Talk2Data – Conversational BI Platform on AWS Bedrock

Talk2Data is an agentic Business Intelligence platform that lets your team query databases using plain English — no SQL knowledge required. Powered by AWS Bedrock (Claude) and a locked 3-agent Agentcore runtime, it acts as an always-available data analyst that thinks, questions assumptions, and delivers interpreted insights rather than raw query results.

How it works ? Users submit natural language questions through a React web interface. The Parent Agent classifies intent, enriches the request with business context (glossary terms, SOPs, schema metadata), generates and validates SQL, executes it against your databases, and returns a formatted, business-friendly response — all in a single conversational turn.

Key Capabilities

• Natural Language to SQL: Automatically generates, validates, and executes SQL against PostgreSQL (OLTP) and Amazon Redshift (OLAP) databases.
• Business Context Enrichment: Maps business terminology to database schema using your own glossary and standard operating procedures stored in S3.
• Human-in-the-Loop Checkpoints: Pauses before high-impact queries to confirm intent with the user, backed by Redis-persisted state so sessions survive restarts.
• 5 Output Formats: Results delivered as Markdown summaries, data tables, Vega-Lite charts, KPI cards, or insight narratives — ready for React rendering.
• Multi-Step Chained Analysis: Chains multiple query workflows together, passing results between steps for complex analytical sequences.
• Scheduled Analysis and Anomaly Alerting: Run recurring analyses on a schedule and receive alerts when metrics deviate from expected ranges.
• Natural Language to Dashboard: Describe a dashboard in plain English and get a fully composed, panel-based dashboard spec back.
• Collaborative Sessions: Multiple users can share a live analysis session with real-time presence and event streaming.
• Data Lineage Tracking: Understand where your data came from and how it was transformed across query steps.
• Multi-Tenant Data Source Routing: Add new database instances with config and environment variables only — no new containers or code changes required.

Architecture Talk2Data runs on a locked, extensible 3-agent topology in AWS Agentcore Runtime:

• Parent Agent: ReAct orchestrator handling routing, guardrails, memory, and HITL decisions.
• Business Context Agent: A2A collaborator retrieving schema and business metadata to ground SQL generation.
• Result Formatter Agent: A2A collaborator transforming execution results into structured output formats.

The SQL pipeline (Generator, Validator, Executor) runs as a LangGraph workflow inside a Specialists MCP server on ECS Fargate, keeping orchestration logic cleanly separated from data access. Database MCP servers (PostgreSQL, Redshift, S3, Orchestration) are type-based and multi-tenant, routing to instance-based data sources via configuration.

Security and Governance

• AWS Bedrock Guardrails (Layer 1 + Layer 2): Content filtering, PII detection, prompt injection prevention, and jailbreak detection on all input and output paths.
• Role-Based Access Control: SQL execution is gated by user roles defined in configuration.
• JWT Authentication: Secure session management with refresh token support.
• Secrets Manager and SSM Parameter Store integration for production credential management.
• All endpoints, model IDs, and agent IDs are config-driven — nothing hardcoded.

Extensibility Talk2Data is built to grow without touching core orchestration:

• Add a new analysis capability by creating a graph package and registering one MCP tool — no changes to the parent agent.
• Add a new database instance with a config entry and environment variables — no new containers.
• Add a new A2A collaborator by publishing an agent card — the parent routing table picks it up automatically.

Technology Stack AWS Bedrock (Claude) | AWS Agentcore | LangGraph | FastMCP | FastAPI | React + Vite | PostgreSQL | Amazon Redshift | Amazon ElastiCache (Redis) | Amazon S3 | Python 3.12