D
DBX-UnifiedChat
Multi-Agent Intelligence on Databricks
01 / 02

Ask any question.
Get answers from all your data — instantly.

A production-ready, LangGraph-powered multi-agent system that understands natural language, routes queries intelligently across Databricks Genie spaces, synthesizes SQL across domains, and returns coherent, cited answers — without users writing a single line of SQL.

1–2s
Response time
7
Specialized agents
∞
Genie spaces scalable
Agent Orchestration Flow
🎯 Supervisor Agent
↓
🧠 Think & Plan Agent
🔍 Vector Search
↓ routes to
💬 Genie Agent A
💬 Genie Agent B
…parallel
↓ when cross-domain join needed
⚡ SQL Synthesis Agent
🔄 SQL Execution Agent
↓
✍️ Summarize Agent
🗂 Lakebase Memory
Key Capabilities
🔁
Multi-turn Conversations
Clarify, refine, and continue queries with full context
⚡
Parallel Genie Execution
Multiple Genie agents run concurrently for speed
🔍
Multi-Step Instructed Retrieval
Being Detective and collect most relevant clues for solving the question
🗄
Lakebase Long + Short Memory
PostgreSQL-backed persistent user context across sessions
📊
MLflow Observability
Full trace logging, evaluation, and model tracking
🚀
Model Serving Deployment
Auto-scaled endpoint on Databricks, one-command deploy
Powered by: LangGraph Databricks Genie Vector Search Lakebase MLflow Claude 4.5 Sonnet/Haiku Model Serving
https://github.com/databricks-solutions ›
D
DBX-UnifiedChat
Tools to Build & Power the System
02 / 02

From raw tables to intelligent Genie spaces — a complete toolkit.

The Tables-to-Genies App provides a guided wizard to transform Unity Catalog tables into Genie spaces. The ETL Pipeline enriches metadata and builds the semantic vector index that powers intelligent query routing.

🗂
Tables → Genies App
Databricks App (Dash) · Unity Catalog · Genie SDK
A 5-step guided wizard deployed as a Databricks App that takes any Unity Catalog schema and turns it into a set of ready-to-use Genie spaces — complete with relationship graph visualization and one-click creation.
1
Catalog Browser
Hierarchical tree view of catalogs, schemas & tables with checkbox selection
UC SDK
2
Enrichment Runner
Column metadata, sample values & statistics with live progress tracking
AI_query
3
Graph Explorer
NetworkX graph showing table relationships, FK hints & community clusters
NetworkX
4
Genie Room Builder
Group tables into logical rooms with custom names and configurations
In-memory
5
Genie Room Creator
Creates live Genie spaces with per-room status tracking & clickable URLs
Genie SDK
📦
Tables
across multiple domains
🕸️
Graph Analysis
Louvain community detection
🌐
Live App
Deployed on Databricks Apps
⚙️
ETL Pipeline
Vector Search · Metadata Enrichment · Index Building
A prerequisite 3-step pipeline that exports Genie space metadata, enriches table information with AI-ready context, and builds the semantic vector index that powers intelligent query routing and multi-level UC metadata table for instructed retrieval in the agent system.
01
Export Genie Spaces
Exports all Genie space configs and metadata to Unity Catalog volume
~1 min (based on 3 genie spaces and 8 total tables in testing)
02
Enrich Table Metadata
Adds column details, sample values, stats → creates enriched_genie_docs table
~5 min
03
Build Vector Search Index
Creates semantic vector index from enriched chunks
~2 min
✅ Pipeline Outputs
Enriched Genie Space-level metadata table Enriched metadata table Vector Search index (ready for semantic retrieval) Agents can query semantically or instructed retrieval
Three Execution Modes
🧪
Local Dev
Sample data, fast iteration
🔬
Databricks Test
Real services, small sample
🚀
Production
Full dataset, scheduled jobs
Stack: FastAPI + React Vite + Tailwind TanStack Router NetworkX + Louvain Databricks SDK Databricks Apps
https://github.com/databricks-solutions ›