Reduce AI costs: switch tagging to gpt-4o-mini, add custom base URL support

- Change AI tagging to use AI_MODELS.QUICK (gpt-4o-mini) instead of gpt-4o for
  10-15x cost reduction on classification tasks
- Add openai_base_url system setting for OpenAI-compatible providers
  (OpenRouter, Groq, Together AI, local models)
- Reset OpenAI client singleton when API key, base URL, or model changes
- Add base URL field to AI settings form with provider examples

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

src/app/(admin)/admin/projects/[id]/page.tsx

@@ -105,14 +105,13 @@ function ProjectDetailContent({ projectId }: { projectId: string }) {
   // Extract all rounds from the competition
   const competitionRounds = competition?.rounds || []
 
-  // Fetch requirements for each round
-  const requirementQueries = competitionRounds.map((round: { id: string; name: string }) =>
-    trpc.file.listRequirements.useQuery({ roundId: round.id })
+  // Fetch requirements for all rounds in a single query (avoids dynamic hook violation)
+  const roundIds = competitionRounds.map((r: { id: string }) => r.id)
+  const { data: allRequirements = [] } = trpc.file.listRequirementsByRounds.useQuery(
+    { roundIds },
+    { enabled: roundIds.length > 0 }
   )
 
-  // Combine requirements from all rounds
-  const allRequirements = requirementQueries.flatMap((q: { data?: unknown[] }) => q.data || [])
-
   const utils = trpc.useUtils()
 
   if (isLoading) {
@@ -592,7 +591,7 @@ function ProjectDetailContent({ projectId }: { projectId: string }) {
                                   </p>
                                 )}
                                 <div className="flex items-center gap-2 text-xs text-muted-foreground mt-0.5">
-                                  {req.acceptedMimeTypes.length > 0 && (
+                                  {req.acceptedMimeTypes?.length > 0 && (
                                     <span>
                                       {req.acceptedMimeTypes.map((mime: string) => {
                                         if (mime === 'application/pdf') return 'PDF'

src/app/(admin)/admin/projects/page.tsx

@@ -366,8 +366,9 @@ export default function ProjectsPage() {
   }
 
   const handleCloseTaggingDialog = () => {
-    if (!taggingInProgress) {
     setAiTagDialogOpen(false)
+    // Only reset job state if not in progress (preserve polling for background jobs)
+    if (!taggingInProgress) {
       setActiveTaggingJobId(null)
       setSelectedRoundForTagging('')
       setSelectedProgramForTagging('')
@@ -618,9 +619,22 @@ export default function ProjectsPage() {
           </p>
         </div>
         <div className="flex flex-wrap gap-2">
-          <Button variant="outline" onClick={() => setAiTagDialogOpen(true)}>
+          <Button
+            variant="outline"
+            onClick={() => setAiTagDialogOpen(true)}
+            className={taggingInProgress ? 'border-amber-400 bg-amber-50 dark:bg-amber-950/20' : ''}
+          >
+            {taggingInProgress ? (
+              <Loader2 className="mr-2 h-4 w-4 animate-spin text-amber-600" />
+            ) : (
               <Bot className="mr-2 h-4 w-4" />
+            )}
             AI Tags
+            {taggingInProgress && (
+              <span className="ml-1.5 text-[10px] text-amber-600 font-medium">
+                {taggingProgressPercent}%
+              </span>
+            )}
           </Button>
           <Button variant="outline" asChild>
             <Link href="/admin/projects/pool">
@@ -1833,9 +1847,8 @@ export default function ProjectsPage() {
                 <Button
                   variant="outline"
                   onClick={handleCloseTaggingDialog}
-                  disabled={taggingInProgress}
                 >
-                  Cancel
+                  {taggingInProgress ? 'Run in Background' : 'Cancel'}
                 </Button>
                 <Button
                   onClick={handleStartTagging}

src/components/settings/ai-settings-form.tsx

@@ -36,6 +36,7 @@ const formSchema = z.object({
   ai_model: z.string(),
   ai_send_descriptions: z.boolean(),
   openai_api_key: z.string().optional(),
+  openai_base_url: z.string().optional(),
 })
 
 type FormValues = z.infer<typeof formSchema>
@@ -47,6 +48,7 @@ interface AISettingsFormProps {
     ai_model?: string
     ai_send_descriptions?: string
     openai_api_key?: string
+    openai_base_url?: string
   }
 }
 
