fix: respect Ollama Modelfile num_ctx configuration (#7798)

Co-authored-by: Roo Code <[email protected]>
Co-authored-by: Hannes Rudolph <[email protected]>
Co-authored-by: daniel-lxs <[email protected]>

Author: 4 people

packages/types/src/provider-settings.ts

@@ -258,6 +258,7 @@ const ollamaSchema = baseProviderSettingsSchema.extend({
 	ollamaModelId: z.string().optional(),
 	ollamaBaseUrl: z.string().optional(),
 	ollamaApiKey: z.string().optional(),
+	ollamaNumCtx: z.number().int().min(128).optional(),
 })
 
 const vsCodeLmSchema = baseProviderSettingsSchema.extend({

src/api/providers/__tests__/native-ollama.spec.ts

@@ -73,6 +73,61 @@ describe("NativeOllamaHandler", () => {
 			expect(results[2]).toEqual({ type: "usage", inputTokens: 10, outputTokens: 2 })
 		})
 
+		it("should not include num_ctx by default", async () => {
+			// Mock the chat response
+			mockChat.mockImplementation(async function* () {
+				yield { message: { content: "Response" } }
+			})
+
+			const stream = handler.createMessage("System", [{ role: "user" as const, content: "Test" }])
+
+			// Consume the stream
+			for await (const _ of stream) {
+				// consume stream
+			}
+
+			// Verify that num_ctx was NOT included in the options
+			expect(mockChat).toHaveBeenCalledWith(
+				expect.objectContaining({
+					options: expect.not.objectContaining({
+						num_ctx: expect.anything(),
+					}),
+				}),
+			)
+		})
+
+		it("should include num_ctx when explicitly set via ollamaNumCtx", async () => {
+			const options: ApiHandlerOptions = {
+				apiModelId: "llama2",
+				ollamaModelId: "llama2",
+				ollamaBaseUrl: "http://localhost:11434",
+				ollamaNumCtx: 8192, // Explicitly set num_ctx
+			}
+
+			handler = new NativeOllamaHandler(options)
+
+			// Mock the chat response
+			mockChat.mockImplementation(async function* () {
+				yield { message: { content: "Response" } }
+			})
+
+			const stream = handler.createMessage("System", [{ role: "user" as const, content: "Test" }])
+
+			// Consume the stream
+			for await (const _ of stream) {
+				// consume stream
+			}
+
+			// Verify that num_ctx was included with the specified value
+			expect(mockChat).toHaveBeenCalledWith(
+				expect.objectContaining({
+					options: expect.objectContaining({
+						num_ctx: 8192,
+					}),
+				}),
+			)
+		})
+
 		it("should handle DeepSeek R1 models with reasoning detection", async () => {
 			const options: ApiHandlerOptions = {
 				apiModelId: "deepseek-r1",
@@ -120,6 +175,49 @@ describe("NativeOllamaHandler", () => {
 			})
 			expect(result).toBe("This is the response")
 		})
+
+		it("should not include num_ctx in completePrompt by default", async () => {
+			mockChat.mockResolvedValue({
+				message: { content: "Response" },
+			})
+
+			await handler.completePrompt("Test prompt")
+
+			// Verify that num_ctx was NOT included in the options
+			expect(mockChat).toHaveBeenCalledWith(
+				expect.objectContaining({
+					options: expect.not.objectContaining({
+						num_ctx: expect.anything(),
+					}),
+				}),
+			)
+		})
+
+		it("should include num_ctx in completePrompt when explicitly set", async () => {
+			const options: ApiHandlerOptions = {
+				apiModelId: "llama2",
+				ollamaModelId: "llama2",
+				ollamaBaseUrl: "http://localhost:11434",
+				ollamaNumCtx: 4096, // Explicitly set num_ctx
+			}
+
+			handler = new NativeOllamaHandler(options)
+
+			mockChat.mockResolvedValue({
+				message: { content: "Response" },
+			})
+
+			await handler.completePrompt("Test prompt")
+
+			// Verify that num_ctx was included with the specified value
+			expect(mockChat).toHaveBeenCalledWith(
+				expect.objectContaining({
+					options: expect.objectContaining({
+						num_ctx: 4096,
+					}),
+				}),
+			)
+		})
 	})
 
 	describe("error handling", () => {

src/api/providers/native-ollama.ts

@@ -8,6 +8,11 @@ import { getOllamaModels } from "./fetchers/ollama"
 import { XmlMatcher } from "../../utils/xml-matcher"
 import type { SingleCompletionHandler, ApiHandlerCreateMessageMetadata } from "../index"
 
+interface OllamaChatOptions {
+	temperature: number
+	num_ctx?: number
+}
+
 function convertToOllamaMessages(anthropicMessages: Anthropic.Messages.MessageParam[]): Message[] {
 	const ollamaMessages: Message[] = []
 
