Allow multiple messages (context window) to be passed

__tests__/schema/ai.ts

@@ -33,16 +33,20 @@ const user = { ...baseUser, roles: ['de_reviewer'] }
 
 const query = new Client({ userId: user.id }).prepareQuery({
   query: gql`
-    query ($prompt: String!) {
+    query ($messages: [ChatCompletionMessageParam!]!) {
       ai {
-        executePrompt(prompt: $prompt) {
+        executePrompt(messages: $messages) {
           success
           record
         }
       }
     }
   `,
-  variables: { prompt: 'Generate exercise for 7th grade math in json' },
+  variables: {
+    messages: [
+      { role: 'user', content: 'Generate exercise for 7th grade math in json' },
+    ],
+  },
 })
 
 beforeAll(() => {

packages/server/src/model/ai.ts

@@ -22,12 +22,16 @@ export async function executePrompt(args: {
   // OpenAI, we can pass the user to the model. See
   // https://platform.openai.com/docs/guides/safety-best-practices/end-user-ids
   userId: number | null
-  prompt: string
+  messages: OpenAI.Chat.Completions.ChatCompletionMessageParam[]
 }): Promise<Record<string, unknown>> {
-  const { userId, prompt } = args
+  const { userId, messages } = args
 
-  if (!prompt || prompt.trim() === '') {
-    throw new UserInputError('Missing prompt parameter')
+  if (
+    !messages[0].content ||
+    (typeof messages[0].content === 'string' &&
+      messages[0].content.trim() === '')
+  ) {
+    throw new UserInputError('Missing prompt within message')
   }
 
   try {
@@ -40,28 +44,23 @@ export async function executePrompt(args: {
 
     const response = await openai.chat.completions.create({
       model: 'gpt-4-1106-preview',
-      messages: [
-        {
-          role: 'user',
-          content: prompt,
-        },
-      ],
+      messages,
       temperature: 0.4,
       user: String(userId),
       response_format: { type: 'json_object' },
     })
 
-    // As we now have the response_format defined as json_object, we shouldn't
-    // need to call JSON.parse on the stringMessage. However, right now the OpenAI
-    // types seem to be broken (thinking the API is returning a string or null).
-    // Instead of fighting the types, we can simply adjust this in the next
-    // version.
     const stringMessage = response.choices[0].message.content
 
     if (!stringMessage) {
       throw new Error('No content received from LLM!')
     }
 
+    // As we now have the response_format defined as json_object, we shouldn't
+    // need to call JSON.parse on the stringMessage. However, right now the OpenAI
+    // types seem to be broken (thinking the API is returning a string or null).
+    // Instead of fighting the types, we can simply adjust this in the next
+    // version.
     const message = JSON.parse(stringMessage) as unknown
 
     if (!t.UnknownRecord.is(message)) {

packages/server/src/schema/ai/resolvers.ts

@@ -1,13 +1,28 @@
 import * as auth from '@serlo/authorization'
 import { Scope } from '@serlo/authorization'
+import { either as E } from 'fp-ts'
+import * as t from 'io-ts'
 
+import { UserInputError } from '~/errors'
 import {
   assertUserIsAuthenticated,
   assertUserIsAuthorized,
   createNamespace,
   Queries,
 } from '~/internals/graphql'
 
+const ChatCompletionMessageParamType = t.type({
+  // Restricts role to 'user' or 'system'. Right now, we don't want to allow
+  // assistant-, tool-, or function calls. See
+  // https://github.com/openai/openai-node/blob/a048174c0e53269a01993a573a10f96c4c9ec79e/src/resources/chat/completions.ts#L405
+  role: t.keyof({ user: null, system: null }),
+  content: t.string,
+})
+
+const ExecutePromptRequestType = t.type({
+  messages: t.array(ChatCompletionMessageParamType),
+})
+
 export const resolvers: Queries<'ai'> = {
   Query: {
     ai: createNamespace(),
@@ -21,9 +36,18 @@ export const resolvers: Queries<'ai'> = {
         message: 'Insufficient role to execute the prompt.',
         dataSources,
       })
+      const { messages } = payload
+
+      const validationResult = ExecutePromptRequestType.decode({ messages })
+      if (E.isLeft(validationResult)) {
+        throw new UserInputError(
+          'Must contain exclusively user or system messages',
+        )
+      }
 