@@ -61,6 +63,7 @@ export function AISettingsForm({ settings }: AISettingsFormProps) {
       ai_model: settings.ai_model || 'gpt-4o',
       ai_send_descriptions: settings.ai_send_descriptions === 'true',
       openai_api_key: '',
+      openai_base_url: settings.openai_base_url || '',
     },
   })
 
@@ -113,6 +116,9 @@ export function AISettingsForm({ settings }: AISettingsFormProps) {
       settingsToUpdate.push({ key: 'openai_api_key', value: data.openai_api_key })
     }
 
+    // Save base URL (empty string clears it)
+    settingsToUpdate.push({ key: 'openai_base_url', value: data.openai_base_url?.trim() || '' })
+
     updateSettings.mutate({ settings: settingsToUpdate })
   }
 
@@ -208,6 +214,27 @@ export function AISettingsForm({ settings }: AISettingsFormProps) {
           )}
         />
 
+        <FormField
+          control={form.control}
+          name="openai_base_url"
+          render={({ field }) => (
+            <FormItem>
+              <FormLabel>API Base URL (Optional)</FormLabel>
+              <FormControl>
+                <Input
+                  placeholder="https://api.openai.com/v1"
+                  {...field}
+                />
+              </FormControl>
+              <FormDescription>
+                Custom base URL for OpenAI-compatible providers. Leave blank for OpenAI.
+                Use <code className="text-xs bg-muted px-1 rounded">https://openrouter.ai/api/v1</code> for OpenRouter (access Claude, Gemini, Llama, etc.)
+              </FormDescription>
+              <FormMessage />
+            </FormItem>
+          )}
+        />
+
         <FormField
           control={form.control}
           name="ai_model"

src/components/settings/settings-content.tsx

@@ -84,6 +84,7 @@ export function SettingsContent({ initialSettings, isSuperAdmin = true }: Settin
     'ai_model',
     'ai_send_descriptions',
     'openai_api_key',
+    'openai_base_url',
   ])
 
   const brandingSettings = getSettingsByKeys([

src/lib/openai.ts

@@ -187,7 +187,25 @@ async function getOpenAIApiKey(): Promise<string | null> {
 }
 
 /**
- * Create OpenAI client instance
+ * Get custom base URL for OpenAI-compatible providers.
+ * Supports OpenRouter, Together AI, Groq, local models, etc.
+ * Set via Settings → AI or OPENAI_BASE_URL env var.
+ */
+async function getBaseURL(): Promise<string | undefined> {
+  try {
+    const setting = await prisma.systemSettings.findUnique({
+      where: { key: 'openai_base_url' },
+    })
+    return setting?.value || process.env.OPENAI_BASE_URL || undefined
+  } catch {
+    return process.env.OPENAI_BASE_URL || undefined
+  }
+}
+
+/**
+ * Create OpenAI client instance.
+ * Supports custom baseURL for OpenAI-compatible providers
+ * (OpenRouter, Groq, Together AI, local models, etc.)
  */
 async function createOpenAIClient(): Promise<OpenAI | null> {
   const apiKey = await getOpenAIApiKey()
@@ -197,8 +215,15 @@ async function createOpenAIClient(): Promise<OpenAI | null> {
     return null
   }
 
+  const baseURL = await getBaseURL()
+
+  if (baseURL) {
+    console.log(`[OpenAI] Using custom base URL: ${baseURL}`)
+  }
+
   return new OpenAI({
     apiKey,
+    ...(baseURL ? { baseURL } : {}),
   })
 }
 
@@ -221,6 +246,15 @@ export async function getOpenAI(): Promise<OpenAI | null> {
   return client
 }
 
+/**
+ * Reset the OpenAI client singleton (e.g., after settings change).
+ * Next call to getOpenAI() will create a fresh client.
+ */
+export function resetOpenAIClient(): void {
+  globalForOpenAI.openai = undefined
+  globalForOpenAI.openaiInitialized = false
+}
+
 /**
  * Check if OpenAI is configured and available
  */

src/server/routers/file.ts

@@ -818,6 +818,20 @@ export const fileRouter = router({
       })
     }),
 
+  /**
+   * List file requirements for multiple rounds in a single query.
+   * Avoids dynamic hook violations when fetching requirements per-round.
+   */
+  listRequirementsByRounds: protectedProcedure
+    .input(z.object({ roundIds: z.array(z.string()).max(50) }))
+    .query(async ({ ctx, input }) => {
+      if (input.roundIds.length === 0) return []
+      return ctx.prisma.fileRequirement.findMany({
+        where: { roundId: { in: input.roundIds } },
+        orderBy: { sortOrder: 'asc' },
+      })
+    }),
+
   /**
    * Create a file requirement for a stage (admin only)
    */

src/server/routers/settings.ts

@@ -201,6 +201,12 @@ export const settingsRouter = router({
         clearStorageProviderCache()
       }
 
