livedash-node/refactor/ai-session-processing-pipeline.md

Refactor AI Session Processing Pipeline

This is a significant but valuable refactoring project. A detailed, well-structured prompt is key for getting a good result from a code-focused AI like Claude. Project: LiveDash-Node (~/Projects/livedash-node-max-branch) Objective: Refactor our AI session processing pipeline to use the OpenAI Batch API for cost savings and higher throughput. Implement a new internal admin API under /api/admin/legacy/* to monitor and manage this new asynchronous workflow. Assignee: Claude Code

Context

Our current system processes AI analysis requests (like sentiment analysis, summarization, etc.) in a synchronous or simple asynchronous loop, likely via processingScheduler.ts. This is inefficient and costly at scale. We are moving to OpenAI's Batch API, which is fully asynchronous and requires a stateful, multi-stage processing architecture.

The term "legacy" in the API path /api/admin/legacy/* is intentional. It refers to the fact that our current method of getting data (CSV imports) is the "legacy" workflow. In the future, we plan to introduce a new API for clients to submit session data directly. This admin API is for monitoring the processing of data from our legacy import system.

Please follow the phased plan below precisely.

---

Phase 1: Database Schema Changes (prisma/schema.prisma)

First, we need to update our database schema to track the state of batch jobs and the individual requests within them.

1. Add the AIBatchRequest model and AIBatchRequestStatus enum. This table will track the status of each batch job submitted to OpenAI.

   // Add this new model to your schema.prisma
   model AIBatchRequest {
       id              String   @id @default(cuid())
       companyId       String
       company         Company  @relation(fields: [companyId], references: [id])
   
       // OpenAI specific IDs
       openaiBatchId   String   @unique
       inputFileId     String
       outputFileId    String?
       errorFileId     String?
   
       // Our internal status tracking
       status          AIBatchRequestStatus @default(PENDING)
   
       // Timestamps
       createdAt       DateTime @default(now())
       completedAt     DateTime?
       processedAt     DateTime? // When we finished processing the results
   
       // Relation to the individual requests included in this batch
       processingRequests AIProcessingRequest[]
       @@index([companyId, status])
   }
   
   // prettier-ignore
   enum AIBatchRequestStatus {
       PENDING           // We have created the batch in our DB, preparing to send to OpenAI
       UPLOADING         // Uploading the .jsonl file
       VALIDATING        // OpenAI is validating the file
       IN_PROGRESS       // OpenAI is processing the batch
       FINALIZING        // OpenAI is finalizing the results
       COMPLETED         // OpenAI job is done, results are available for download
       PROCESSED         // We have successfully downloaded and processed all results
       FAILED            // The batch failed validation or expired
       CANCELLED         // The batch was cancelled
   }
   

2. Update the AIProcessingRequest model and add the AIRequestStatus enum. We need to track the state of each individual request as it moves through the batching pipeline.

   // In your existing AIProcessingRequest model, add the new fields and enum.
   
   model AIProcessingRequest {
       // ... all existing fields (id, sessionId, token counts, etc.)
       // === ADD THESE NEW FIELDS ===
       processingStatus  AIRequestStatus @default(PENDING_BATCHING)
       batchId           String?
       batch             AIBatchRequest? @relation(fields: [batchId], references: [id])
       // ============================
       @@index([processingStatus]) // Add this index for efficient querying
   }
   
   // prettier-ignore
   enum AIRequestStatus {
       PENDING_BATCHING      // Default state: waiting to be picked up by the batch creator
       BATCHING_IN_PROGRESS  // It has been assigned to a batch that is currently running
       PROCESSING_COMPLETE   // The batch finished and we successfully got a result for this request
       PROCESSING_FAILED     // The batch finished but this specific request failed
   }
   

After modifying the schema, please run pnpm prisma:generate.

---

Phase 2: Implement the Batch Processing Schedulers

The core of this refactor is to replace the existing logic in lib/processingScheduler.ts with a two-stage scheduler system. You can create new files for this logic (e.g., lib/batchCreator.ts, lib/batchPoller.ts) and integrate them into lib/schedulers.ts.

Scheduler 1: Batch Creation (lib/batchCreator.ts)

This scheduler runs periodically (e.g., every 10 minutes) to bundle pending requests into a batch.

