Idempotency & Request Ordering Issue in Chat Flow

[MostafaOsmanFathi]

Idempotency Issue: Concurrent Requests Before Document Ingestion / Response Completion

Description

There is an idempotency and request-ordering issue in the current system flow.

When a user requests to add a document, the system still allows:

• Sending chat/messages before the document is fully added
• Sending multiple messages concurrently without waiting for the response of the previous request

This can lead to:

• Inconsistent RAG context
• Duplicate or invalid requests
• Responses generated without the required document being available
• Race conditions in backend processing

This should be handled through the WebSocket session to ensure proper session-level sequencing.

Current Behavior

1. User requests to add a document
2. System accepts additional messages before document ingestion is completed
3. User can send another message before receiving the response to the previous one
4. Requests may be processed out of order

Expected Behavior

The system should enforce strict sequential processing per WebSocket session:

Document Ingestion Lock

• Once a document add request is initiated, block all chat/message requests on the same WebSocket session
• Allow messages only after document ingestion completes successfully

Message Request Lock (Idempotency)

• Prevent sending another message until the response to the previous message is received
• Ensure only one in-flight request at a time per WebSocket session

Importance

• Ensures idempotent behavior
• Prevents duplicate or overlapping requests
• Guarantees responses are generated using the correct document context
• Maintains WebSocket session consistency
• Improves system reliability and user experience

Suggested Solutions

• Introduce a session-level request state flag (e.g., isProcessing)
• Use backend request locking or queueing per WebSocket session
• Return an error (e.g., 409 Conflict or 429 Too Many Requests) if a request is already in progress
• Disable frontend input while document ingestion is running or a message response is pending