+      // Reset OpenAI client if API key or base URL changed
+      if (input.settings.some((s) => s.key === 'openai_api_key' || s.key === 'openai_base_url' || s.key === 'ai_model')) {
+        const { resetOpenAIClient } = await import('@/lib/openai')
+        resetOpenAIClient()
+      }
+
       // Audit log
       await logAudit({
         prisma: ctx.prisma,

src/server/routers/tag.ts

@@ -5,6 +5,7 @@ import { prisma } from '@/lib/prisma'
 import { logAudit } from '../utils/audit'
 import {
   tagProject,
+  tagProjectsBatch,
   getTagSuggestions,
   addProjectTag,
   removeProjectTag,
@@ -17,7 +18,7 @@ import {
   NotificationTypes,
 } from '../services/in-app-notification'
 
-// Background job runner for tagging
+// Background job runner for tagging — uses batched API calls for efficiency
 async function runTaggingJob(jobId: string, userId: string) {
   const job = await prisma.taggingJob.findUnique({
     where: { id: jobId },
@@ -28,7 +29,7 @@ async function runTaggingJob(jobId: string, userId: string) {
     return
   }
 
-  console.log(`[AI Tagging Job] Starting job ${jobId}...`)
+  console.log(`[AI Tagging Job] Starting job ${jobId} (batched mode)...`)
 
   // Mark as running
   await prisma.taggingJob.update({
@@ -56,7 +57,7 @@ async function runTaggingJob(jobId: string, userId: string) {
 
     const allProjects = await prisma.project.findMany({
       where: whereClause,
-      select: { id: true, title: true, tags: true },
+      select: { id: true, title: true, tags: true, projectTags: { select: { tagId: true } } },
     })
 
     const untaggedProjects = allProjects.filter(p => p.tags.length === 0)
@@ -83,48 +84,33 @@ async function runTaggingJob(jobId: string, userId: string) {
       return
     }
 
-    let taggedCount = 0
-    let failedCount = 0
-    const errors: string[] = []
     const startTime = Date.now()
 
-    for (let i = 0; i < untaggedProjects.length; i++) {
-      const project = untaggedProjects[i]
-      console.log(`[AI Tagging Job] Processing ${i + 1}/${untaggedProjects.length}: "${project.title.substring(0, 40)}..."`)
-
-      try {
-        const result = await tagProject(project.id, userId)
-        taggedCount++
-        console.log(`[AI Tagging Job] ✓ Tagged with ${result.applied.length} tags`)
-      } catch (error) {
-        failedCount++
-        const errorMsg = error instanceof Error ? error.message : 'Unknown error'
-        errors.push(`${project.title}: ${errorMsg}`)
-        console.error(`[AI Tagging Job] ✗ Failed: ${errorMsg}`)
-      }
-
-      // Update progress
+    // Use batched tagging — processes 10 projects per API call, 3 concurrent calls
+    const { results, totalTokens } = await tagProjectsBatch(
+      untaggedProjects,
+      userId,
+      async (processed, total) => {
+        // Update job progress on each batch completion
+        const taggedSoFar = results?.length ?? processed
         await prisma.taggingJob.update({
           where: { id: jobId },
           data: {
-          processedCount: i + 1,
-          taggedCount,
-          failedCount,
-          errorsJson: errors.length > 0 ? errors.slice(0, 20) : undefined, // Keep last 20 errors
+            processedCount: processed,
+            taggedCount: taggedSoFar,
           },
         })
 
