Ollama monitoring
This is the existing Agentless Ollama monitor. It reads service/model metadata and is separate from OTel application instrumentation or AI Gateway provider configuration.
Configure the monitor
Choose Ollama in the monitor creation catalog. Set a host reachable from the selected collector, the port (template default 11434) and the matching TLS setting. If your endpoint or access proxy requires a Bearer credential, the optional API Key field supplies it. This field does not create authentication in an otherwise unauthenticated service.
Run detection before saving, then inspect a fresh current sample after a configured collection interval. A collector container cannot reach another container through its own loopback address. Limit network access according to your deployment and keep keys out of public diagnostics.
Source-backed collection contract
The template hertzbeat-manager/src/main/resources/define/app-ollama.yml issues three GET requests:
| Metric group | Endpoint | Evidence |
|---|---|---|
version_info | /api/version | Service version; priority-zero availability check |
models | /api/tags | Installed model names, sizes and descriptive metadata |
running_models | /api/ps | Currently loaded model rows, size/VRAM fields and expiry |
The model-list and running-model contracts are described in the official tags and ps references. HertzBeat reads those rows; it does not load or download a model. Template size fields are converted from bytes to MB. Empty running-model rows do not imply a failed service, and missing fields are not measured zeroes.
Diagnose missing results
Version collection must succeed before later metric groups run. Distinguish connection/TLS failures, authentication failures, incompatible response shapes and a valid empty model list. Compare returned fields with the template for your server version. These groups do not measure inference error rate, request latency, model quality or complete memory utilization.
A real Ollama deployment was not part of this alpha's MySQL/Java intake acceptance. This guide records the implemented template and expected upstream shape, not a verified-version matrix. Record actual detection/current samples for your deployment. Use the separate AI Gateway guide if you want to configure a model provider.