fix: moderation persist scores [AI-696]

packages/aila/src/features/moderation/AilaModeration.ts

@@ -71,6 +71,7 @@ export class AilaModeration implements AilaModerationFeature {
       appSessionId: chatId,
       messageId: lastAssistantMessage.id,
       categories: moderationResult.categories,
+      scores: moderationResult.scores,
       justification: moderationResult.justification,
       lesson: lessonPlan,
     });
@@ -195,10 +196,11 @@ export class AilaModeration implements AilaModerationFeature {
     messages: Message[];
     lessonPlan: LooseLessonPlan;
     retries: number;
-  }) {
+  }): Promise<ModerationResult> {
     if (retries < 1) {
       return {
         categories: [],
+        scores: undefined,
         justification: "Failed to parse moderation response",
       };
     }

packages/aila/src/features/moderation/moderators/OpenAiModerator.ts

@@ -123,10 +123,7 @@ export class OpenAiModerator extends AilaModerator {
     );
 
     const log = aiLogger("aila:moderation:response");
-    log.info(
-      "Moderation response: ",
-      JSON.stringify(moderationResponse, null, 2),
-    );
+    log.info(JSON.stringify(moderationResponse));
 
     const response = moderationResponseSchema.safeParse(
       JSON.parse(moderationResponse.choices[0]?.message.content ?? "null"),
@@ -152,6 +149,7 @@ export class OpenAiModerator extends AilaModerator {
 
     return {
       justification,
+      scores,
       categories: categories.filter((category) => {
         /**
          * We only want to include the category if the parent category scores below a certain threshold.

packages/core/src/models/moderations.ts

@@ -34,13 +34,15 @@ export class Moderations {
     appSessionId,
     messageId,
     categories,
+    scores,
     justification,
     lesson,
   }: {
     userId: string;
     appSessionId: string;
     messageId: string;
     categories: ModerationResult["categories"];
+    scores: ModerationResult["scores"];
     justification?: string;
     lesson: Snapshot;
   }): Promise<Moderation> {
@@ -58,6 +60,7 @@ export class Moderations {
         userId,
         categories,
         justification,
+        scores,
         appSessionId,
         messageId,
         lessonSnapshotId,

packages/core/src/utils/ailaModeration/moderationSchema.ts

@@ -44,19 +44,21 @@ export const moderationCategoriesSchema = z.array(
 
 const likertScale = z.number().int().min(1).max(5);
 
+const moderationScoresSchema = z.object({
+  l: likertScale.describe("Language and discrimination score"),
+  v: likertScale.describe("Violence and crime score"),
+  u: likertScale.describe("Upsetting, disturbing and sensitive score"),
+  s: likertScale.describe("Nudity and sex score"),
+  p: likertScale.describe("Physical activity and safety score"),
+  t: likertScale.describe("Toxic score"),
+});
+
 /**
  * Schema for the moderation response from the LLM.
  * Note: it's important that 'categories' is the last field in the schema
  */
 export const moderationResponseSchema = z.object({
-  scores: z.object({
-    l: likertScale.describe("Language and discrimination score"),
-    v: likertScale.describe("Violence and crime score"),
-    u: likertScale.describe("Upsetting, disturbing and sensitive score"),
-    s: likertScale.describe("Nudity and sex score"),
-    p: likertScale.describe("Physical activity and safety score"),
-    t: likertScale.describe("Toxic score"),
-  }),
+  scores: moderationScoresSchema,
   justification: z.string().describe("Add justification for your scores."),
   categories: moderationCategoriesSchema,
 });
@@ -65,8 +67,9 @@ export const moderationResponseSchema = z.object({
  * Schema for the moderation result, once parsed from the moderation response
  */
 export const moderationResultSchema = z.object({
-  categories: moderationCategoriesSchema,
   justification: z.string().optional(),
+  scores: moderationScoresSchema.optional(),
+  categories: moderationCategoriesSchema,
 });
 
 export type ModerationResult = z.infer<typeof moderationResultSchema>;

packages/db/prisma/migrations/20241202151820_moderation_scores/migration.sql

@@ -0,0 +1,2 @@
+-- AlterTable
+ALTER TABLE "public"."moderations" ADD COLUMN     "scores" JSONB;

packages/db/prisma/schema.prisma

@@ -789,6 +789,7 @@ model Moderation {
   appSessionId     String    @map("app_session_id")
   messageId        String    @map("message_id")
   categories       Json[]
+  scores           Json?     @map("scores") @db.JsonB
   justification    String?
   lessonSnapshotId String?   @map("lesson_snapshot_id")
   // A user's comment in relation to the moderation. Likely they are contesting it.