-      // Log progress every 10 projects
-      if ((i + 1) % 10 === 0) {
         const elapsed = ((Date.now() - startTime) / 1000).toFixed(0)
-        const avgTime = (Date.now() - startTime) / (i + 1) / 1000
-        const remaining = avgTime * (untaggedProjects.length - i - 1)
-        console.log(`[AI Tagging Job] Progress: ${i + 1}/${untaggedProjects.length} (${elapsed}s elapsed, ~${remaining.toFixed(0)}s remaining)`)
-      }
+        console.log(`[AI Tagging Job] Progress: ${processed}/${total} (${elapsed}s elapsed)`)
       }
+    )
+
+    const taggedCount = results.filter(r => r.applied.length > 0).length
+    const failedCount = untaggedProjects.length - results.length
 
     const totalTime = ((Date.now() - startTime) / 1000).toFixed(1)
-    console.log(`[AI Tagging Job] Complete: ${taggedCount} tagged, ${failedCount} failed in ${totalTime}s`)
+    console.log(`[AI Tagging Job] Complete: ${taggedCount} tagged, ${failedCount} failed in ${totalTime}s (${totalTokens} tokens)`)
 
     // Mark as completed
     await prisma.taggingJob.update({
@@ -132,7 +118,9 @@ async function runTaggingJob(jobId: string, userId: string) {
       data: {
         status: 'COMPLETED',
         completedAt: new Date(),
-        errorsJson: errors.length > 0 ? errors : undefined,
+        processedCount: results.length,
+        taggedCount,
+        failedCount,
       },
     })
 
@@ -144,7 +132,7 @@ async function runTaggingJob(jobId: string, userId: string) {
       linkUrl: '/admin/projects',
       linkLabel: 'View Projects',
       priority: 'normal',
-      metadata: { jobId, taggedCount, failedCount, skippedCount },
+      metadata: { jobId, taggedCount, failedCount, skippedCount, totalTokens },
     })
 
   } catch (error) {

src/server/services/ai-filtering.ts

@@ -142,7 +142,7 @@ interface FilteringRuleInput {
 const DEFAULT_BATCH_SIZE = 20
 const MAX_BATCH_SIZE = 50
 const MIN_BATCH_SIZE = 1
-const DEFAULT_PARALLEL_BATCHES = 1
+const DEFAULT_PARALLEL_BATCHES = 3
 const MAX_PARALLEL_BATCHES = 10
 
 // Structured system prompt for AI screening

src/server/services/ai-shortlist.ts

@@ -344,8 +344,8 @@ export async function generateShortlist(
     let totalTokens = 0
     const allErrors: string[] = []
 
-    // Run each category independently
-    for (const cat of categories) {
+    // Run categories in parallel for efficiency
+    const categoryPromises = categories.map(async (cat) => {
       const catTopN = cat === 'STARTUP'
         ? (startupTopN ?? topN)
         : (conceptTopN ?? topN)
@@ -357,6 +357,12 @@ export async function generateShortlist(
         prisma,
       )
 
+      return { cat, result }
+    })
+
+    const categoryResults = await Promise.all(categoryPromises)
+
+    for (const { cat, result } of categoryResults) {
       if (cat === 'STARTUP') {
         allRecommendations.STARTUP = result.recommendations
       } else {

src/server/services/ai-tagging.ts

@@ -5,7 +5,7 @@
  *
  * Features:
  * - Single project tagging (on-submit or manual)
- * - Batch tagging for rounds
+ * - Batch tagging with concurrent processing (10 projects/batch, 3 concurrent)
  * - Confidence scores for each tag
  * - Additive only - never removes existing tags
  *
@@ -16,7 +16,7 @@
  */
 
 import { prisma } from '@/lib/prisma'
-import { getOpenAI, getConfiguredModel, buildCompletionParams } from '@/lib/openai'
+import { getOpenAI, getConfiguredModel, buildCompletionParams, AI_MODELS } from '@/lib/openai'
 import { logAIUsage, extractTokenUsage } from '@/server/utils/ai-usage'
 import { classifyAIError, createParseError, logAIError } from './ai-errors'
 import {
@@ -53,8 +53,10 @@ interface AvailableTag {
 
 const CONFIDENCE_THRESHOLD = 0.5
 const DEFAULT_MAX_TAGS = 5
+const BATCH_SIZE = 10 // Projects per API call
+const BATCH_CONCURRENCY = 3 // Concurrent API calls
 
-// System prompt optimized for tag suggestion
+// System prompt optimized for single-project tag suggestion
 const TAG_SUGGESTION_SYSTEM_PROMPT = `You are an expert at categorizing ocean conservation and sustainability projects.
 
