feat(WHE-23): add hybrid purchase parser with persisted parse metadata

apps/bot/package.json

@@ -10,6 +10,7 @@
     "lint": "oxlint \"src\""
   },
   "dependencies": {
+    "@household/application": "workspace:*",
     "@household/db": "workspace:*",
     "drizzle-orm": "^0.44.7",
     "grammy": "1.41.1"

apps/bot/src/config.ts

@@ -8,6 +8,8 @@ export interface BotRuntimeConfig {
   telegramHouseholdChatId?: string
   telegramPurchaseTopicId?: number
   purchaseTopicIngestionEnabled: boolean
+  openaiApiKey?: string
+  parserModel: string
 }
 
 function parsePort(raw: string | undefined): number {
@@ -66,7 +68,8 @@ export function getBotRuntimeConfig(env: NodeJS.ProcessEnv = process.env): BotRu
     telegramBotToken: requireValue(env.TELEGRAM_BOT_TOKEN, 'TELEGRAM_BOT_TOKEN'),
     telegramWebhookSecret: requireValue(env.TELEGRAM_WEBHOOK_SECRET, 'TELEGRAM_WEBHOOK_SECRET'),
     telegramWebhookPath: env.TELEGRAM_WEBHOOK_PATH ?? '/webhook/telegram',
-    purchaseTopicIngestionEnabled
+    purchaseTopicIngestionEnabled,
+    parserModel: env.PARSER_MODEL?.trim() || 'gpt-4.1-mini'
   }
 
   if (databaseUrl !== undefined) {
@@ -81,6 +84,10 @@ export function getBotRuntimeConfig(env: NodeJS.ProcessEnv = process.env): BotRu
   if (telegramPurchaseTopicId !== undefined) {
     runtime.telegramPurchaseTopicId = telegramPurchaseTopicId
   }
+  const openaiApiKey = parseOptionalValue(env.OPENAI_API_KEY)
+  if (openaiApiKey !== undefined) {
+    runtime.openaiApiKey = openaiApiKey
+  }
 
   return runtime
 }

apps/bot/src/index.ts

@@ -2,6 +2,7 @@ import { webhookCallback } from 'grammy'
 
 import { createTelegramBot } from './bot'
 import { getBotRuntimeConfig } from './config'
+import { createOpenAiParserFallback } from './openai-parser-fallback'
 import {
   createPurchaseMessageRepository,
   registerPurchaseTopicIngestion
@@ -17,6 +18,7 @@ let closePurchaseRepository: (() => Promise<void>) | undefined
 if (runtime.purchaseTopicIngestionEnabled) {
   const purchaseRepositoryClient = createPurchaseMessageRepository(runtime.databaseUrl!)
   closePurchaseRepository = purchaseRepositoryClient.close
+  const llmFallback = createOpenAiParserFallback(runtime.openaiApiKey, runtime.parserModel)
 
   registerPurchaseTopicIngestion(
     bot,
@@ -25,7 +27,12 @@ if (runtime.purchaseTopicIngestionEnabled) {
       householdChatId: runtime.telegramHouseholdChatId!,
       purchaseTopicId: runtime.telegramPurchaseTopicId!
     },
-    purchaseRepositoryClient.repository
+    purchaseRepositoryClient.repository,
+    llmFallback
+      ? {
+          llmFallback
+        }
+      : {}
   )
 } else {
   console.warn(

apps/bot/src/openai-parser-fallback.ts

@@ -0,0 +1,119 @@
+import type { PurchaseParserLlmFallback } from '@household/application'
+
+interface OpenAiStructuredResult {
+  amountMinor: string
+  currency: 'GEL' | 'USD'
+  itemDescription: string
+  confidence: number
+  needsReview: boolean
+}
+
+function asBigInt(value: string): bigint | null {
+  if (!/^[0-9]+$/.test(value)) {
+    return null
+  }
+
+  const parsed = BigInt(value)
+  return parsed > 0n ? parsed : null
+}
+
+export function createOpenAiParserFallback(
+  apiKey: string | undefined,
+  model: string
+): PurchaseParserLlmFallback | undefined {
+  if (!apiKey) {
+    return undefined
+  }
+
+  return async (rawText: string) => {
+    const response = await fetch('https://api.openai.com/v1/responses', {
+      method: 'POST',
+      headers: {
+        authorization: `Bearer ${apiKey}`,
+        'content-type': 'application/json'
+      },
+      body: JSON.stringify({
+        model,
+        input: [
+          {
+            role: 'system',
+            content:
+              'Extract a shared household purchase from text. Return only valid JSON with amountMinor, currency, itemDescription, confidence, needsReview.'
+          },
+          {
+            role: 'user',
+            content: rawText
+          }
+        ],
+        text: {
+          format: {
+            type: 'json_schema',
+            name: 'purchase_parse',
+            schema: {
+              type: 'object',
+              additionalProperties: false,
+              properties: {
+                amountMinor: {
+                  type: 'string'
+                },
+                currency: {
+                  type: 'string',
+                  enum: ['GEL', 'USD']
+                },
+                itemDescription: {
+                  type: 'string'
+                },
+                confidence: {
+                  type: 'number',
+                  minimum: 0,
+                  maximum: 100
+                },
+                needsReview: {
+                  type: 'boolean'
+                }
+              },
+              required: ['amountMinor', 'currency', 'itemDescription', 'confidence', 'needsReview']
+            }
+          }
+        }
+      })
+    })
+
+    if (!response.ok) {
+      return null
+    }
+
+    const payload = (await response.json()) as {
+      output_text?: string
+    }
+
+    if (!payload.output_text) {
+      return null
+    }
+
+    let parsedJson: OpenAiStructuredResult
+    try {
+      parsedJson = JSON.parse(payload.output_text) as OpenAiStructuredResult
+    } catch {
+      return null
+    }
+
+    const amountMinor = asBigInt(parsedJson.amountMinor)
+    if (!amountMinor) {
+      return null
+    }
+
+    if (parsedJson.itemDescription.trim().length === 0) {
+      return null
+    }
+
+    return {
+      amountMinor,
+      currency: parsedJson.currency,
+      itemDescription: parsedJson.itemDescription,
+      confidence: Math.max(0, Math.min(100, Math.round(parsedJson.confidence))),
+      parserMode: 'llm',
+      needsReview: parsedJson.needsReview
+    }
+  }
+}

