System Architecture

Understanding the Verdana AI agent architecture with PostgreSQL vector storage, OpenAI embeddings, and web verification powering the EU Green Policies Chatbot

Verdana AI Agent Architecture Overview

My system uses Verdana, an intelligent AI agent that combines query classification, automatic language detection, PostgreSQL vector search, real-time web verification, and OpenAI Whisper speech processing to deliver accurate, context-aware responses about EU Green Deal policies.

PostgreSQL + pgvector

High-performance vector database storing 3072-dimensional OpenAI embeddings of 24+ official EU policy documents for semantic search.

OpenAI Integration

Uses GPT-4 for response generation, text-embedding-3-large for vectorization, and Whisper API for speech-to-text processing.

Verdana AI Agent

Intelligent agent with query classification, language detection, conversation context, and proactive web search capabilities.

Comprehensive Knowledge Base

The system contains 24+ official EU policy documents, processed into 800-token chunks with 300-token overlap, embedded using OpenAI text-embedding-3-large for precise semantic retrieval of relevant policy information.

Official Sources

All documents sourced directly from European Commission official publications and legal databases.

Vector Search

Uses cosine similarity on 3072-dimensional embeddings with 0.3 threshold to find the most relevant policy content for each query.

Semantic Understanding

Advanced semantic search understands context and intent, not just keyword matching.

Verdana Agent Processing Workflow

Verdana handles the complete interaction lifecycle with intelligent query classification, automatic language detection, conversation context awareness, vector search, real-time web verification, and comprehensive source attribution.

1. Query Classification

Verdana classifies queries as casual conversation, identity, or EU Green Deal policy queries and detects user language.

2. Vector Search

PostgreSQL pgvector cosine similarity search using OpenAI embeddings to find relevant EU policy content.

3. Web Verification

Dual web search via Tavily API - EU domain-restricted search plus broader policy research for comprehensive coverage.

4. Context Integration

Verdana integrates conversation history, vector search results, and web verification data for context-aware responses.

5. Source Attribution

Response delivered with comprehensive source attribution, relevance scores, and deduplicated reference list.

Unified Chat Agent Architecture

This streamlined architecture uses a single intelligent agent that combines all necessary capabilities in one efficient system, eliminating the complexity of multiple agent coordination.

Unified Chat Agent

✅ Active - Production System

The unified agent consolidates all AI capabilities into a single, efficient system that handles the complete interaction lifecycle. It processes voice input via local Whisper, performs RAG retrieval, conducts web verification, and generates comprehensive responses all within one streamlined workflow.

Speech Processing

  • • Local Whisper transcription
  • • Multi-language support
  • • High-quality audio processing
  • • Real-time conversion

RAG Retrieval

  • • Semantic document search
  • • Vector similarity matching
  • • Context-aware retrieval
  • • Source ranking & scoring

Web Verification

  • • Real-time source checking
  • • EU official site priority
  • • Policy update detection
  • • Accuracy validation

Response Generation

  • • Comprehensive synthesis
  • • Source attribution
  • • Confidence scoring
  • • EU AI Act compliance

Architecture Benefits

  • Simplified Architecture: Single agent reduces complexity
  • Faster Processing: No inter-agent communication overhead
  • Better Consistency: Unified decision-making process
  • Enhanced Reliability: Fewer points of failure
  • Easier Maintenance: Single codebase to manage
  • Cost Efficiency: Reduced resource requirements

Technical Stack & Infrastructure

Database Layer

  • • PostgreSQL 15+ with pgvector
  • • Vector similarity search
  • • 1536-dimensional embeddings
  • • Efficient indexing & caching

AI & ML Services

  • • OpenAI GPT-4o-mini
  • • OpenAI Whisper speech-to-text
  • • text-embedding-3-large
  • • Tavily search API

Backend Services

  • • FastAPI with async support
  • • LangChain framework
  • • WebSocket real-time chat
  • • Structured logging

Local Whisper Speech Processing

The system uses OpenAI Whisper deployed locally for high-quality, privacy-focused speech-to-text processing. Audio is processed server-side ensuring consistent quality and eliminating browser compatibility issues.

Whisper Implementation

Frontend Audio Capture

MediaRecorder API captures high-quality audio in multiple formats (WebM, MP4, OGG) for optimal compatibility.

Server-Side Processing

Audio files are sent to backend where Whisper processes them with language detection and confidence scoring.

Docker Integration

Whisper runs in containerized environment with FFmpeg support and optimized resource allocation.

Technical Benefits

QualityEnterprise-grade accuracy

Whisper provides state-of-the-art speech recognition quality across multiple languages.

PrivacyLocal processing

Audio data is processed locally, not sent to external speech recognition services.

ReliabilityBrowser-independent

Eliminates Web Speech API limitations and browser compatibility issues.

Supported Features

Audio Formats
  • • MP3, WAV, WebM
  • • M4A, OGG, FLAC
  • • Up to 25MB file size
Languages
  • • English, French, German
  • • Spanish, Italian, Portuguese
  • • Dutch, Polish, Romanian
Processing
  • • Real-time transcription
  • • Automatic language detection
  • • Confidence scoring

User Experience & Accessibility

Voice Input

  • • OpenAI Whisper integration
  • • Multi-language voice recognition
  • • Server-side speech-to-text processing
  • • High-quality AI transcription

Visual Feedback

  • • Recording state indicators
  • • Toast notifications
  • • Loading animations
  • • Error state handling

Session Management

  • • Multiple conversation sessions by topic
  • • Browser localStorage persistence (private)
  • • No external server storage for privacy
  • • Auto-session naming and organization
  • • Context continuity across sessions
  • • Data cleared only with browser cache

Intelligent Verification System

Every user query triggers an automatic verification process that cross-checks RAG results with current web sources, ensuring responses are both comprehensive and up-to-date.

How Verification Works

Step 1: RAG Retrieval

System searches knowledge base for relevant EU policy information using semantic similarity.

Step 2: Web Verification

Simultaneously searches EU official sources for latest updates, policy changes, and current information.

Step 3: Confidence Analysis

Compares RAG and web results, calculates confidence scores, and determines optimal response strategy.

Step 4: Enhanced Response

Generates final response combining verified knowledge base information with current web insights.

Verification Strategies

High Confidence (90%+)Use RAG

When web sources confirm RAG information is current and accurate.

Medium Confidence (60-89%)Combine Sources

When minor updates are detected, blend RAG knowledge with recent web findings.

Lower Confidence (<60%)Prioritize Web

When significant policy changes are detected, emphasize current web information.

System Performance & Metrics

24+
EU Policy Documents
24
EU Languages Supported
3072
Vector Dimensions
0.3
Similarity Threshold