 Analyze the project and suggest the most relevant expertise tags from the provided list.
@@ -78,6 +80,36 @@ Rules:
 - Maximum 7 suggestions per project
 - Be conservative - only suggest tags that truly apply`
 
+// System prompt optimized for batch tagging (multiple projects in one call)
+const BATCH_TAG_SYSTEM_PROMPT = `You are an expert at categorizing ocean conservation and sustainability projects.
+
+Analyze EACH project and suggest the most relevant expertise tags from the provided list.
+Consider each project's focus areas, technology, methodology, and domain.
+
+Return JSON with this format:
+{
+  "projects": [
+    {
+      "project_id": "PROJECT_001",
+      "suggestions": [
+        {
+          "tag_name": "exact tag name from list",
+          "confidence": 0.0-1.0,
+          "reasoning": "brief explanation"
+        }
+      ]
+    }
+  ]
+}
+
+Rules:
+- Only suggest tags from the provided list (exact names)
+- Order by relevance (most relevant first)
+- Confidence should reflect how well the tag matches
+- Maximum 7 suggestions per project
+- Be conservative - only suggest tags that truly apply
+- Return results for ALL projects provided`
+
 // ─── Helper Functions ────────────────────────────────────────────────────────
 
 /**
@@ -132,7 +164,8 @@ export async function getAvailableTags(): Promise<AvailableTag[]> {
 // ─── AI Tagging Core ─────────────────────────────────────────────────────────
 
 /**
- * Call OpenAI to get tag suggestions for a project
+ * Call OpenAI to get tag suggestions for a single project
+ * Used for on-demand single-project tagging
  */
 async function getAISuggestions(
   anonymizedProject: AnonymizedProjectForAI,
@@ -145,9 +178,10 @@ async function getAISuggestions(
     return { suggestions: [], tokensUsed: 0 }
   }
 
-  const model = await getConfiguredModel()
+  // Use QUICK model — tag classification is simple, doesn't need expensive reasoning
+  const model = await getConfiguredModel(AI_MODELS.QUICK)
 
-  // Build tag list for prompt
+  // Build compact tag list for prompt
   const tagList = availableTags.map((t) => ({
     name: t.name,
     category: t.category,
@@ -155,10 +189,10 @@ async function getAISuggestions(
   }))
 
   const userPrompt = `PROJECT:
-${JSON.stringify(anonymizedProject, null, 2)}
+${JSON.stringify(anonymizedProject)}
 
