Compendium.Adapters.Ollama 1.0.0-preview.0

compendium-adapter-ollama

Ollama AI provider adapter for the Compendium event-sourcing framework. Implements IAIProvider from Compendium.Abstractions.AI against the Ollama HTTP API for local LLM inference — designed for developer workflows and air-gapped production environments.

Extracted from sassy-solutions/compendium per ADR-0006 (multi-repo adapter split).

Install

dotnet add package Compendium.Adapters.Ollama

services.AddCompendiumOllama(builder.Configuration.GetSection("Ollama"));

Quick-start

1. Install Ollama (one of):

   # macOS
   brew install ollama && brew services start ollama
   # Linux
   curl -fsSL https://ollama.com/install.sh | sh
   # Docker
   docker run -d -p 11434:11434 --name ollama ollama/ollama
   

2. Pull a model:

   ollama pull llama3.2          # general chat + tool calling
   ollama pull nomic-embed-text  # embeddings
   

3. Wire the adapter into your DI container:

   services.AddCompendiumOllama(opt =>
   {
       opt.BaseUrl = "http://localhost:11434";
       opt.DefaultModel = "llama3.2";
       opt.DefaultEmbeddingModel = "nomic-embed-text";
   });
   

4. Use IAIProvider like any other Compendium AI provider:

   var result = await provider.CompleteAsync(new CompletionRequest
   {
       Model = "llama3.2",
       Messages = new[] { Message.User("Tell me a joke.") },
   });
   

A runnable example lives in samples/01-local-chat.

Options

Tool calling

Some Ollama models advertise OpenAI-compatible tool calling. Use the WithTools() extension:

using Compendium.Adapters.Ollama.Tools;

var tools = new List<AgentTool>
{
    new("get_weather", "Returns the current weather.",
        """{"type":"object","properties":{"city":{"type":"string"}},"required":["city"]}"""),
};
var request = new CompletionRequest
{
    Model = "llama3.2",
    Messages = new[] { Message.User("What's the weather in Paris?") },
}.WithTools(tools);

var result = await provider.CompleteAsync(request);
var calls = result.Value.GetToolCalls(); // IReadOnlyList<AgentToolInvocation>

Model support is uneven. Tool calling generally works on:

Older models silently ignore the tools field; the response will simply have no tool_calls. Verify against your specific model + Ollama version before relying on tool flow in production.

Streaming

Ollama streams newline-delimited JSON (not Server-Sent Events). The adapter parses chunks automatically — consume them as any other IAsyncEnumerable<Result<CompletionChunk>>:

await foreach (var chunk in provider.StreamCompleteAsync(request, ct))
{
    if (chunk.IsFailure) break;
    Console.Write(chunk.Value.ContentDelta);
}

Auto-pull

Set opt.AutoPullModel = true to automatically pull a missing model the first time it's requested. Useful for dev and one-shot ops; not recommended for production because pulls block the request for the duration of the download (often gigabytes).

Security

Ollama runs unauthenticated by default on localhost. For production:

• Never expose the raw API to the public internet — there is no auth, no rate-limiting, no per-tenant isolation.
• Put it behind a reverse proxy that handles authentication (e.g. nginx + OIDC, or a Kubernetes ingress with mTLS).
• Bind Ollama to 127.0.0.1 only and let the proxy do the public-facing TLS termination.
• Treat the model surface as untrusted code execution context — anyone with API access can pull and run arbitrary models.

Recommended models

Versioning

Versions are driven by git tags via MinVer. The first published version is 1.0.0-preview.0 (the orchestrator tags on merge — do not tag from a feature branch).

Build & test locally

dotnet restore
dotnet build -c Release
dotnet test  -c Release --filter "FullyQualifiedName!~IntegrationTests"

Integration tests boot a real Ollama container and pull tiny models — opt in via:

RUN_OLLAMA_INTEGRATION=1 dotnet test -c Release --filter "FullyQualifiedName~IntegrationTests"

Expect them to take several minutes the first time the image and models are pulled.

License

MIT — Copyright © 2026 Sassy Solutions.