AI architecture

roles · customization

Where the LLM can come from

Training and model files

Templates in s3://kmai1km-templates

InferenceServices on instdb2 (11)

Storage and models

Tenant (aiservice-inst-user)

AI Service (aiservice-inst)

MAS Core + Manage

Who asks

tools/call

MCP

REST as key user

AI broker

training export

templates

reads template

deploys

-gpt configs

holds

deploys

deploys

deploys

deploys

deploys

pipeline.yaml

writes model

loads

pipeline.yaml

writes model

loads

pipeline.yaml

writes model

loads

loads

loads

loads

pipeline.yaml

writes model

loads

pipeline.yaml

writes model

loads

loads

loads

5 need it for every answer

3 for training only

today

model swap

same API, on-cluster

not supported

lab patch

APP Maximo Assistant
chat in Manage

APP Role-based apps
similar WOs, classification

CLIENT MCP clients
Control Room, IDE agents

CR AiCfg
suite AI config

POD ALM agent
plans multi-step asks

POD MCP server
inst-mcp, McpiCfg

APP Manage
AI configurations,
mxe.int.aibroker props

SETTING Training data
queries + filters

POD AI broker API
tenantid + apikey

POD km-controller
builds models

POD km-store
template catalog

POD km-watcher

SECRET wx-secret
watsonx key, URL, project

CR KServe predictors
one per AI configuration

CR almwatsonx predictor
LLM proxy

STORAGE Object storage
km-s3-secret: templates,
training data, logs

EXTERNAL watsonx.ai
foundation models

EXTERNAL IBM entitled registry
predictor images

CR TICKETSIMILARITY
[email protected]
● runs locally

CR WOSIMILARITY
[email protected]
● runs locally

CR WOPROBLEMCODE
[email protected]
◐ watsonx for training

CR RSSTRATEGYASSISTANT
[email protected]
○ needs watsonx

CR MREF_LEASE_ABSTRACT
[email protected]
○ needs watsonx

CR IBMDOCS
[email protected]
○ needs watsonx

CR PLUSGINCTYPE
[email protected]
◐ watsonx for training

CR PLUSGINCIDENTGROUP
[email protected]
◐ watsonx for training

CR INSIGHT
[email protected]
○ needs watsonx

CR ASSISTANT
[email protected]
○ needs watsonx

CONFIG docsearch 1.0.5
data/data.parquet.zip + manifest.json · 20 args

CONFIG fieldextractor 1.0.2
manifest.json · 8 args

CONFIG fmea 1.3.2-gpt
manifest.json · 30 args

CONFIG insightsgenerator 1.3.2-gpt
manifest.json · 15 args

CONFIG mcc 1.7.1-gpt
manifest.json + pipeline.yaml · 22 args

CONFIG nl2oslc 1.4.2-gpt
manifest.json · 36 args

CONFIG pcc 1.8.2-gpt
manifest.json + pipeline.yaml · 17 args

CONFIG similarity 1.1.1
manifest.json + pipeline.yaml · 11 args

POD TICKETSIMILARITY training
pipeline pod

STORAGE s3://kmai1170a6fe0ba8611f
model files

POD WOSIMILARITY training
pipeline pod

STORAGE s3://kmai11e421fb0ba8611f
model files

POD WOPROBLEMCODE training
pipeline pod

STORAGE s3://kmai1899753a0ba9711f
model files

STORAGE s3://kmai193828120ba8b11f
model files

STORAGE s3://kmai1c7d9db80ba8b11f
model files

STORAGE s3://kmai1c8eb99a0ba8b11f
model files

POD PLUSGINCTYPE training
pipeline pod

STORAGE s3://kmai1d8d5d870ba9611f
model files

POD PLUSGINCIDENTGROUP training
pipeline pod

STORAGE s3://kmai1dd3f2600ba9611f
model files

STORAGE s3://kmai1f958f320ba8c11f
model files

STORAGE s3://kmai1fe613fd0ba8c11f
model files

EXTERNAL watsonx.ai SaaS
IBM Cloud · today
supported

EXTERNAL Other watsonx models
Llama · Mistral · Granite
model swap

EXTERNAL watsonx.ai software
on your cluster · Hugging Face models
supported · data stays

EXTERNAL Claude / ChatGPT
Anthropic · OpenAI APIs
not supported

EXTERNAL Local open model
Ollama / vLLM
lab only

customize here
customize only via CR or builder tools
managed by IBM / the operator
dashed arrow = optional · future · indirect