Conversation Troubleshooting

Overview

Conversational issues usually stem from network connectivity hiccups, excessive system prompt size, restrictive similarity thresholds, or unhandled browser tab lifecycle events.

Slow Response Times & Time to First Token (TTFT)

Expected Benchmarks

• Fast Tier (Llama 3.1 / Groq / GPT-4o-mini): TTFT < 180ms

• Quality Tier (Claude 3.5 Sonnet / GPT-4o): TTFT < 380ms

Troubleshooting High Latency (> 1500ms):

1. Reduce Knowledge Chunk Count: In Project Settings > Knowledge, lower maximum retrieved context chunks from 8 to 4.

1. Switch to Edge Streaming: Ensure WebSocket transport is active instead of long-polling.

1. Inspect AI Provider Status: Check for provider rate limits or latency spikes in the Asiyst status page.

Streaming Abruptly Terminates

If an assistant message stops generating mid-sentence:

1. Token Limits: Check if the maxTokens setting in configuration is set too low (e.g., 150 tokens). Increase to at least 800 tokens.

1. Network Interruption: On mobile devices switching from Wi-Fi to cellular data, the WebSocket connection resets. The SDK will automatically request the remaining message transcript.

Assistant Forgets Earlier Turns

If the assistant fails to recall information provided 3 messages ago:

1. Verify that session.persist is not set to false.

1. Ensure you are not generating a new conversationId on each page navigation.

1. Confirm that the browser is not in private browsing mode with sessionStorage disabled.

Cross-Origin & CORS Errors

If DevTools logs Access-Control-Allow-Origin header missing:

• Verify your website domain is in the Allowed Domains list in the dashboard.

• Ensure you are calling https://api.asiyst.com, not an outdated internal development URL.

Diagnostic Steps

1. Enable debug logging:

      Asiyst.init({ projectId, publicKey, options: { debug: true } });

1. Inspect the Network tab for POST /api/v1/sdk/chat/stream.

1. Check the Diagnostics tab in the dashboard for active error rates.