Skip to content

setup.md

Latest commit

 

History

History
88 lines (64 loc) · 3.28 KB

setup.md

File metadata and controls

88 lines (64 loc) · 3.28 KB
Raw

Running Agent mode against a local model

From nothing to a finished agent run, on your own machine, with no key and no traffic leaving it.

1. Install Ollama and pull a model

# macOS
brew install ollama
ollama serve            # leave running

ollama pull qwen3:4b    # 2.5 GB, passes all three workflows

qwen3:4b is the smallest model measured that does the whole job. For the alternatives and what each costs in disk, see the index.

2. Point LibreDB Studio at it

In .env.local:

LLM_PROVIDER=ollama
LLM_MODEL=qwen3:4b
LLM_API_URL=http://localhost:11434/v1

LLM_API_URL must end in /v1: that is Ollama's OpenAI-compatible endpoint, which is what the agent runtime posts to. The native /api/chat endpoint is not used.

No API key is needed. Nothing reaches a hosted provider on this path — what a run sends and where is documented in ../AGENT_DATA_FLOW.md.

Restart the server after editing the file; these are read server-side at startup.

3. Check the model can drive a run at all

Agent mode needs tool calling. A model that cannot call a tool will produce fluent prose and answer nothing, so it is worth ten seconds to find out first:

curl -s http://localhost:11434/api/chat -d '{
  "model": "qwen3:4b",
  "messages": [{"role": "user", "content": "Check this database health."}],
  "tools": [{"type": "function", "function": {
    "name": "inspect_operations",
    "description": "Read live database operational state.",
    "parameters": {"type": "object",
      "properties": {"kind": {"type": "string", "enum": ["health"]}},
      "required": ["kind"]}}}],
  "stream": false
}' | grep -o '"tool_calls".\{0,120\}'

A model that is usable answers with a tool_calls array naming inspect_operations. One that is not answers in prose — that is what every deepseek-r1 size does, and its page has the detail.

The application runs its own capability probe before a run and refuses a model whose tool calling is established as absent, so this check only saves you the wait.

4. Run one

Open Agent mode, pick a connection, and ask a question that needs the database. The embedded SQLite sample is there by default.

A finished run leaves a report with cited claims. If a workflow presents answers, it also leaves the result beside the report.

When it does not work

What you see Where to look
Report with an empty answer pane the model reported without presenting; see the verdict on the ledger
A run that ends having recorded nothing no-report — the model narrated its findings. Common below 4B
model-timeout one turn exceeded 90 seconds. Check the power mode first: pmset -g | grep powermode
The model refuses, saying it has no tools it may be unstable rather than incapable; see mistral-small3.2

Every run writes its ledger to .workflow-data, and that is the authority on what a run actually did — which tools it called, what came back, and why it was scored the way it was.

Using something other than Ollama

LLM_PROVIDER=custom with LLM_API_URL pointing at any OpenAI-compatible endpoint works the same way; openai and gemini take a key. All variables are documented in .env.example.