 AVAILABLE TAGS:
-${JSON.stringify(tagList, null, 2)}
+${JSON.stringify(tagList)}
 
 Suggest relevant tags for this project.`
 
@@ -246,6 +280,161 @@ Suggest relevant tags for this project.`
   }
 }
 
+/**
+ * Call OpenAI to get tag suggestions for a batch of projects in one API call.
+ * Returns a map of project_id -> TagSuggestion[].
+ */
+async function getAISuggestionsBatch(
+  anonymizedProjects: AnonymizedProjectForAI[],
+  availableTags: AvailableTag[],
+  userId?: string
+): Promise<{ suggestionsMap: Map<string, TagSuggestion[]>; tokensUsed: number }> {
+  const openai = await getOpenAI()
+  if (!openai) {
+    console.warn('[AI Tagging] OpenAI not configured')
+    return { suggestionsMap: new Map(), tokensUsed: 0 }
+  }
+
+  // Use QUICK model — tag classification is simple, doesn't need expensive reasoning
+  const model = await getConfiguredModel(AI_MODELS.QUICK)
+  const suggestionsMap = new Map<string, TagSuggestion[]>()
+
+  // Build compact tag list (sent once for entire batch)
+  const tagList = availableTags.map((t) => ({
+    name: t.name,
+    category: t.category,
+    description: t.description,
+  }))
+
+  const userPrompt = `PROJECTS (${anonymizedProjects.length}):
+${JSON.stringify(anonymizedProjects)}
+
+AVAILABLE TAGS:
+${JSON.stringify(tagList)}
+
+Suggest relevant tags for each project.`
+
+  const MAX_PARSE_RETRIES = 2
+  let parseAttempts = 0
+
+  try {
+    const params = buildCompletionParams(model, {
+      messages: [
+        { role: 'system', content: BATCH_TAG_SYSTEM_PROMPT },
+        { role: 'user', content: userPrompt },
+      ],
+      jsonMode: true,
+      temperature: 0.1,
+      maxTokens: Math.min(4000, anonymizedProjects.length * 500),
+    })
+
+    let response = await openai.chat.completions.create(params)
+    let usage = extractTokenUsage(response)
+    let totalTokens = usage.totalTokens
+
+    // Parse with retry logic
+    let parsed: {
+      projects: Array<{
+        project_id: string
+        suggestions: Array<{
+          tag_name: string
+          confidence: number
+          reasoning: string
+        }>
+      }>
+    }
+
+    while (true) {
+      try {
+        const content = response.choices[0]?.message?.content
+        if (!content) throw new Error('Empty response from AI')
+
+        const raw = JSON.parse(content)
+        parsed = raw.projects ? raw : { projects: Array.isArray(raw) ? raw : [] }
+        break
+      } catch (parseError) {
+        if (parseError instanceof SyntaxError && parseAttempts < MAX_PARSE_RETRIES) {
+          parseAttempts++
+          console.warn(`[AI Tagging Batch] JSON parse failed, retrying (${parseAttempts}/${MAX_PARSE_RETRIES})`)
+          const retryParams = buildCompletionParams(model, {
+            messages: [
+              { role: 'system', content: BATCH_TAG_SYSTEM_PROMPT },
+              { role: 'user', content: userPrompt + '\n\nIMPORTANT: Please ensure valid JSON output.' },
+            ],
+            jsonMode: true,
+            temperature: 0.1,
+            maxTokens: Math.min(4000, anonymizedProjects.length * 500),
+          })
+          response = await openai.chat.completions.create(retryParams)
+          const retryUsage = extractTokenUsage(response)
+          totalTokens += retryUsage.totalTokens
+          continue
+        }
+        throw parseError
+      }
+    }
+
+    // Log usage for the entire batch
+    await logAIUsage({
+      userId,
+      action: 'PROJECT_TAGGING',
+      entityType: 'Project',
+      model,
+      promptTokens: usage.promptTokens,
+      completionTokens: usage.completionTokens,
+      totalTokens,
+      batchSize: anonymizedProjects.length,
+      itemsProcessed: parsed.projects?.length || 0,
+      status: 'SUCCESS',
+    })
+
+    // Map results back to TagSuggestion format
+    for (const projectResult of parsed.projects || []) {
+      const suggestions: TagSuggestion[] = []
+      for (const s of projectResult.suggestions || []) {
+        const tag = availableTags.find(
+          (t) => t.name.toLowerCase() === s.tag_name.toLowerCase()
+        )
+        if (tag) {
+          suggestions.push({
+            tagId: tag.id,
+            tagName: tag.name,
+            confidence: Math.max(0, Math.min(1, s.confidence)),
+            reasoning: s.reasoning || '',
+          })
+        }
+      }
+      suggestionsMap.set(projectResult.project_id, suggestions)
+    }
+
+    return { suggestionsMap, tokensUsed: totalTokens }
+  } catch (error) {
+    if (error instanceof SyntaxError) {
+      const parseError = createParseError(error.message)
+      logAIError('Tagging', 'getAISuggestionsBatch', parseError)
+    }
+
+    const classified = classifyAIError(error)
+    logAIError('Tagging', 'getAISuggestionsBatch', classified)
+
+    await logAIUsage({
+      userId,
+      action: 'PROJECT_TAGGING',
+      entityType: 'Project',
+      model: 'unknown',
+      promptTokens: 0,
+      completionTokens: 0,
+      totalTokens: 0,
+      batchSize: anonymizedProjects.length,
+      itemsProcessed: 0,
+      status: 'ERROR',
+      errorMessage: error instanceof Error ? error.message : 'Unknown error',
+    })
+
+    throw error
+  }
+}
+
 // ─── Public API ──────────────────────────────────────────────────────────────
 