@@ -184,15 +189,22 @@ export class NativeOllamaHandler extends BaseProvider implements SingleCompletio
 		)
 
 		try {
+			// Build options object conditionally
+			const chatOptions: OllamaChatOptions = {
+				temperature: this.options.modelTemperature ?? (useR1Format ? DEEP_SEEK_DEFAULT_TEMPERATURE : 0),
+			}
+
+			// Only include num_ctx if explicitly set via ollamaNumCtx
+			if (this.options.ollamaNumCtx !== undefined) {
+				chatOptions.num_ctx = this.options.ollamaNumCtx
+			}
+
 			// Create the actual API request promise
 			const stream = await client.chat({
 				model: modelId,
 				messages: ollamaMessages,
 				stream: true,
-				options: {
-					num_ctx: modelInfo.contextWindow,
-					temperature: this.options.modelTemperature ?? (useR1Format ? DEEP_SEEK_DEFAULT_TEMPERATURE : 0),
-				},
+				options: chatOptions,
 			})
 
 			let totalInputTokens = 0
@@ -274,13 +286,21 @@ export class NativeOllamaHandler extends BaseProvider implements SingleCompletio
 			const { id: modelId } = await this.fetchModel()
 			const useR1Format = modelId.toLowerCase().includes("deepseek-r1")
 
+			// Build options object conditionally
+			const chatOptions: OllamaChatOptions = {
+				temperature: this.options.modelTemperature ?? (useR1Format ? DEEP_SEEK_DEFAULT_TEMPERATURE : 0),
+			}
+
+			// Only include num_ctx if explicitly set via ollamaNumCtx
+			if (this.options.ollamaNumCtx !== undefined) {
+				chatOptions.num_ctx = this.options.ollamaNumCtx
+			}
+
 			const response = await client.chat({
 				model: modelId,
 				messages: [{ role: "user", content: prompt }],
 				stream: false,
-				options: {
-					temperature: this.options.modelTemperature ?? (useR1Format ? DEEP_SEEK_DEFAULT_TEMPERATURE : 0),
-				},
+				options: chatOptions,
 			})
 
 			return response.message?.content || ""

src/shared/api.ts

@@ -18,6 +18,12 @@ export type ApiHandlerOptions = Omit<ProviderSettings, "apiProvider"> & {
 	 * Defaults to true; set to false to disable summaries.
 	 */
 	enableGpt5ReasoningSummary?: boolean
+	/**
+	 * Optional override for Ollama's num_ctx parameter.
+	 * When set, this value will be used in Ollama chat requests.
+	 * When undefined, Ollama will use the model's default num_ctx from the Modelfile.
+	 */
+	ollamaNumCtx?: number
 }
 
 // RouterName

webview-ui/src/components/settings/providers/Ollama.tsx

@@ -130,6 +130,26 @@ export const Ollama = ({ apiConfiguration, setApiConfigurationField }: OllamaPro
 					))}
 				</VSCodeRadioGroup>
 			)}
+			<VSCodeTextField
+				value={apiConfiguration?.ollamaNumCtx?.toString() || ""}
+				onInput={(e: any) => {
+					const value = e.target?.value
+					if (value === "") {
+						setApiConfigurationField("ollamaNumCtx", undefined)
+					} else {
+						const numValue = parseInt(value, 10)
+						if (!isNaN(numValue) && numValue >= 128) {
+							setApiConfigurationField("ollamaNumCtx", numValue)
+						}
+					}
+				}}
+				placeholder="e.g., 4096"
+				className="w-full">
+				<label className="block font-medium mb-1">{t("settings:providers.ollama.numCtx")}</label>
+				<div className="text-xs text-vscode-descriptionForeground mt-1">
+					{t("settings:providers.ollama.numCtxHelp")}
+				</div>
+			</VSCodeTextField>
 			<div className="text-sm text-vscode-descriptionForeground">
 				{t("settings:providers.ollama.description")}
 				<span className="text-vscode-errorForeground ml-1">{t("settings:providers.ollama.warning")}</span>

webview-ui/src/components/ui/hooks/useSelectedModel.ts

@@ -263,9 +263,17 @@ function getSelectedModel({
 		case "ollama": {
 			const id = apiConfiguration.ollamaModelId ?? ""
 			const info = ollamaModels && ollamaModels[apiConfiguration.ollamaModelId!]
+
+			const adjustedInfo =
+				info?.contextWindow &&
+				apiConfiguration?.ollamaNumCtx &&
+				apiConfiguration.ollamaNumCtx < info.contextWindow
+					? { ...info, contextWindow: apiConfiguration.ollamaNumCtx }
+					: info
+
 			return {
 				id,
-				info: info || undefined,
+				info: adjustedInfo || undefined,
 			}
 		}
 		case "lmstudio": {

webview-ui/src/i18n/locales/ca/settings.json

@@ -380,6 +380,8 @@
 			"modelId": "Model ID",
 			"apiKey": "Ollama API Key",
 			"apiKeyHelp": "Optional API key for authenticated Ollama instances or cloud services. Leave empty for local installations.",
+			"numCtx": "Context Window Size (num_ctx)",
+			"numCtxHelp": "Override the model's default context window size. Leave empty to use the model's Modelfile configuration. Minimum value is 128.",
 			"description": "Ollama allows you to run models locally on your computer. For instructions on how to get started, see their quickstart guide.",
 			"warning": "Note: Roo Code uses complex prompts and works best with Claude models. Less capable models may not work as expected."
 		},