apps/bot/src/purchase-topic-ingestion.ts

@@ -1,3 +1,4 @@
+import { parsePurchaseMessage, type PurchaseParserLlmFallback } from '@household/application'
 import { and, eq } from 'drizzle-orm'
 import type { Bot, Context } from 'grammy'
 
@@ -25,7 +26,10 @@ export interface PurchaseTopicRecord extends PurchaseTopicCandidate {
 }
 
 export interface PurchaseMessageIngestionRepository {
-  save(record: PurchaseTopicRecord): Promise<'created' | 'duplicate'>
+  save(
+    record: PurchaseTopicRecord,
+    llmFallback?: PurchaseParserLlmFallback
+  ): Promise<'created' | 'duplicate'>
 }
 
 export function extractPurchaseTopicCandidate(
@@ -52,6 +56,10 @@ export function extractPurchaseTopicCandidate(
   }
 }
 
+function needsReviewAsInt(value: boolean): number {
+  return value ? 1 : 0
+}
+
 export function createPurchaseMessageRepository(databaseUrl: string): {
   repository: PurchaseMessageIngestionRepository
   close: () => Promise<void>
@@ -62,7 +70,7 @@ export function createPurchaseMessageRepository(databaseUrl: string): {
   })
 
   const repository: PurchaseMessageIngestionRepository = {
-    async save(record) {
+    async save(record, llmFallback) {
       const matchedMember = await db
         .select({ id: schema.members.id })
         .from(schema.members)
@@ -75,6 +83,30 @@ export function createPurchaseMessageRepository(databaseUrl: string): {
         .limit(1)
 
       const senderMemberId = matchedMember[0]?.id ?? null
+      let parserError: string | null = null
+
+      const parsed = await parsePurchaseMessage(
+        {
+          rawText: record.rawText
+        },
+        llmFallback
+          ? {
+              llmFallback
+            }
+          : {}
+      ).catch((error) => {
+        parserError = error instanceof Error ? error.message : 'Unknown parser error'
+        return null
+      })
+
+      const processingStatus =
+        parserError !== null
+          ? 'parse_failed'
+          : parsed === null
+            ? 'needs_review'
+            : parsed.needsReview
+              ? 'needs_review'
+              : 'parsed'
 
       const inserted = await db
         .insert(schema.purchaseMessages)
@@ -89,7 +121,14 @@ export function createPurchaseMessageRepository(databaseUrl: string): {
           telegramThreadId: record.threadId,
           telegramUpdateId: String(record.updateId),
           messageSentAt: record.messageSentAt,
-          processingStatus: 'pending'
+          parsedAmountMinor: parsed?.amountMinor,
+          parsedCurrency: parsed?.currency,
+          parsedItemDescription: parsed?.itemDescription,
+          parserMode: parsed?.parserMode,
+          parserConfidence: parsed?.confidence,
+          needsReview: needsReviewAsInt(parsed?.needsReview ?? true),
+          parserError,
+          processingStatus
         })
         .onConflictDoNothing({
           target: [
@@ -151,7 +190,10 @@ function toCandidateFromContext(ctx: Context): PurchaseTopicCandidate | null {
 export function registerPurchaseTopicIngestion(
   bot: Bot,
   config: PurchaseTopicIngestionConfig,
-  repository: PurchaseMessageIngestionRepository
+  repository: PurchaseMessageIngestionRepository,
+  options: {
+    llmFallback?: PurchaseParserLlmFallback
+  } = {}
 ): void {
   bot.on('message:text', async (ctx) => {
     const candidate = toCandidateFromContext(ctx)
@@ -165,7 +207,7 @@ export function registerPurchaseTopicIngestion(
     }
 
     try {
-      const status = await repository.save(record)
+      const status = await repository.save(record, options.llmFallback)
 
       if (status === 'created') {
         console.log(