Creating an AI Server for LM Studio models

This guide describes how to register a running LM Studio instance as an AI Server in AI Knowledge Fabric, so its models become selectable for Topics, GraphRAG Topics, and Agent Pipelines.

Prerequisite: LM Studio is installed, the models are downloaded, and the local server is running with Serve on Local Network enabled. If that is not yet the case, work through Getting started local model first.

1. Background: Managed vs. External

Every AI Server has a type that determines whether the platform controls its container:

Server Type Meaning

Managed (deployed by this app)

The platform generates a Compose file and starts/stops an Ollama container itself. Locked to the Ollama provider — this is the only provider it knows how to deploy.

External (existing instance)

The platform only talks to an already-running endpoint. Start/Stop do nothing; only reachability is checked.

LM Studio is always registered as External. The platform cannot start or stop LM Studio — you control that on your desktop.

2. Create the AI Server

Navigate to Admin → AI Servers and click Add AI Server.

Fill in the form

Field Value Notes

Server Type

External (existing instance)

Required — Managed would lock the provider to Ollama

Provider

OpenAI API (self-hosted)

LM Studio speaks the same language (protocol) as the OpenAI API

Name

e.g. LM_Studio

Must be unique. Letters and only, at most 50 characters — a space becomes

Host

see the table below

Must be reachable from inside the Topic containers

Port

1234

LM Studio’s default server port

Use TLS (HTTPS)

off

LM Studio serves plain HTTP. Turn it on only for a server reached over https://

Default Model ID

qwen2.5-coder-7b-instruct

Pre-fills the model picker in the Topic wizard; the user can still choose another model

Leave API Key and Management API Key Env Var empty. LM Studio needs no key by default. If you protected your server with a key, enter it as API Key.

Choosing the Host value

This is the field that most often goes wrong. The value must be reachable from the Topic container, not from your browser:

Your setup Host value

Management app in Docker (Docker Desktop or Linux), LM Studio on the same machine

host.docker.internal

Management app started directly from the IDE / JAR, on the same machine as LM Studio

localhost

LM Studio on a different machine in the network

That machine’s IP or hostname, e.g. 192.168.1.42

Management app in Kubernetes

An address of the LM Studio machine that the pods can reach — never localhost

A pasted address is fine. If you paste a full address such as 192.168.1.42:1234/v1, the form splits it when you save: the host stays in Host, 1234 goes into Port, http or https sets Use TLS, and /v1 is dropped, because the platform adds it by itself.

Click Save.

3. Verify the connection

The server detail page has three tabs:

Tab Contents

Configuration

The values you entered — editable

Models

The models LM Studio reports

Actions

Reachability check and deletion. Start/Stop do nothing for an external server

Open the Models tab. Both models should be listed:

  • qwen2.5-coder-7b-instruct

  • text-embedding-nomic-embed-text-v1.5

If the list is empty, the endpoint is not reachable — see Troubleshooting.

Pull New Model is only offered for Ollama servers. LM Studio models are downloaded in LM Studio itself, not from this UI.

4. Use the server

Once the server is saved and reachable, its models appear in every model picker:

Where What to select

Topic wizard, model step

Chat model: LM_Studio → qwen2.5-coder-7b-instruct · Embedding model: LM_Studio → text-embedding-nomic-embed-text-v1.5

Topic detail → Configuration

Per-topic override of the chat/embedding model

Pipeline Designer, AGENT node

Pick LM_Studio and a model in the model selection. This fills in provider, model and address of the module

Notes for local inference:

  • Set a timeout that suits local hardware. A 7B model on CPU is far slower than a cloud API — 5 minutes or more is realistic. Never leave the timeout empty.

  • Keep the embedding model stable. Changing it invalidates all existing vectors in a Topic and the documents have to be ingested again.

  • The same server can be used simultaneously by several Topics and Agents. LM Studio processes the requests sequentially, so parallel load increases response times noticeably.

5. Troubleshooting

Symptom Cause and fix

Models tab is empty, server shown as unreachable

LM Studio server not running, or Serve on Local Network disabled. LM Studio binds to localhost only by default, which containers cannot reach

"Cannot resolve hostname"

Wrong Host value — see Choosing the Host value

"Connection refused"

Host resolves but nothing is listening on port 1234. Check LM Studio’s port setting

Model list shows only one model

The other model is not loaded in LM Studio and Just-In-Time loading is disabled

Chat works, document upload fails

Wrong model selected as the embedding model, or the embedding model is not loaded

A name is rejected

An AI server with that name already exists, or the name contains unsupported characters

Related