Serve¶

Use IBM Confluent Real-Time Context Engine (RTCE) and Tableflow to serve continuously current business state to operational applications, AI agents, and open lakehouse query engines with sub-second latency.

Product mapping

IBM Confluent — RTCE & Tableflow — Real-Time Context Engine (low-latency state serving via Model Context Protocol / REST for AI agents) and Confluent Tableflow (zero-ETL materialization of Kafka topics as Apache Iceberg tables).

GitHub Repository

The complete source code and examples are available in the GitHub repository:

Building Blocks - Streamhouse Serve

Included Assets¶

Asset	Description
live-context-for-supply-chain-resilience	Full-stack operational AI application combining Confluent Real-Time Context Engine, watsonx Orchestrate AI agents, and an IBM Carbon React control tower
streamhouse-continous-rag	FactoryPulse Continuous RAG demo — Kafka/Flink RAG profiles, FastAPI backend, IBM Carbon UI, and IBM Code Engine deployment

Quick Start¶

Live Context for Supply Chain Resilience:

cd data/streamhouse/serve/assets/live-context-for-supply-chain-resilience/backend
cp .env.example .env
# Configure IBM Confluent, watsonx.ai, and watsonx Orchestrate credentials
pip install -r requirements.txt
python run.py

See assets/live-context-for-supply-chain-resilience/README.md for full setup including the React UI and agent configuration.

Streamhouse Continuous RAG:

cd data/streamhouse/serve/assets/streamhouse-continous-rag
cp .env.example .env
# Configure Confluent Cloud, watsonx.ai embedding model, and IBM COS
pip install -r requirements.txt
python -m app.main

See assets/streamhouse-continous-rag/README.md for Flink RAG profile setup and UI options.

Bob Skills¶

Skill	Description
data-streaming-confluent	Configures Confluent Tableflow, Apache Iceberg sink integrations, and RTCE endpoints for application consumption
streamhouse-continuous-rag	Combines Real-Time live operational state with RAG enterprise knowledge — designs and implements continuous RAG pipelines that ground AI agents in real-time context

Installing skills

Download the skill .zip files and copy the skill folders to ~/.bob/skills (global) or <project>/.bob/skills (project-level). See the Data Skills and Modes page for full installation instructions.

Why It Matters¶

Data is only as valuable as the decisions it empowers. In the Streamhouse architecture, Serve provides the dual consumption layer that bridges data in motion with both operational AI/applications and analytical lakehouses:

For AI Agents & Operational Apps: The Real-Time Context Engine provides sub-second query access to the latest state of any business entity (e.g., current location of a shipment, real-time risk score, latest fraud alert) via standard protocols like Model Context Protocol (MCP) and REST.
For Lakehouse Analytics: Confluent Tableflow automatically materializes Kafka topics as open Apache Iceberg tables without custom ETL pipelines, making streaming data instantly queryable in IBM watsonx.data, Presto, and Apache Spark.

Business Value¶

Key outcomes

Outcome	What It Means
Ground AI in real-time truth	AI agents receive live operational facts rather than stale batch data when formulating recommendations and taking actions
Zero-ETL lakehouse ingestion	Tableflow converts streaming topics into Apache Iceberg tables in object storage with zero code
Standardized agent integration (MCP)	Expose real-time state directly as Model Context Protocol (MCP) tools for LLMs and watsonx Orchestrate
Unified data architecture	A single streaming pipeline feeds both low-latency operational systems and analytical data engines simultaneously
Open formats & no vendor lock-in	Apache Iceberg tables generated by Tableflow are readable by any Iceberg-compatible query engine

When to Use¶

Use Serve when:

AI agents or applications need low-latency access to the current state of business entities.
You are building Continuous RAG architectures combining live streaming facts with document embeddings.
You want to eliminate custom batch ETL pipelines to get streaming data into Apache Iceberg tables in watsonx.data.
Real-time operational dashboards require sub-second state queries.

