DEMO

FactoryPulse — Continuous RAG on Confluent Cloud Streamhouse

FactoryPulse shows Continuous RAG as a live Streamhouse workload on Confluent Cloud — factory exceptions trigger AI-grounded answers that refresh the moment a technician closes a work order, with no nightly index rebuild.

Building Blocks used:

data.pipelines

data.lakehouse

Skills used

Download skills for IBM Bob →

Products Used

1 Review use case

Manufacturing operators and maintenance engineers are increasingly reliant on AI assistants to diagnose machine faults, recommend corrective actions, and surface the right procedure at the right moment. The critical limitation of today's RAG deployments is staleness: AI retrieval is only as current as the last scheduled index rebuild, which means a technician's resolution note from this morning is invisible to the model until tonight's batch run. In a fast-moving factory environment — where CNC machines degrade in hours and work-order knowledge must flow immediately to the next shift — that lag creates operational risk.

FactoryPulse solves this by treating enterprise knowledge as a continuous event stream on Confluent Cloud. Machine telemetry, production exceptions, quality events, maintenance work orders, and SOPs all flow through the same Kafka event backbone. Confluent Cloud Flink continuously chunks new documents with ML_RECURSIVE_TEXT_SPLITTER and embeds them with AI_EMBEDDING, writing durable chunk-and-vector records back to a compacted Kafka topic. The FastAPI service rebuilds its retrieval index by replaying that topic, so the moment a technician closes work order WO-11023, the resolution is retrievable without a rebuild. watsonx.ai synthesizes a grounded natural-language answer and the UI returns every evidence chunk so the operator can inspect the grounding path.

The demo targets manufacturing operations managers, maintenance engineers, and IT architects evaluating real-time AI in industrial environments. Confluent Cloud Tableflow optionally materializes factory state and exception topics as Apache Iceberg or Delta Lake tables, connecting the live operational story to the analytical lakehouse for shift comparison, MTTR trending, and predictive model training — all governed by the same Schema Registry that enforces data contracts at every Kafka topic boundary.

Start Here

Watch this short overview

Preview the application

Explore key screens from the demo

2 Try the interactive application

Experience the interactive demo and explore its features. A demo script is provided to help when presenting the application to partners.

Try the demo

Have questions, want a walkthrough or found a bug?

Contact the demo owner for assistance.

Submit a request →

3 Explore how it was built

FactoryPulse is built on Confluent Cloud as the shared event backbone. Machine telemetry, production events, quality results, and maintenance work orders flow into Kafka topics. Two Confluent Cloud Flink jobs run continuously: the operational pipeline (factory_operations.sql) aggregates telemetry into 30-second windows, detects threshold breaches, emits factory.exceptions, and joins telemetry + quality + production into factory.state; the Continuous RAG pipeline (continuous_rag.sql) chunks new knowledge documents with ML_RECURSIVE_TEXT_SPLITTER and embeds them with AI_EMBEDDING, writing durable chunk-and-vector records to rag.knowledge.embeddings. A Python 3.12 / FastAPI service deployed on IBM Cloud Code Engine consumes the embedding topic and maintains a Kafka-replayed retrieval index; when a question arrives it performs cosine-similarity retrieval, optionally calls watsonx.ai for answer synthesis, and returns the answer with evidence chunks to the dependency-free web UI. Schema Registry enforces data contracts at every topic boundary. Tableflow materializes factory.state and factory.exceptions as Apache Iceberg or Delta Lake tables for historical analytics, shift comparison, and model development — without touching the low-latency RAG request path.

Architecture

4 Download and modify with IBM Bob

Download the source code and specifications for this application to set up and run it locally. Once downloaded, open the code in IBM Bob to modify and adapt it to your own use case.

Specifications

Source Code

FactoryPulse — Seller Demo Script (8–10 minutes)

1. Set the scene (45 s)

Key message: “We are not showing Kafka as a message pipe. We are showing a Streamhouse where operational state and enterprise knowledge are continuously updated on the same event backbone.”

Open the Operations Overview tab and click Baseline.
Point out: CNC-03 is healthy, production is above target, quality is normal.

2. Degradation begins (60 s)

Key message: Machine condition, production, and quality indicators are continuously combined in a single view — no single source system owns this answer.

Click Degrade.
Explain: machine vibration is rising; the current state now combines condition, production, and quality indicators in a single continuously derived view.

3. Business exception (60 s)

Key message: “No single source system owns this combined answer. The Streamhouse continuously represents what the factory knows right now.”

Click Critical.
Call out on screen: vibration 5.8 mm/s, cycle time increased, defect rate elevated, projected output drops to 870/1000.

4. Continuous RAG (2 min)

Key message: The answer is grounded in current maintenance SOPs and work-order history, not general model memory.

Click Seed knowledge if not already done.
Open the Continuous RAG tab and ask: “Why is CNC-03 showing high vibration and what should the technician inspect first?”
Show the evidence cards returned with the answer.

5. The continuous-learning moment (2 min)

Key message: “Traditional RAG answers from what the index knew when it was last refreshed. Continuous RAG answers from what the enterprise knows now.”

Return to Overview and click Resolve + Learn.
Explain: this publishes synthetic work-order WO-11023 resolution as a new Kafka event — tool-holder imbalance found and corrected.
Wait a few seconds, return to Continuous RAG and ask: “What changed after WO-11023 was resolved?”
Call out: the resolution is retrievable immediately — no rebuild required.

6. Explain the Streamhouse topology (90 s)

Key message: Every velocity of factory data — telemetry, documents, events — normalized into one ordered event backbone.

Open the Streamhouse tab and walk through:
- Kafka = durable event backbone
- Flink = two concurrent jobs (operational correlation + continuous chunking/embedding with ML_RECURSIVE_TEXT_SPLITTER and AI_EMBEDDING)
- Schema Registry = data contracts at every topic boundary
- Tableflow = analytical/historical Iceberg path

7. Close (30 s)

Key message: “The value is not just lower latency. It is closing the gap between a business change, the enterprise’s shared understanding of that change, and the action an operator or AI assistant can take.”