fix: sequential AI parsing to prevent free-tier rate limit exhaustion

- CONCURRENT_PARSE: 25 -> 1 (sequential, one email at a time)
- CONCURRENT_FETCH: 20 -> 5 (gentler on Gmail API)
- Added 200ms delay between AI parse calls
- Firing 25 simultaneous Groq/Mistral calls instantly hits RPM limits
and disables all provider slots, causing Max retries exceeded on all emails

packages/server/src/services/scanService.ts

@@ -16,8 +16,9 @@ const DISABLE_SCREENSHOTS = process.env.DISABLE_SCREENSHOTS === 'true';
 const logger = pino({ level: 'info', transport: { target: 'pino-pretty', options: { colorize: true } } });
 
 const BATCH_SIZE = 100;
-const CONCURRENT_FETCH = 20;
-const CONCURRENT_PARSE = 25;
+const CONCURRENT_FETCH = 5;
+const CONCURRENT_PARSE = 1;  // Sequential: free-tier APIs (Groq/Mistral) have strict RPM limits
+const PARSE_DELAY_MS = 200; // ms between AI calls to stay under rate limits
 
 /** Extract a bare email address from a header like "Name <email@domain>" or "email@domain" */
 function extractEmailAddress(header: string): string {
@@ -264,7 +265,7 @@ async function runScan(jobId: string, scanRequest: ScanRequest, userId: string,
         }
       }
 
-      // Phase 3: AI parsing with concurrency limit
+      // Phase 3: AI parsing — sequential with delay to respect free-tier RPM limits
       for (let j = 0; j < emails.length; j += CONCURRENT_PARSE) {
         const parseChunk = emails.slice(j, j + CONCURRENT_PARSE);
         const parseResults = await Promise.allSettled(
@@ -273,6 +274,10 @@ async function runScan(jobId: string, scanRequest: ScanRequest, userId: string,
             return { email, parsed };
           })
         );
+        // Small delay between batches to avoid rate limiting
+        if (j + CONCURRENT_PARSE < emails.length) {
+          await new Promise(resolve => setTimeout(resolve, PARSE_DELAY_MS));
+        }
 
         for (const result of parseResults) {
           if (result.status === 'fulfilled') {