Recipe: Integrate Ollama
March 25, 2026 ยท View on GitHub
Goal
Set up regression tracking for local Ollama models with minimal cloud dependency.
Decision rule:
- If your task involves changing how EvalView talks to Ollama, use Path A.
- If your task is setting up an existing project to use Ollama, use Path B.
Read These Files First
evalview/adapters/ollama_adapter.pyevalview/core/adapter_factory.pyevalview/commands/shared.pyREADME.md
Existing Support
EvalView already has an Ollama adapter at evalview/adapters/ollama_adapter.py.
It sends requests to:
http://localhost:11434/v1/chat/completions
and converts the OpenAI-compatible response into ExecutionTrace.
Typical Setup Paths
Path A: direct adapter work
Use or extend OllamaAdapter if you need:
- new request parameters
- better token or trace capture
- health-check behavior
Path B: project setup
Use EvalView against a project configured for Ollama by:
- setting adapter/config values
- creating or updating tests
- running
snapshot - running
check
Useful Commands
evalview check tests --dry-run
evalview snapshot
evalview check
If you need local model selection in eval flows, inspect:
evalview/commands/shared.pyevalview/core/config.pyevalview/core/llm_provider.py
Done Criteria
- EvalView can execute tests against an Ollama-backed agent
- traces preserve output and token/latency metadata where available
snapshotandcheckrun without adapter-specific hacks elsewhere
Common Pitfalls
- confusing an Ollama-backed HTTP agent with the built-in Ollama adapter
- hardcoding model-specific logic into diffing or reports
- forgetting to test local health/failure behavior