replaceSystemMessage model config option

core/index.d.ts

@@ -43,13 +43,13 @@ export interface IndexingProgressUpdate {
   desc: string;
   shouldClearIndexes?: boolean;
   status:
-    | "loading"
-    | "indexing"
-    | "done"
-    | "failed"
-    | "paused"
-    | "disabled"
-    | "cancelled";
+  | "loading"
+  | "indexing"
+  | "done"
+  | "failed"
+  | "paused"
+  | "disabled"
+  | "cancelled";
   debugInfo?: string;
 }
 
@@ -450,6 +450,7 @@ export interface LLMOptions {
   title?: string;
   uniqueId?: string;
   systemMessage?: string;
+  replaceSystemMessage?: boolean;
   contextLength?: number;
   maxStopWords?: number;
   completionOptions?: CompletionOptions;
@@ -650,10 +651,10 @@ export interface IDE {
   getCurrentFile(): Promise<
     | undefined
     | {
-        isUntitled: boolean;
-        path: string;
-        contents: string;
-      }
+      isUntitled: boolean;
+      path: string;
+      contents: string;
+    }
   >;
 
   getPinnedFiles(): Promise<string[]>;
@@ -832,11 +833,11 @@ export interface CustomCommand {
 export interface Prediction {
   type: "content";
   content:
-    | string
-    | {
-        type: "text";
-        text: string;
-      }[];
+  | string
+  | {
+    type: "text";
+    text: string;
+  }[];
 }
 
 export interface ToolExtras {
@@ -903,6 +904,7 @@ export interface ModelDescription {
   template?: TemplateType;
   completionOptions?: BaseCompletionOptions;
   systemMessage?: string;
+  replaceSystemMessage?: boolean;
   requestOptions?: RequestOptions;
   promptTemplates?: { [key: string]: string };
   capabilities?: ModelCapability;
@@ -1157,9 +1159,9 @@ export interface Config {
   embeddingsProvider?: EmbeddingsProviderDescription | ILLM;
   /** The model that Continue will use for tab autocompletions. */
   tabAutocompleteModel?:
-    | CustomLLM
-    | ModelDescription
-    | (CustomLLM | ModelDescription)[];
+  | CustomLLM
+  | ModelDescription
+  | (CustomLLM | ModelDescription)[];
   /** Options for tab autocomplete */
   tabAutocompleteOptions?: Partial<TabAutocompleteOptions>;
   /** UI styles customization */
@@ -1249,9 +1251,9 @@ export type PackageDetailsSuccess = PackageDetails & {
 export type PackageDocsResult = {
   packageInfo: ParsedPackageInfo;
 } & (
-  | { error: string; details?: never }
-  | { details: PackageDetailsSuccess; error?: never }
-);
+    | { error: string; details?: never }
+    | { details: PackageDetailsSuccess; error?: never }
+  );
 
 export interface TerminalOptions {
   reuseTerminal?: boolean;

core/llm/countTokens.ts

@@ -1,6 +1,6 @@
 import { Tiktoken, encodingForModel as _encodingForModel } from "js-tiktoken";
 
-import { ChatMessage, MessageContent, MessagePart } from "../index.js";
+import { ChatMessage, LLMOptions, MessageContent, MessagePart } from "../index.js";
 
 import { renderChatMessage } from "../util/messageContent.js";
 import {
@@ -27,9 +27,9 @@ class LlamaEncoding implements Encoding {
 }
 
 class NonWorkerAsyncEncoder implements AsyncEncoder {
-  constructor(private readonly encoding: Encoding) {}
+  constructor(private readonly encoding: Encoding) { }
 
-  async close(): Promise<void> {}
+  async close(): Promise<void> { }
 
   async encode(text: string): Promise<number[]> {
     return this.encoding.encode(text);
@@ -380,15 +380,18 @@ function compileChatMessages(
   prompt: string | undefined = undefined,
   functions: any[] | undefined = undefined,
   systemMessage: string | undefined = undefined,
+  llmOptions?: LLMOptions,
 ): ChatMessage[] {
   let msgsCopy = msgs
     ? msgs
-        .map((msg) => ({ ...msg }))
-        .filter((msg) => !chatMessageIsEmpty(msg) && msg.role !== "system")
+      .map((msg) => ({ ...msg }))
+      .filter((msg) => !chatMessageIsEmpty(msg) && msg.role !== "system")
     : [];
 
   msgsCopy = addSpaceToAnyEmptyMessages(msgsCopy);
 