Core Capabilities¶

Capability	Component	What It Does
Live State Serving for AI	Real-Time Context Engine (RTCE)	Exposes entity state from stream tables to AI agents and apps via low-latency REST and Model Context Protocol (MCP)
Zero-ETL Table Materialization	Confluent Tableflow	Automatically writes Kafka event streams as Apache Iceberg tables in object storage with continuous catalog synchronization
Iceberg Sink Integration	Confluent Apache Iceberg Sink	Connects Kafka topics directly to IBM watsonx.data Apache Iceberg catalogs
Continuous RAG Grounding	RTCE + OpenRAG Integration	Merges real-time operational context from streams with semantic retrieval from enterprise documents

Reference Architecture¶

flowchart LR
    KAFKA["Governed Kafka Topics & Flink State<br/>(IBM Confluent)"]

    subgraph SERVE_LAYER["Streamhouse Serve Layer"]
        RTCE["Real-Time Context Engine (RTCE)<br/>• Fast State Queries<br/>• Model Context Protocol (MCP)<br/>• REST Endpoints"]
        TABLEFLOW["Confluent Tableflow<br/>• Zero-ETL Open Tables<br/>• Auto Iceberg Metadata"]
    end

    subgraph CONSUMERS["Downstream Consumers"]
        subgraph OPERATIONAL_AI["Operational AI & Applications"]
            AGENTS["watsonx Orchestrate / AI Agents"]
            APPS["Carbon Control Tower / Apps"]
        end

        subgraph LAKEHOUSE_ANALYTICS["Lakehouse Analytics"]
            WXD["IBM watsonx.data<br/>(Presto / Spark Engines)"]
            BI["BI Dashboards & Reports"]
        end
    end

    KAFKA --> RTCE
    KAFKA --> TABLEFLOW

    RTCE -->|MCP / REST| AGENTS
    RTCE -->|REST / SSE| APPS

    TABLEFLOW -->|Apache Iceberg Tables| WXD
    WXD --> BI

Continuous RAG Pattern: Real-Time + Lakehouse¶

Context Type	Provider	Characteristics	Example
Static / Unstructured Context	OpenRAG / OpenSearch	Long-term enterprise documents, policies, product guides	"What is our standard supplier SLA agreement?"
Live / Operational Context	Real-Time RTCE	Real-time events, current sensor metrics, live transit delays	"Where is supplier shipment #8942 right now and what is its risk status?"

Combining both layers gives AI agents full contextual awareness — what the enterprise knows plus what is happening right now.

flowchart TD
    RTCE["RTCE — live operational state\ncurrent orders · alerts · inventory · risk scores"]
    RAG["RAG Pipeline — enterprise knowledge\nhistorical documents · policies · contracts · specs"]

    RTCE --> AGENT["AI Agent\nwatsonx.ai"]
    RAG  --> AGENT

    AGENT --> OUT["Grounded response with\nlive + historical context"]

What to Demonstrate¶

Configure Tableflow on an active Kafka topic to automatically create and sync an Apache Iceberg table in object storage.
Query the generated Iceberg table in IBM watsonx.data using Presto SQL to show zero-ETL data availability.
Launch the Real-Time Context Engine (RTCE) exposing live stream state as an MCP server.
Connect an AI agent (e.g., watsonx Orchestrate or IBM Bob) to the MCP server.
Ask the AI agent for current operational status and show it retrieving real-time state directly from the stream.

Design Considerations¶

Serving design guidelines

Use RTCE for point lookups and low-latency state retrieval; use Tableflow/Iceberg for multi-record analytics and aggregation queries.
Tune Iceberg table compaction and snapshot retention settings to balance write performance and query speed.
Secure RTCE endpoints with mutual TLS (mTLS) or API keys when integrating with external AI agent runtimes.
Structure MCP tool parameters clearly so LLMs can easily formulate accurate state queries.

IBM Products Used¶

Product	Role
IBM Confluent — RTCE	Real-Time Context Engine delivering low-latency operational state for applications and AI agents
Confluent Tableflow	Converts Kafka topics into Apache Iceberg tables for lakehouse analytics
IBM watsonx.data	Open hybrid lakehouse querying Iceberg tables generated by Tableflow via Presto and Spark
IBM watsonx Orchestrate	AI agent platform consuming live state from RTCE via Model Context Protocol (MCP)