⚡

Enterprise Multi-Agent RAG Platform

Multi-Agent Backend & Cost Optimization Platform

🏠 Home

Hybrid FAISS + BM25 Search

15,000+ Chunks

FlatIP + BM25 Keyword Index

Semantic Cache Hits

0.0%

<10ms Instant Hits ($0)

Total API Cost

$0.0000

Dynamic Model Routing

Avg Pipeline Latency

0 ms

Parallel Execution

RAGAS Eval Score

--

Faithfulness & Relevance

💬 Enterprise Query Workbench

Employee Q&A View

Sample Scenarios:

🤖

Hello! I am your Enterprise AI Assistant backed by 15,000+ indexed document chunks, SQL analytics, and live API actions. How can I assist you today?

🎯

RAGAS Benchmark Performance Gain

Live Multi-Agent Optimization

Metric Name	Baseline	Improved	Gain (%)
Overall RAGAS	0.5648	0.7200	+27.5%
Faithfulness	1.0000	0.9800	-2.0%
Answer Relevance	0.5104	0.7500	+46.9%
Context Precision	0.4800	0.7800	+62.5%
Context Recall	0.2688	0.6500	+141.8%

📊 Multi-Agent Execution Trace

Ready

Routed Model: --

Model Tier: --

Query Cost: $0.0000

Total Latency: 0 ms

🤖 Dynamic Model Router rationale will appear here.

Execution Waterfall Spans

Submit a query to visualize real-time agent execution spans.

📌 What is this Project?

An enterprise-grade, production-deployed Autonomous Multi-Agent RAG & Automation Platform. Built on a decoupled 12-layer pipeline, it seamlessly orchestrates specialized AI agents (RAG Vector Search across 15,000+ chunks, AST-guarded SQL relational analytics across 6 schemas, and structured REST API tool actions). Features sub-millisecond semantic caching, cost-aware dynamic LLM routing, strict hallucination validation, ReportLab executive PDF generation, real SMTP email dispatching, and automated n8n workflow integration.

🏗️ Platform Architecture & Execution Pipeline

👤 USER

│
▼

🌐 ENTERPRISE AI WORKBENCH UI

│
▼

⚡ FASTAPI ENTERPRISE SERVICE

│
▼

🛡️ INPUT VALIDATION & SECURITY

│
▼

🚦 INTENT CLASSIFICATION AGENT

│

│
▼

│
▼

│
▼

🔍 RAG AGENT

│
▼

📦 FAISS + BM25 HYBRID SEARCH
Multi-Query + RRF Reranker

🗄️ SQL ANALYTICS

│
▼

📊 DATABASE ENGINE
6 Schemas + AST SQL Guard

⚡ API ACTION AGENT

│
▼

🔌 REST API REGISTRY
Structured Function Calling

│

│

│

│
▼

📦 RETRIEVED / PREPARED DATA

│
▼

💰 DYNAMIC COST & MODEL ROUTER
Payload Size + Complexity (Tier 1 vs Tier 2)

│
▼

🤖 REASONING & SYNTHESIS AGENT

│
▼

🛡️ VALIDATION AGENT
Groundedness & Hallucination Check

│
▼

📝 RESPONSE AGENT

│

│
▼

│
▼

│
▼

📚 Grounded Citations

📈 Chart Analytics

⚡ Actions

│
▼

🔄 n8n Workflows

📊 Observability & Continuous Evaluation (Cross-Cutting Platform Layer)

Operates across all agent execution layers to monitor telemetry, audit LLM performance, enforce safety, and track RAGAS benchmarks in real time.

