Skip to content

Build your personal knowledge base with Trilium Notes

feat(llm): choose DeepSeek V4's reasoning effort from the chat

Elian Doran4d0f53bff801
diff --git a/apps/client/src/widgets/type_widgets/options/llm/ModelSelection.spec.tsx b/apps/client/src/widgets/type_widgets/options/llm/ModelSelection.spec.tsx
index ac439e26e4..bcc227a877 100644
--- a/apps/client/src/widgets/type_widgets/options/llm/ModelSelection.spec.tsx
+++ b/apps/client/src/widgets/type_widgets/options/llm/ModelSelection.spec.tsx
@@ -96,6 +96,19 @@ describe("ModelSelection", () => {
expect(onChange).not.toHaveBeenCalled();
});
+ it("refreshes the stored metadata of selected models from the listing", async () => {
+ // Stored before the provider listed effort levels, with an old price; `gone` is no longer listed.
+ const stale: LlmModelInfo = { id: "gpt-4.1", name: "GPT-4.1", pricing: { input: 1, output: 1 } };
+ const gone: LlmModelInfo = { id: "gone", name: "Gone" };
+ const fresh: LlmModelInfo = { ...MODELS[0], reasoningEfforts: ["low", "high"], defaultReasoningEffort: "high" };
+ fetchProviderModelsMock.mockResolvedValue([fresh, MODELS[1], MODELS[2]]);
+ const onChange = vi.fn();
+
+ await renderSelection({ query: { provider: "openai" }, selected: [stale, gone], onChange, autoSelectDefaults: true });
+ expect(onChange).toHaveBeenCalledOnce();
+ expect(onChange).toHaveBeenCalledWith([fresh, gone]);
+ });
+
it("leaves an empty selection untouched when auto-select is off", async () => {
// A provider deliberately emptied to "hide all" — reopening its editor
// must not silently re-populate it (see shouldSeedDefaultModels).
diff --git a/apps/client/src/widgets/type_widgets/options/llm/ModelSelection.tsx b/apps/client/src/widgets/type_widgets/options/llm/ModelSelection.tsx
index 23f9075219..cb8267cdad 100644
--- a/apps/client/src/widgets/type_widgets/options/llm/ModelSelection.tsx
+++ b/apps/client/src/widgets/type_widgets/options/llm/ModelSelection.tsx
@@ -55,6 +55,11 @@ export default function ModelSelection({ query, timeoutMs, selected, onChange, a
setLoading(false);
if (autoSelectDefaults && selected.length === 0 && fetched.length > 0) {
onChange(defaultSelectedModels(fetched));
+ return;
+ }
+ const refreshed = refreshSelectedModels(selected, fetched);
+ if (refreshed) {
+ onChange(refreshed);
}
})
.catch(err => {
@@ -129,6 +134,17 @@ function defaultSelectedModels(models: LlmModelInfo[]): LlmModelInfo[] {
return models.filter(model => model.recommended);
}
+/**
+ * The selection with each stored model replaced by its listed entry, so saving the provider picks
+ * up new metadata such as `reasoningEfforts` or a changed price. A model the listing no longer
+ * carries stays as stored. Returns undefined when nothing changed.
+ */
+function refreshSelectedModels(selected: LlmModelInfo[], fetched: LlmModelInfo[]): LlmModelInfo[] | undefined {
+ const listed = new Map(fetched.map(model => [model.id, model]));
+ const refreshed = selected.map(model => listed.get(model.id) ?? model);
+ return JSON.stringify(refreshed) === JSON.stringify(selected) ? undefined : refreshed;
+}
+
/** Row label: model name plus a cost hint (per-Mtok price / subscription) when known. */
function ModelLabel({ model }: { model: LlmModelInfo }) {
const cost = formatModelCost(model);
diff --git a/apps/server/src/assets/doc_notes/en/User Guide/User Guide/AI.html b/apps/server/src/assets/doc_notes/en/User Guide/User Guide/AI.html
index a74a269c9e..6ec3da068c 100644
--- a/apps/server/src/assets/doc_notes/en/User Guide/User Guide/AI.html
+++ b/apps/server/src/assets/doc_notes/en/User Guide/User Guide/AI.html
@@ -119,6 +119,24 @@ class="admonition note">
<p>Currently only the search native to the LLM provider is supported. External
search providers such as Exa, Tavily &amp; SearXNG are not yet supported.</p>
</aside>
+<h3>Thinking</h3>
+<p>Some models reason before they answer. Their reasoning appears above the
+ reply as a collapsible <em>Thought process</em> line.</p>
+<p>How much a model thinks is set per chat, in one of two ways depending
+ on the model:</p>
+<ul>
+ <li>Most models have an <em>Extended thinking</em> switch, next to <em>Web search</em>.</li>
+ <li>Models that offer several levels of reasoning (DeepSeek V4, OpenAI Codex,
+ Antigravity) have a <span class="tn-icon bx bx-brain"></span> reasoning effort
+ dropdown next to the model selector instead. <em>None</em>, where offered,
+ turns thinking off; higher levels give better answers to hard questions,
+ but take longer and cost more.</li>
+</ul>
+<aside class="admonition note">
+ <p>The effort dropdown only appears once Trilium knows the model's levels.
