -m agent.rag_defense

Logging

Logs are written to logs/ directory with timestamps.

Code Quality

• Type hints throughout
• Pydantic validation
• Structured error handling
• Modular architecture

Future Improvements

• Add comprehensive test suite
• Implement configuration validation
• Add monitoring and metrics
• Enhance prompt injection defense
• Support dynamic persona loading

License

MIT License - see LICENSE file for details.