⚡ Real-Time Telemetry & Cost Tracking

Execution Traces: OpenTelemetry waterfall spans per request
Latency: Per-agent breakdown & total pipeline latency
Token Usage: Prompt & completion token budgets
API Cost: Live USD cost monitoring ($0 for cache hits)
Model Selection: Dynamic Tier 1 / Tier 2 routing audit
Cache Performance: Hit ratio & sub-ms similarity tracking

🎯 RAGAS Benchmarking Framework

RAGAS Benchmark: 200+ query evaluation dataset
Faithfulness (Groundedness): Boosted via Reciprocal Rank Fusion (RRF) Reranking
Answer Relevance: Boosted via Multi-Query Expansion (4 Semantic Variations)
Context Precision: Boosted via RRF Reranking & Dynamic Score Cutoffs
Context Recall: Boosted via Hybrid Search (Dense FAISS + Sparse BM25 Keywords)
Phase 2 Benchmark: SQL & Tool execution success metrics

📂 Codebase Verification & Architecture Mapping

Direct 1-to-1 mapping connecting every architectural component in the diagram to its corresponding backend Python module:

Architecture Component	Backend Python Module	Responsibilities & Functionality
🛡️ Input Security & AST Guard	security/sanitizer.py	PII redaction, prompt injection guardrails, rate limiting
⚡ Semantic Vector Cache	cost_optimization/semantic_cache.py	<1ms sub-millisecond cache hits for recurring queries (>=0.92 cosine sim)
🚦 Intent Classification Agent	agents/classification_agent.py	Traffic controller routing query to RAG, SQL DB, or REST API Action
🔍 RAG Retrieval Agent	agents/rag_agent.py & rag_pipeline/	Multi-Query Expansion (4 variations) + Hybrid Search (FAISS + BM25) + RRF Reranking
🗄️ SQL Analytics Agent	agents/sql_agent.py & tools/sql_engine.py	AST parse tree validation (100% read-only) + SQLite query execution
⚡ API Action Agent	agents/api_agent.py & tools/function_registry.py	Structured JSON schema tool calling for PTO, Invoices, and HR actions
💰 Dynamic Cost & Model Router	llm_router/model_router.py	Payload & complexity analysis routing between Tier 1 Flash vs Tier 2 Pro
🤖 Reasoning & Synthesis Agent	agents/reasoning_agent.py	LLM generation & multi-source context synthesis
🛡️ Validation & Groundedness Agent	agents/validation_agent.py	Hallucination prevention & claim verification with automatic retries
📝 Response Agent & Visualizer	agents/response_agent.py	Persona view formatting (Employee Q&A vs Workbench) + SVG bar charts
🔄 n8n Enterprise Workflows	workflows/n8n_integrator.py & pdf_generator.py	ReportLab executive PDF generation, SMTP email dispatching, webhook triggers
🎯 RAGAS & Phase 2 Benchmarks	evaluation/ragas_evaluator.py & phase2_evaluator.py	Faithfulness, Answer Relevance, Context Precision/Recall, SQL accuracy

⚙️ How It Works (Chronological 6-Stage Pipeline)

1. Security Guard & Vector Cache

Queries undergo PII redaction and prompt injection checks, followed by a FAISS similarity cache scan returning sub-millisecond (<1ms) instant hits for recurring queries.

2. Intent Classification Agent

Traffic Controller Agent evaluates query intent and routes to specialized domain agents (RAG Vector Search, Relational SQL Analytics, or REST API Actions).

3. Specialized Data Retrieval

Executes FAISS vector retrieval across 15,000+ chunks with BM25 reranking, AST-guarded read-only SQL queries across 6 business schemas, or structured REST API tool calls.

4. Dynamic Cost & Model Router

Evaluates retrieved payload size and query complexity to route dynamically between Tier 1 Flash ($0.00015) vs Tier 2 Pro ($0.002) LLM models to optimize budget spend.

5. Reasoning & Groundedness Validation

Reasoning Agent synthesizes answer, while Validation Agent audits claim overlap against context to eliminate hallucinations with automated retry loops.

6. Response Persona & n8n Automation

Formats Employee Q&A vs Agent-Assist Workbench views, renders SVG data charts, attaches citations, and triggers n8n PDF compilation & SMTP email dispatching.