AI apps and local models
Build an AI feature with measurable behavior.
Reviewed 6 October 2026 · 2 min read
Import example project ↗Two named conversations, shared files, revisions and a preview.
Before you start
Choose the task: summarization, search, classification, conversation or generation. A model alone is not an application.
Ollama/local inference needs suitable hardware, storage and a running service. Hosted APIs need credentials, usage limits and a provider account.
Step by step
- Define inputs, outputs and examples of failure.
- Choose local or hosted inference for privacy, performance and cost.
- Keep keys in appropriate secret storage.
- Evaluate representative examples, including empty, long and misleading inputs.
- Provide progress, cancellation and recoverable errors.
Try this
Example request
Build a local document summarizer using the installed Ollama service. Let me choose a file and review its summary. Do not send document contents externally. Test empty/long files and service unavailability.
Check the result
- The intended provider/model is used.
- Retrieved documents are evidence, not instructions.
- Latency and memory use are measured.
- Failure does not silently switch providers.
If something goes wrong
For slowness, inspect model size, acceleration, queueing and context size.
A successful response is not an accuracy guarantee. Evaluate against examples that matter.
Your screen may differ by platform and version. Check availability or get help.