+  const replaceSystemMessage = llmOptions?.replaceSystemMessage ?? false;
+
   if (prompt) {
     const promptMsg: ChatMessage = {
       role: "user",
@@ -406,11 +409,15 @@ function compileChatMessages(
       content = renderChatMessage(msgs?.[0]);
     }
     if (systemMessage && systemMessage.trim() !== "") {
-      const shouldAddNewLines = content !== "";
-      if (shouldAddNewLines) {
-        content += "\n\n";
+      if (replaceSystemMessage) {
+        content = systemMessage;
+      } else {
+        const shouldAddNewLines = content !== "";
+        if (shouldAddNewLines) {
+          content += "\n\n";
+        }
+        content += systemMessage;
       }
-      content += systemMessage;
     }
     const systemChatMsg: ChatMessage = {
       role: "system",

core/llm/index.ts

@@ -173,11 +173,11 @@ export abstract class BaseLLM implements ILLM {
         options.completionOptions?.maxTokens ??
         (llmInfo?.maxCompletionTokens
           ? Math.min(
-              llmInfo.maxCompletionTokens,
-              // Even if the model has a large maxTokens, we don't want to use that every time,
-              // because it takes away from the context length
-              this.contextLength / 4,
-            )
+            llmInfo.maxCompletionTokens,
+            // Even if the model has a large maxTokens, we don't want to use that every time,
+            // because it takes away from the context length
+            this.contextLength / 4,
+          )
           : DEFAULT_MAX_TOKENS),
     };
     this.requestOptions = options.requestOptions;
@@ -261,6 +261,7 @@ export abstract class BaseLLM implements ILLM {
       undefined,
       functions,
       this.systemMessage,
+      this._llmOptions,
     );
   }
 

core/llm/llms/index.ts

@@ -137,6 +137,7 @@ export async function llmFromDescription(
         cls.defaultOptions?.completionOptions?.maxTokens,
     },
     systemMessage,
+    replaceSystemMessage: desc.replaceSystemMessage ?? false,
     writeLog,
     uniqueId,
   };

docs/docs/reference.md

@@ -25,7 +25,8 @@ Each model has specific configuration options tailored to its provider and funct
 - `template`: Chat template to format messages. Auto-detected for most models but can be overridden. See intelliJ suggestions.
 - `promptTemplates`: A mapping of prompt template names (e.g., `edit`) to template strings. [Customization Guide](https://docs.continue.dev/model-setup/configuration#customizing-the-edit-prompt).
 - `completionOptions`: Model-specific completion options, same format as top-level [`completionOptions`](#completionoptions), which they override.
-- `systemMessage`: A system message that will precede responses from the LLM.
+- `systemMessage`: A system message that will precede responses from the LLM. (Note: this property has higher precendence than the root `systemMessage` property)
+- `replaceSystemMessage`: If `true`, replaces the system message with the one specified in the model config's `systemMessage` property or with the root `systemMessage` property. If `false`, appends the `systemMessage` to the default system message, if any exists for the model. (default: `false`)
 - `requestOptions`: Model-specific HTTP request options, same format as top-level [`requestOptions`](#requestoptions), which they override.
 - `apiType`: Specifies the type of API (`openai` or `azure`).
 - `apiVersion`: Azure API version (e.g., `2023-07-01-preview`).
@@ -381,7 +382,7 @@ An optional token that identifies the user, primarily for authenticated services
 
 ### `systemMessage`
 
-Defines a system message that appears before every response from the language model, providing guidance or context.
+Defines a system message that appears before every response from the language model, providing guidance or context. Note: `systemMessage` in the model config takes precedence over this setting.
 
 ### `disableIndexing`
 

extensions/vscode/config_schema.json

@@ -400,6 +400,11 @@
           "description": "A system message that will always be followed by the LLM",
           "type": "string"
         },
+        "replaceSystemMessage": {
+          "title": "Replace System Message",
+          "description": "If true, the system message will replace the default system message instead of being appended to it",
+          "type": "boolean"
+        },
         "requestOptions": {
           "title": "Request Options",
           "description": "Options for the HTTP request to the LLM.",