Skip to content

Build your personal knowledge base with Trilium Notes

fix(llm): generate chat titles without reasoning

Elian Dorana98023b4d8fc
diff --git a/packages/trilium-core/src/services/llm/providers/base_provider.spec.ts b/packages/trilium-core/src/services/llm/providers/base_provider.spec.ts
index 293c391573..e9fd8c84ae 100644
--- a/packages/trilium-core/src/services/llm/providers/base_provider.spec.ts
+++ b/packages/trilium-core/src/services/llm/providers/base_provider.spec.ts
@@ -601,6 +601,7 @@ describe("BaseProvider chat / pricing / models / title", () => {
expect(provider.createdModelIds).toContain("cheap");
const args = generateTextMock.mock.calls[0][0] as any;
expect(args.maxOutputTokens).toBe(30);
+ expect(args.reasoning).toBe("none");
expect(args.messages[0].content).toContain("Some long first message");
});
diff --git a/packages/trilium-core/src/services/llm/providers/base_provider.ts b/packages/trilium-core/src/services/llm/providers/base_provider.ts
index dcd6ab9e2b..5067f174e8 100644
--- a/packages/trilium-core/src/services/llm/providers/base_provider.ts
+++ b/packages/trilium-core/src/services/llm/providers/base_provider.ts
@@ -518,11 +518,19 @@ export abstract class BaseProvider implements LlmProvider {
*/
private titleNeedsRoomToThink = false;
+ /**
+ * Whether a title call asks the model not to reason (`reasoning: "none"`). Each
+ * provider package maps that to its own switch, such as DeepSeek's
+ * `thinking: { type: "disabled" }`.
+ */
+ protected titleSkipsReasoning = true;
+
/** One title call, with whatever the caller is willing to spend on it. */
private async requestTitle(firstMessage: string, maxOutputTokens: number) {
const { text, finishReason, usage } = await generateText({
model: this.createModel(this.titleModel),
maxOutputTokens,
+ ...(this.titleSkipsReasoning && { reasoning: "none" as const }),
telemetry: TELEMETRY_OFF,
messages: [
{
diff --git a/packages/trilium-core/src/services/llm/providers/local.spec.ts b/packages/trilium-core/src/services/llm/providers/local.spec.ts
index 46f0226a92..9f91ef09b2 100644
--- a/packages/trilium-core/src/services/llm/providers/local.spec.ts
+++ b/packages/trilium-core/src/services/llm/providers/local.spec.ts
@@ -16,6 +16,15 @@ vi.mock("@ai-sdk/openai", () => ({
}
}));
+const { generateTextMock } = vi.hoisted(() => ({
+ generateTextMock: vi.fn(async (..._args: unknown[]) => ({ text: "A title", finishReason: "stop", usage: {} }) as any)
+}));
+
+vi.mock("ai", async (importOriginal) => {
+ const actual = await importOriginal<typeof import("ai")>();
+ return { ...actual, generateText: generateTextMock };
+});
+
import { LocalProvider } from "./local.js";
import { installGlobalFetchAsApiTransport } from "../../../test/request_provider.js";
import { llmFetch } from "./fetch.js";
@@ -464,5 +473,15 @@ describe("LocalProvider", () => {
expect(fetchMock.mock.calls.length).toBe(probes);
generateTitle.mockRestore();
});
+
+ it("leaves the title call's reasoning to the server", async () => {
+ // Servers disagree on `reasoning_effort: "none"`, so a thinking model gets the larger retry instead.
+ fetchMock.mockImplementation(routes({ "/api/tags": ollamaTags([{ name: "tiny", parameter_size: "1B" }]) }));
+ generateTextMock.mockClear();
+
+ await expect(new LocalProvider("ollama").generateTitle("Hello")).resolves.toBe("A title");
+ expect(generateTextMock).toHaveBeenCalledOnce();
+ expect(generateTextMock.mock.calls[0][0]).not.toHaveProperty("reasoning");
+ });
});
});
diff --git a/packages/trilium-core/src/services/llm/providers/local.ts b/packages/trilium-core/src/services/llm/providers/local.ts
index 1c4fbaf598..b5ed4e1b47 100644
--- a/packages/trilium-core/src/services/llm/providers/local.ts
+++ b/packages/trilium-core/src/services/llm/providers/local.ts
@@ -73,6 +73,11 @@ export class LocalProvider extends BaseProvider {
// whatever the user installed, so there is nothing sensible to hardcode.
protected defaultModel = "";
protected titleModel = "";
+ /**
+ * The OpenAI package sends `reasoning_effort: "none"` to any model, and self-hosted
+ * servers disagree on what that means, so a thinking model gets the larger retry instead.
+ */
+ protected override titleSkipsReasoning = false;
private readonly kind: LocalProviderKind;
private openai: OpenAISDKProvider;