+ For a DeepSeek provider set up with an earlier version of Trilium, edit
+ the provider in the model selection box and press <em>Save</em> once.</p>
+</aside>
<h3>Note access (tools)</h3>
<p>Tools allow the agentic AI to understand and operate on notes directly
within your Trilium instance.</p>
diff --git a/docs/User Guide/User Guide/AI.md b/docs/User Guide/User Guide/AI.md
index 73742b9d47..226a400bc1 100644
--- a/docs/User Guide/User Guide/AI.md
+++ b/docs/User Guide/User Guide/AI.md
@@ -90,6 +90,18 @@ This feature is on by default but it can easily be disabled by clicking on the m
> [!NOTE]
> Currently only the search native to the LLM provider is supported. External search providers such as Exa, Tavily & SearXNG are not yet supported.
+### Thinking
+
+Some models reason before they answer. Their reasoning appears above the reply as a collapsible _Thought process_ line.
+
+How much a model thinks is set per chat, in one of two ways depending on the model:
+
+* Most models have an _Extended thinking_ switch, next to _Web search_.
+* Models that offer several levels of reasoning (DeepSeek V4, OpenAI Codex, Antigravity) have a <span class="tn-icon bx bx-brain"></span> reasoning effort dropdown next to the model selector instead. _None_, where offered, turns thinking off; higher levels give better answers to hard questions, but take longer and cost more.
+
+> [!NOTE]
+> The effort dropdown only appears once Trilium knows the model's levels. For a DeepSeek provider set up with an earlier version of Trilium, edit the provider in the model selection box and press _Save_ once.
+
### Note access (tools)
Tools allow the agentic AI to understand and operate on notes directly within your Trilium instance.
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 5067f174e8..3c5b3832aa 100644
--- a/packages/trilium-core/src/services/llm/providers/base_provider.ts
+++ b/packages/trilium-core/src/services/llm/providers/base_provider.ts
@@ -413,6 +413,10 @@ export abstract class BaseProvider implements LlmProvider {
onError: () => {},
telemetry: TELEMETRY_OFF
};
+ const providerOptions = this.chatProviderOptions(config);
+ if (providerOptions) {
+ streamOptions.providerOptions = providerOptions;
+ }
const tools = this.buildTools(config);
if (Object.keys(tools).length > 0) {
@@ -424,6 +428,11 @@ export abstract class BaseProvider implements LlmProvider {
return streamText(streamOptions);
}
+ /** Provider-specific `providerOptions` for a chat turn, such as a reasoning effort. */
+ protected chatProviderOptions(_config: LlmProviderConfig): Parameters<typeof streamText>[0]["providerOptions"] {
+ return undefined;
+ }
+
/**
* This provider's slice of the committed price table. Overridable so tests
* can inject a table without a real JSON file.
