Logo Event-Driven Architecture & Industrial Edge EN ES

AI Playground + Kafka MCP

OpenShift AI 3.3 includes the Gen AI Studio (Playground) for interacting with deployed LLMs. Combined with the Kafka MCP Server registered through Kuadrant’s MCP Gateway, the Playground becomes an AI agent capable of querying live Kafka data.

Architecture

Component Role

RHOAI Playground

Chat UI connected to the Granite 3.1 2B InferenceService (KServe)

Granite 3.1 2B Instruct

Small-footprint LLM running via llama.cpp on KServe — handles natural-language reasoning

MCP Gateway (Kuadrant)

Streamable-HTTP proxy that exposes registered MCP servers under a unified endpoint

Kafka MCP Server

Python FastMCP server with 4 tools: kafka_list_topics, kafka_describe_topic, kafka_read_messages, kafka_consumer_groups

Kafka Cluster

CDC cluster (cdc-cluster) in kafka-cdc namespace with 12+ topics

MCP Gateway Endpoint

URL:  https://mcp-gateway.apps.<CLUSTER_DOMAIN>/mcp

Available Kafka Tools

Tool Description Parameters

kafka_list_topics

Lists all non-internal Kafka topics with partition counts

none

kafka_describe_topic

Shows partition offsets, message counts, and watermarks for a topic

topic_name (string)

kafka_read_messages

Reads the latest N messages from a topic (max 20)

topic_name (string), count (int, default 5)

kafka_consumer_groups

Lists all consumer groups with their states

none

Available Kafka Topics

Topic Content

cv.face.detections

NeuroFace CV events: person_registered, person_detected, ovms_model_status

cdc.public.customers

CDC change events from the customers table

cdc.public.orders

CDC change events from the orders table

bpm.approval.requests

BPM approval workflow events from Kogito/SonataFlow

bpm.onboarding.metrics

BPM onboarding metrics and process telemetry

dlq.cdc-errors

Dead-letter queue for failed CDC events

dlq.cv-errors

Dead-letter queue for failed CV pipeline events

dlq.cdc-camel-errors

Dead-letter queue for failed Camel K integrations

Playground Prompts

Prompt 1 — Inventory Overview

List all Kafka topics available in the cluster.
For each topic, show the name and number of partitions.
Highlight any topics related to face detection or computer vision.

Prompt 2 — CV Pipeline Analysis

Read the last 10 messages from the topic cv.face.detections.
Summarize:
- How many unique people were detected?
- What detection methods are being used?
- What is the average confidence score?
- Are there any OVMS model status events? If so, is the model healthy?

Prompt 3 — CDC Event Inspection

Describe the topic cdc.public.orders and tell me how many total messages it has.
Then read the last 5 messages and summarize the order data:
- What fields are in each record?
- Are these INSERT, UPDATE, or DELETE operations?
- Show a brief table with customer_id, amount, and operation type.

Prompt 4 — Consumer Group Health

List all consumer groups in the Kafka cluster.
For each group, show the state (Stable, Empty, Dead, etc.).
Identify any consumer groups that might indicate a problem
(e.g., Dead or with rebalancing state).

Prompt 5 — Dead Letter Queue Analysis

Check the following dead-letter queue topics:
- dlq.cdc-errors
- dlq.cv-errors
- dlq.cdc-camel-errors

For each one, tell me how many messages are in it.
If there are messages, read the latest 3 and summarize
what errors occurred and which components generated them.

Prompt 6 — Security Face Detection Report

Read the last 20 messages from cv.face.detections.
Generate a security report with:
1. Timeline of detections (person name + timestamp)
2. Total unique identities currently registered
3. Detection frequency (events per minute estimate)
4. Model server health status
Format the report as a professional security briefing.

Prompt 7 — BPM Workflow Events

Describe the topics bpm.approval.requests and bpm.onboarding.metrics.
Read the last 5 messages from each and explain:
- What business processes are generating events?
- What is the approval workflow doing?
- Are there any onboarding metrics available?

Prompt 8 — Cluster-wide Topic Health

List all topics and describe each one that has more than 0 messages.
Create a dashboard-style summary:
- Topic name
- Total messages
- Number of partitions
- Category (CDC / CV / BPM / DLQ / Internal)
Flag any topic that looks unhealthy (e.g., too many DLQ messages
relative to the source topic).

Prompt 9 — NeuroFace End-to-End Pipeline

I want to understand the NeuroFace CV pipeline end-to-end.
1. Read the latest messages from cv.face.detections
2. Check if there are any errors in dlq.cv-errors
3. Describe the cv.face.detections topic to see throughput

Based on this data, explain:
- Is the pipeline healthy?
- How many events are being generated per hour (estimate)?
- Are any detections failing?

Prompt 10 — Cross-system Correlation

Read the last 5 messages from these topics:
- cv.face.detections
- cdc.public.orders
- bpm.approval.requests

Find any correlations:
- Are detections happening around the same time as orders?
- Is there any pattern between face detection events and BPM workflows?
Provide a timeline visualization in text format.

References