 /**
@@ -355,6 +544,153 @@ export async function tagProject(
   }
 }
 
+/**
+ * Tag a batch of projects using batched API calls with concurrency.
+ * Much more efficient than tagging one-by-one for bulk operations.
+ *
+ * @param projects Array of { id, projectTags } to tag
+ * @param userId The user initiating the tagging
+ * @param onProgress Callback for progress updates
+ * @returns Array of TaggingResult
+ */
+export async function tagProjectsBatch(
+  projects: Array<{
+    id: string
+    title: string
+    projectTags: Array<{ tagId: string }>
+  }>,
+  userId: string,
+  onProgress?: (processed: number, total: number) => Promise<void>
+): Promise<{ results: TaggingResult[]; totalTokens: number }> {
+  const settings = await getTaggingSettings()
+  if (!settings.enabled) {
+    return { results: [], totalTokens: 0 }
+  }
+
+  const availableTags = await getAvailableTags()
+  if (availableTags.length === 0) {
+    return { results: [], totalTokens: 0 }
+  }
+
+  // Fetch full project data for all projects at once (single DB query)
+  const fullProjects = await prisma.project.findMany({
+    where: { id: { in: projects.map((p) => p.id) } },
+    include: {
+      projectTags: true,
+      files: { select: { fileType: true } },
+      _count: { select: { teamMembers: true, files: true } },
+    },
+  })
+
+  const projectMap = new Map(fullProjects.map((p) => [p.id, p]))
+
+  // Anonymize all projects at once
+  const projectsWithRelations = fullProjects.map(toProjectWithRelations)
+  const { anonymized, mappings } = anonymizeProjectsForAI(projectsWithRelations, 'FILTERING')
+
+  if (!validateAnonymizedProjects(anonymized)) {
+    throw new Error('GDPR compliance check failed: PII detected in anonymized data')
+  }
+
+  // Build mapping from anonymous ID to real project
+  const anonToRealMap = new Map<string, string>()
+  for (const mapping of mappings) {
+    anonToRealMap.set(mapping.anonymousId, mapping.realId)
+  }
+
+  // Split into batches
+  const batches: AnonymizedProjectForAI[][] = []
+  for (let i = 0; i < anonymized.length; i += BATCH_SIZE) {
+    batches.push(anonymized.slice(i, i + BATCH_SIZE))
+  }
+
+  const allResults: TaggingResult[] = []
+  let totalTokens = 0
+  let processedCount = 0
+
+  // Process batches with concurrency
+  for (let i = 0; i < batches.length; i += BATCH_CONCURRENCY) {
+    const concurrentBatches = batches.slice(i, i + BATCH_CONCURRENCY)
+
+    const batchPromises = concurrentBatches.map(async (batch) => {
+      try {
+        const { suggestionsMap, tokensUsed } = await getAISuggestionsBatch(
+          batch,
+          availableTags,
+          userId
+        )
+        return { suggestionsMap, tokensUsed, error: null }
+      } catch (error) {
+        console.error('[AI Tagging Batch] Batch failed:', error)
+        return { suggestionsMap: new Map<string, TagSuggestion[]>(), tokensUsed: 0, error }
+      }
+    })
+
+    const batchResults = await Promise.all(batchPromises)
+
+    // Process results from all concurrent batches
+    for (const { suggestionsMap, tokensUsed } of batchResults) {
+      totalTokens += tokensUsed
+
+      for (const [anonId, suggestions] of suggestionsMap) {
+        const realId = anonToRealMap.get(anonId)
+        if (!realId) continue
+
+        const project = projectMap.get(realId)
+        if (!project) continue
+
+        // Filter by confidence
+        const validSuggestions = suggestions.filter(
+          (s) => s.confidence >= settings.confidenceThreshold
+        )
+
+        // Get existing tags
+        const existingTagIds = new Set(project.projectTags.map((pt) => pt.tagId))
+        const currentTagCount = project.projectTags.length
+        const remainingSlots = Math.max(0, settings.maxTags - currentTagCount)
+
+        const newSuggestions = validSuggestions
+          .filter((s) => !existingTagIds.has(s.tagId))
+          .slice(0, remainingSlots)
+
+        // Apply tags
+        const applied: TagSuggestion[] = []
+        for (const suggestion of newSuggestions) {
+          try {
+            await prisma.projectTag.create({
+              data: {
+                projectId: realId,
+                tagId: suggestion.tagId,
+                confidence: suggestion.confidence,
+                source: 'AI',
+              },
+            })
+            applied.push(suggestion)
+          } catch {
+            // Skip duplicates
+          }
+        }
+
+        allResults.push({
+          projectId: realId,
+          suggestions,
+          applied,
+          tokensUsed: 0, // Token tracking is per-batch, not per-project
+        })
+
+        processedCount++
+      }
+    }
+
+    // Report progress after each concurrent chunk
+    if (onProgress) {
+      await onProgress(processedCount, projects.length)
+    }
+  }
+
+  return { results: allResults, totalTokens }
+}
+
 /**
  * Get tag suggestions for a project without applying them
  * Useful for preview/review before applying