Build a local AI ticket classifier.

Install the local model.

Install or update Ollama to 0.35 or newer and keep its local server running. Open PowerShell and check the version:

ollama --version

Review Nimble’s current model listing against your hardware before downloading. It currently lists roughly 9.3–9.5 GB of model files. Download size does not establish the memory required or how fast it will run.

ollama pull nimble

Send one sample ticket.

Copy this complete PowerShell example. It sends a synthetic message to your local decision endpoint and prints the selected category and its probability distribution.

$ticketClassifierRequest = @{
    model = 'nimble'
    state = @{ ticket = 'I was charged twice. Please refund the extra payment.' }
    questions = @{
        team = @{
            type = 'choice'
            instructions = 'Which team should handle this ticket?'
            criteria = @{
                billing = 'Payments, charges, and refunds'
                technical = 'Software bugs and integrations'
                other = 'No clear match, insufficient information, or another topic'
            }
        }
    }
}

$ticketClassifierJson = $ticketClassifierRequest | ConvertTo-Json -Depth 8
$ticketClassifierResponse = Invoke-RestMethod `
    -Uri 'http://localhost:11434/v1/systemone' `
    -Method Post `
    -ContentType 'application/json' `
    -Body $ticketClassifierJson `
    -TimeoutSec 180

$ticketClassifierResponse.answers.team.choice
$ticketClassifierResponse.answers.team.probabilities | ConvertTo-Json

The official launch example selects billing for a duplicate-charge request. Your output may differ. The first request can take longer while the model loads.

Bring ten examples you already labeled.

Write ten synthetic messages and assign a category yourself before running them through the model. Change only the ticket text for each run.

• Include straightforward billing and technical requests.
• Add an unrelated question that should land in Other.
• Try ambiguous wording and requests that overlap categories.
• Compare the predicted category with your label. Investigate mismatches.

Ten examples are a quick exploration, not a production validation set. Improve overlapping descriptions, retain the Other option, and test a larger representative set before connecting the model to real work.

Keep the person in the workflow.

Display the message, suggested category, and probabilities together for review. Separate application code can use the chosen label to propose a queue; the API itself does not move tickets or issue refunds.

Nimble’s confidence measures the concentration of its probabilities. It is not the chance an answer is correct, and a probability of 0.9 does not establish 90% accuracy on your examples. Validate any review threshold on your own data.

What was checked: this example’s PowerShell syntax and request JSON were checked. Local model execution and performance have not been tested by AI Anonymous.

If the example does not run.

• Command missing: finish the Ollama installation and open a new terminal.
• Connection refused: check that Ollama is running locally.
• Endpoint missing: verify the installed version supports /v1/systemone.
• Model missing: finish the ollama pull nimble download.
• Loading fails or runs slowly: check available memory and model requirements. Increase the request timeout if model loading exceeds the example’s limit.

This guide uses HTTP because the current model page contains inconsistent statements about native Ollama Python and JavaScript support. The launch documentation explicitly supports HTTP and the TypeSafe SDK.

Follow the evidence.

• Ollama’s decision-model announcement — release, setup, endpoint, and sample request.
• Nimble model documentation — version requirement, model size, API types, and limitations.
• TypeSafe confidence documentation — interpreting probability concentration and validating thresholds.

Continue with Prompt Patterns That Hold Up →