feat (ai/core): add system message support in messages list (#1551)

Author: lgrammel

.changeset/dirty-falcons-love.md

@@ -0,0 +1,6 @@
+---
+'@ai-sdk/anthropic': patch
+'ai': patch
+---
+
+feat (ai/core): add system message support in messages list

examples/ai-core/src/generate-text/openai-system-message-a.ts

@@ -0,0 +1,19 @@
+import { openai } from '@ai-sdk/openai';
+import { generateText } from 'ai';
+import dotenv from 'dotenv';
+
+dotenv.config();
+
+async function main() {
+  const result = await generateText({
+    model: openai('gpt-3.5-turbo'),
+    messages: [
+      { role: 'system', content: 'You are a helpful assistant.' },
+      { role: 'user', content: 'What is the capital of France?' },
+    ],
+  });
+
+  console.log(result.text);
+}
+
+main().catch(console.error);

examples/ai-core/src/generate-text/openai-system-message-b.ts

@@ -0,0 +1,17 @@
+import { openai } from '@ai-sdk/openai';
+import { generateText } from 'ai';
+import dotenv from 'dotenv';
+
+dotenv.config();
+
+async function main() {
+  const result = await generateText({
+    model: openai('gpt-3.5-turbo'),
+    system: 'You are a helpful assistant.',
+    messages: [{ role: 'user', content: 'What is the capital of France?' }],
+  });
+
+  console.log(result.text);
+}
+
+main().catch(console.error);

packages/anthropic/src/convert-to-anthropic-messages-prompt.ts

@@ -11,12 +11,18 @@ import {
 export function convertToAnthropicMessagesPrompt(
   prompt: LanguageModelV1Prompt,
 ): AnthropicMessagesPrompt {
-  let system: string | undefined;
+  let system: string | undefined = undefined;
   const messages: AnthropicMessage[] = [];
 
   for (const { role, content } of prompt) {
     switch (role) {
       case 'system': {
+        if (system != null) {
+          throw new UnsupportedFunctionalityError({
+            functionality: 'Multiple system messages',
+          });
+        }
+
         system = content;
         break;
       }

packages/core/core/prompt/convert-to-language-model-prompt.ts

@@ -30,6 +30,10 @@ export function convertToLanguageModelPrompt(
       languageModelMessages.push(
         ...prompt.messages.map((message): LanguageModelV1Message => {
           switch (message.role) {
+            case 'system': {
+              return { role: 'system', content: message.content };
+            }
+
             case 'user': {
               if (typeof message.content === 'string') {
                 return {

packages/core/core/prompt/message.ts

@@ -10,10 +10,20 @@ A message that can be used in the `messages` field of a prompt.
 It can be a user message, an assistant message, or a tool message.
  */
 export type CoreMessage =
+  | CoreSystemMessage
   | CoreUserMessage
   | CoreAssistantMessage
   | CoreToolMessage;
 
+/**
+ A system message. It can contain system information.
+
+ Note: using the "system" part of the prompt is strongly preferred
+ to increase the resilience against prompt injection attacks,
+ and because not all providers support several system messages.
+ */
+export type CoreSystemMessage = { role: 'system'; content: string };
+
 /**
  * @deprecated Use `CoreMessage` instead.
  */