Functionality:

1. Query the database for AIProcessingRequest records with processingStatus: PENDING_BATCHING.
2. Group these requests by the AI model they need to use (e.g., gpt-4-turbo). The Batch API requires one model per batch file.
3. For each model group:
   1. Generate a .jsonl string. Each line must be a valid OpenAI batch request.
   2. Crucially, use our internal AIProcessingRequest.id as the custom_id in each JSON line. This is how we will map results back.
   3. Upload the .jsonl content to OpenAI using openai.files.create({ file: Buffer.from(jsonlContent), purpose: 'batch' }).
   4. Create the batch job using openai.batches.create() with the returned input_file_id.
   5. In a single database transaction:
      1. Create a new AIBatchRequest record in our database, storing the openaiBatchId, inputFileId, and setting the initial status to VALIDATING.
      2. Update all the AIProcessingRequest records included in this batch to set their processingStatus to BATCHING_IN_PROGRESS and link them via the batchId.

Scheduler 2: Result Polling (lib/batchPoller.ts)

This scheduler runs more frequently (e.g., every 2 minutes) to check for and process completed jobs.

Functionality:

1. Query our database for AIBatchRequest records with a status that is still in-flight (e.g., VALIDATING, IN_PROGRESS, FINALIZING).
2. For each active batch, call openai.batches.retrieve(batch.openaiBatchId) to get the latest status from OpenAI.
3. Update the status of our AIBatchRequest record to match the one from OpenAI.
4. If a batch's status becomes completed:
   1. Update its status in our DB and store the output_file_id and error_file_id.
   2. Download the content of the output_file_id from OpenAI.
   3. Parse the resulting .jsonl file line by line. For each line:
      1. Use the custom_id to find our original AIProcessingRequest record.
      2. If the line contains a response, parse the AI content and usage data. Update our AIProcessingRequest record with this data and set its processingStatus to PROCESSING_COMPLETE.
      3. If the line contains an error, log it and set the processingStatus to PROCESSING_FAILED.
   4. Do the same for the error_file_id if it exists.
   5. Once all results are processed, update the parent AIBatchRequest status to PROCESSED and set its processedAt timestamp.

---

Phase 3: Implement the Internal Admin API

Create a new set of internal API endpoints for monitoring and managing this process.

• Location: app/api/admin/legacy/
• Authentication: Protect all these endpoints with our most secure admin-level authentication middleware (e.g., from lib/platform-auth.ts). Access should be strictly limited.

Endpoint 1: Get Summary

• Route: GET /api/admin/legacy/summary

• Description: Returns a count of all AIProcessingRequest records, grouped by processingStatus.

• Response:

  {
    "ok": true,
    "summary": {
      "pending_batching": 15231,
      "batching_in_progress": 2500,
      "processing_complete": 85432,
      "processing_failed": 78
    }
  }
  

Endpoint 2: List Requests

• Route: GET /api/admin/legacy/requests

• Description: Retrieves a paginated list of AIProcessingRequest records, filterable by status.

• Query Params: status (required), limit (optional), cursor (optional).

• Response:

  {
    "ok": true,
    "requests": [
      {
        "id": "...",
        "sessionId": "...",
        "status": "processing_failed",
        "failedAt": "2024-03-15T10:23:45Z",
        "error": "Timeout during processing"
      }
    ],
    "nextCursor": "..."
  }
  

Endpoint 3: Re-queue Failed Requests

• Route: POST /api/admin/legacy/requests/requeue

• Description: Resets the status of specified failed requests back to PENDING_BATCHING so they can be re-processed in a new batch.

• Request Body:

  {
    "requestIds": ["req_id_1", "req_id_2"]
  }
  

• Response:

  {
    "ok": true,
    "requeuedCount": 2,
    "notFoundCount": 0
  }
  

---

Phase 4: Final Integration and Cleanup

1. Update server.ts and lib/schedulers.ts: Disable the old processingScheduler and enable the two new schedulers (batchCreator, batchPoller). Ensure they are controlled by environment variables (e.g., BATCH_CREATION_ENABLED, BATCH_POLLING_ENABLED).
2. Documentation: Add a section to CLAUDE.md or a new file in docs/ explaining the new batch processing architecture and the purpose of the admin API endpoints.
3. Environment Variables: Add any new required environment variables to .env.example.

Please proceed with this refactoring plan. Implement robust logging throughout the new schedulers to ensure we can debug the pipeline effectively.