feat(llm): add Ollama provider support

[Kureii]

Summary

Add Ollama as a new LLM provider, enabling local/self-hosted model support via Ollama's OpenAI-compatible API.

Changes

• New OllamaProvider class using @ai-sdk/openai with custom baseURL
• Dynamic model discovery via GET /api/tags endpoint
• Provider registration in AddProviderModal with base URL configuration
• getModels route changed to asyncApiRoute to support async model fetching
• "Free" label for zero-cost models in the model selector
• Title generation support with automatic model selection

Files (9 files, +253/-26)

• apps/server/src/services/llm/providers/ollama.ts - New: Ollama provider implementation
• apps/server/src/services/llm/index.ts - Provider registration
• apps/server/src/services/llm/chat_title.ts - Ollama model handling
• apps/server/src/routes/api/llm_chat.ts - asyncApiRoute for getModels
• apps/server/src/routes/routes.ts - Route type update
• apps/client/src/widgets/type_widgets/options/llm/AddProviderModal.tsx - Ollama option in UI
• apps/client/src/widgets/type_widgets/llm_chat/ChatInputBar.tsx - Conditional cost display
• apps/client/src/widgets/type_widgets/llm_chat/useLlmChat.ts - "Free" label logic
• apps/client/src/translations/en/translation.json - Translation keys

[gemini-code-assist]

Code Review

This pull request introduces support for Ollama as a self-hosted AI provider. Key changes include updating the provider configuration UI to handle optional API keys and custom base URLs, adding translations for free models, and implementing the OllamaProvider on the server to dynamically fetch models from local instances. Feedback focuses on simplifying redundant logic in model retrieval and optimizing performance by caching loaded models to avoid unnecessary network requests during chat and title generation.

[eliandoran]

Overall seems pretty fine.

However, I've tested it on NixOS which comes with Ollama 0.12.11, before the responses API was introduced (0.13.3), which results in the chat failing:

APICallError [AI_APICallError]: Not Found
    at <anonymous> (/home/elian/Projects/TriliumNext/Notes/node_modules/@ai-sdk/provider-utils/src/response-handler.ts:70:16)
    at process.processTicksAndRejections (node:internal/process/task_queues:104:5)
    at async postToApi (/home/elian/Projects/TriliumNext/Notes/node_modules/@ai-sdk/provider-utils/src/post-to-api.ts:118:28)
    at async OpenAIResponsesLanguageModel.doStream (/home/elian/Projects/TriliumNext/Notes/node_modules/@ai-sdk/openai/src/responses/openai-responses-language-model.ts:1035:50)
    at async fn (/home/elian/Projects/TriliumNext/Notes/node_modules/ai/src/generate-text/stream-text.ts:1686:27)
    at async <anonymous> (/home/elian/Projects/TriliumNext/Notes/node_modules/ai/src/telemetry/record-span.ts:32:24)
    at async _retryWithExponentialBackoff (/home/elian/Projects/TriliumNext/Notes/node_modules/ai/src/util/retry-with-exponential-backoff.ts:96:12)
    at async streamStep (/home/elian/Projects/TriliumNext/Notes/node_modules/ai/src/generate-text/stream-text.ts:1638:17)
    at async fn (/home/elian/Projects/TriliumNext/Notes/node_modules/ai/src/generate-text/stream-text.ts:2166:9)
    at async <anonymous> (/home/elian/Projects/TriliumNext/Notes/node_modules/ai/src/telemetry/record-span.ts:32:24) {
  cause: undefined,
  url: 'http://localhost:11434/v1/responses',

This is not a good UX, we must guard against it.

Seems that we can fall back to chat mode via:

    protected createModel(modelId: string) {
        // Use .chat() to target /v1/chat/completions (Ollama doesn't support /v1/responses)
-        return this.openai(modelId);
+        return this.openai.chat(modelId);
    }

Maybe we could do some version check of Ollama.

[eliandoran]

Too complicated. Extract to a function with simple ifs and returns.