       const record = await dataSources.model.serlo.executePrompt({
         ...payload,
+        messages: validationResult.right.messages,
         userId,
       })
 

packages/server/src/schema/ai/types.graphql

@@ -1,9 +1,16 @@
+input ChatCompletionMessageParam {
+  role: String!
+  content: String!
+}
+
 extend type Query {
   ai: AiQuery!
 }
 
 type AiQuery {
-  executePrompt(prompt: String!): ExecutePromptResponse!
+  executePrompt(
+    messages: [ChatCompletionMessageParam!]!
+  ): ExecutePromptResponse!
 }
 
 type ExecutePromptResponse {

packages/server/src/types.ts

@@ -276,7 +276,7 @@ export type AiQuery = {
 
 
 export type AiQueryExecutePromptArgs = {
-  prompt: Scalars['String']['input'];
+  messages: Array<ChatCompletionMessageParam>;
 };
 
 export type AliasInput = {
@@ -519,6 +519,11 @@ export type CacheUpdateResponse = {
   success: Scalars['Boolean']['output'];
 };
 
+export type ChatCompletionMessageParam = {
+  content: Scalars['String']['input'];
+  role: Scalars['String']['input'];
+};
+
 export type CheckoutRevisionInput = {
   reason: Scalars['String']['input'];
   revisionId: Scalars['Int']['input'];
@@ -2927,6 +2932,7 @@ export type ResolversTypes = {
   CacheSetResponse: ResolverTypeWrapper<ModelOf<CacheSetResponse>>;
   CacheUpdateInput: ResolverTypeWrapper<ModelOf<CacheUpdateInput>>;
   CacheUpdateResponse: ResolverTypeWrapper<ModelOf<CacheUpdateResponse>>;
+  ChatCompletionMessageParam: ResolverTypeWrapper<ModelOf<ChatCompletionMessageParam>>;
   CheckoutRevisionInput: ResolverTypeWrapper<ModelOf<CheckoutRevisionInput>>;
   CheckoutRevisionNotificationEvent: ResolverTypeWrapper<ModelOf<CheckoutRevisionNotificationEvent>>;
   CheckoutRevisionResponse: ResolverTypeWrapper<ModelOf<CheckoutRevisionResponse>>;
@@ -3132,6 +3138,7 @@ export type ResolversParentTypes = {
   CacheSetResponse: ModelOf<CacheSetResponse>;
   CacheUpdateInput: ModelOf<CacheUpdateInput>;
   CacheUpdateResponse: ModelOf<CacheUpdateResponse>;
+  ChatCompletionMessageParam: ModelOf<ChatCompletionMessageParam>;
   CheckoutRevisionInput: ModelOf<CheckoutRevisionInput>;
   CheckoutRevisionNotificationEvent: ModelOf<CheckoutRevisionNotificationEvent>;
   CheckoutRevisionResponse: ModelOf<CheckoutRevisionResponse>;
@@ -3447,7 +3454,7 @@ export type AddRevisionResponseResolvers<ContextType = Context, ParentType exten
 };
 
 export type AiQueryResolvers<ContextType = Context, ParentType extends ResolversParentTypes['AiQuery'] = ResolversParentTypes['AiQuery']> = {
-  executePrompt?: Resolver<ResolversTypes['ExecutePromptResponse'], ParentType, ContextType, RequireFields<AiQueryExecutePromptArgs, 'prompt'>>;
+  executePrompt?: Resolver<ResolversTypes['ExecutePromptResponse'], ParentType, ContextType, RequireFields<AiQueryExecutePromptArgs, 'messages'>>;
   __isTypeOf?: IsTypeOfResolverFn<ParentType, ContextType>;
 };
 

packages/types/src/index.ts

@@ -272,7 +272,7 @@ export type AiQuery = {
 
 
 export type AiQueryExecutePromptArgs = {
-  prompt: Scalars['String']['input'];
+  messages: Array<ChatCompletionMessageParam>;
 };
 
 export type AliasInput = {
@@ -515,6 +515,11 @@ export type CacheUpdateResponse = {
   success: Scalars['Boolean']['output'];
 };
 
+export type ChatCompletionMessageParam = {
+  content: Scalars['String']['input'];
+  role: Scalars['String']['input'];
+};
+
 export type CheckoutRevisionInput = {
   reason: Scalars['String']['input'];
   revisionId: Scalars['Int']['input'];