diff --git a/packages/trilium-core/src/services/llm/providers/deepseek.spec.ts b/packages/trilium-core/src/services/llm/providers/deepseek.spec.ts
index b3dfad4af0..227ff56b63 100644
--- a/packages/trilium-core/src/services/llm/providers/deepseek.spec.ts
+++ b/packages/trilium-core/src/services/llm/providers/deepseek.spec.ts
@@ -30,6 +30,7 @@ import type { LlmStreamChunk } from "@triliumnext/commons";
import { installGlobalFetchAsApiTransport } from "../../../test/request_provider.js";
import { streamToChunks } from "../stream.js";
+import type { LlmProviderConfig } from "../types.js";
import { DeepSeekProvider, deepSeekModelName } from "./deepseek.js";
import { llmFetch } from "./fetch.js";
@@ -213,8 +214,7 @@ describe("DeepSeekProvider streaming", () => {
{ choices: [{ index: 0, delta: { content: "Pick B." } }] },
{ choices: [{ index: 0, delta: {}, finish_reason: "stop" }], usage: { prompt_tokens: 5, completion_tokens: 7 } }
];
- const body = `${events.map(e => `data: ${JSON.stringify({ id: "c1", created: 0, model: "deepseek-v4-pro", ...e })}\n\n`).join("")}data: [DONE]\n\n`;
- const fetchMock = vi.fn(async () => new Response(body, { headers: { "content-type": "text/event-stream" } }));
+ const fetchMock = vi.fn(async () => sseResponse(events));
vi.stubGlobal("fetch", fetchMock);
const provider = new DeepSeekProvider("sk-deep");
@@ -232,6 +232,58 @@ describe("DeepSeekProvider streaming", () => {
});
});
+describe("DeepSeekProvider reasoning effort", () => {
+ const fetchMock = vi.fn();
+ const okJson = (body: unknown) => ({ ok: true, json: async () => body });
+
+ beforeEach(() => {
+ fetchMock.mockReset();
+ vi.stubGlobal("fetch", fetchMock);
+ });
+
+ it("offers the effort levels on V4 models only", async () => {
+ fetchMock.mockResolvedValue(okJson({ data: [{ id: "deepseek-v4-pro" }, { id: "deepseek-flash" }, { id: "deepseek-chat" }] }));
+ const models = await new DeepSeekProvider("sk-deep").listModels();
+
+ for (const id of ["deepseek-v4-pro", "deepseek-flash"]) {
+ expect(models.find(m => m.id === id)).toMatchObject({
+ reasoningEfforts: ["none", "low", "high", "max"],
+ defaultReasoningEffort: "high"
+ });
+ }
+ const legacy = models.find(m => m.id === "deepseek-chat");
+ expect(legacy).toBeDefined();
+ expect(legacy).not.toHaveProperty("reasoningEfforts");
+ });
+
+ it("sends the chosen effort, and the default when the chat has none", async () => {
+ fetchMock.mockImplementation(async () => sseResponse([
+ { choices: [{ index: 0, delta: { role: "assistant", content: "Ok" }, finish_reason: "stop" }] }
+ ]));
+ const provider = new DeepSeekProvider("sk-deep");
+ const requestBody = async (config: LlmProviderConfig) => {
+ fetchMock.mockClear();
+ for await (const _ of streamToChunks(provider.chat([{ role: "user", content: "Hi" }], config))) { /* drain */ }
+ expect(fetchMock).toHaveBeenCalledOnce();
+ return JSON.parse(String((fetchMock.mock.calls[0][1] as RequestInit).body));
+ };
+
+ expect(await requestBody({ model: "deepseek-v4-pro", reasoningEffort: "max" }))
+ .toMatchObject({ thinking: { type: "enabled" }, reasoning_effort: "max" });
+ expect(await requestBody({ model: "deepseek-v4-pro" }))
+ .toMatchObject({ thinking: { type: "enabled" }, reasoning_effort: "high" });
+
+ const none = await requestBody({ model: "deepseek-v4-pro", reasoningEffort: "none" });
+ expect(none).toMatchObject({ thinking: { type: "disabled" } });
+ expect(none).not.toHaveProperty("reasoning_effort");
+
+ // A model without levels is left to DeepSeek's own default.
+ const legacy = await requestBody({ model: "deepseek-chat", reasoningEffort: "max" });
+ expect(legacy).not.toHaveProperty("thinking");
+ expect(legacy).not.toHaveProperty("reasoning_effort");
+ });
+});
+
describe("DeepSeekProvider recommendations", () => {
it("pre-selects the current line but not the superseded coder models", () => {
const provider = new DeepSeekProvider("sk-deep");
@@ -256,3 +308,9 @@ describe("deepSeekModelName", () => {
expect(deepSeekModelName("llama3.2")).toBe("llama3.2");
});
});
+
+/** A Chat Completions stream carrying `events`, as DeepSeek sends it. */
+function sseResponse(events: object[]) {
+ const body = `${events.map(e => `data: ${JSON.stringify({ id: "c1", created: 0, model: "deepseek-v4-pro", ...e })}\n\n`).join("")}data: [DONE]\n\n`;
+ return new Response(body, { headers: { "content-type": "text/event-stream" } });
+}
diff --git a/packages/trilium-core/src/services/llm/providers/deepseek.ts b/packages/trilium-core/src/services/llm/providers/deepseek.ts
index a305db99d1..a16acbb8c2 100644
--- a/packages/trilium-core/src/services/llm/providers/deepseek.ts
+++ b/packages/trilium-core/src/services/llm/providers/deepseek.ts
@@ -1,6 +1,7 @@
-import { createDeepSeek, type DeepSeekProvider as DeepSeekSDKProvider } from "@ai-sdk/deepseek";
+import { createDeepSeek, type DeepSeekLanguageModelChatOptions, type DeepSeekProvider as DeepSeekSDKProvider } from "@ai-sdk/deepseek";
+import type { LlmReasoningEffort } from "@triliumnext/commons";
-import type { ModelInfo } from "../types.js";
+import type { LlmProviderConfig, ModelInfo } from "../types.js";
import { BaseProvider, type RemoteModel } from "./base_provider.js";
import { llmFetch } from "./fetch.js";
@@ -11,6 +12,11 @@ import { llmFetch } from "./fetch.js";
*/
const OFFICIAL_BASE_URL = "https://api.deepseek.com/v1";
+/** The thinking levels of a V4 model, weakest first: `none` turns thinking off, the rest set `reasoning_effort`. */
+const V4_REASONING_EFFORTS = [ "none", "low", "high", "max" ] as const satisfies LlmReasoningEffort[];
+type V4ReasoningEffort = (typeof V4_REASONING_EFFORTS)[number];
+const V4_DEFAULT_REASONING_EFFORT: V4ReasoningEffort = "high";
+
/**
* DeepSeek, over its OpenAI-compatible API. `@ai-sdk/deepseek` rather than
* `@ai-sdk/openai` reads the thinking from `reasoning_content` and sends it back
@@ -48,6 +54,29 @@ export class DeepSeekProvider extends BaseProvider {
return this.deepseek.chat(modelId);
}
+ /** V4 models get the reasoning effort dropdown in place of the extended thinking switch. */
+ override async listModels(): Promise<ModelInfo[]> {
+ const models = await super.listModels();
+ return models.map(m => isDeepSeekV4Model(m.id)
+ ? { ...m, reasoningEfforts: [ ...V4_REASONING_EFFORTS ], defaultReasoningEffort: V4_DEFAULT_REASONING_EFFORT }
+ : m);
+ }
+
+ /**
+ * Sends the chat's effort to a V4 model, or the default when the chat has none, so the
+ * level the dropdown shows is the one DeepSeek runs at.
+ */
+ protected override chatProviderOptions(config: LlmProviderConfig) {
+ if (!isDeepSeekV4Model(config.model || this.defaultModel)) {
+ return undefined;
+ }
+ const effort = V4_REASONING_EFFORTS.find(level => level === config.reasoningEffort) ?? V4_DEFAULT_REASONING_EFFORT;
+ const options: DeepSeekLanguageModelChatOptions = effort === "none"
+ ? { thinking: { type: "disabled" } }
+ : { thinking: { type: "enabled" }, reasoningEffort: effort };
+ return { deepseek: options };
+ }
+
/** Everything DeepSeek lists is a chat model, so nothing is filtered out. */
protected override async fetchRemoteModels(): Promise<RemoteModel[] | null> {
const payload = await this.fetchJson(`${this.baseURL ?? OFFICIAL_BASE_URL}/models`, {
@@ -148,3 +177,8 @@ export function deepSeekModelName(id: string): string {
.map(word => (/^v\d+$/i.test(word) ? word.toUpperCase() : word.charAt(0).toUpperCase() + word.slice(1)));
return `DeepSeek ${words.join(" ")}`;
}
+
+/** The V4 line, which takes `thinking` and `reasoning_effort`; the same test `@ai-sdk/deepseek` applies. */
+function isDeepSeekV4Model(id: string): boolean {
+ return id.includes("deepseek-v4") || id.startsWith("deepseek-flash") || id.startsWith("deepseek-pro");
+}