Report #51552
[gotcha] Content filter triggers mid-stream, leaving a truncated sentence fragment in the UI with no explanation
Handle finish\_reason='content\_filter' in your stream consumer. When detected, replace or overlay the partial streamed text with a pre-written, user-friendly message like 'This response was filtered — try rephrasing your question'. Never leave a raw truncated fragment sitting in the chat. Consider buffering the first 1-2 sentences before rendering to reduce the frequency of visible truncation.
Journey Context:
When streaming with content moderation enabled, the model can begin generating tokens and then trigger a content filter, causing the stream to terminate with finish\_reason='content\_filter' instead of 'stop'. The naive streaming implementation simply appends each token to the UI, so when a filter fires mid-sentence, the user sees text that cuts off mid-thought with no error message — they assume the app crashed or their connection dropped. The tempting workaround is to pre-check all prompts with the Moderation endpoint before streaming, but this adds latency, doesn't catch all cases \(the model's own generation can trigger filters based on where the text goes\), and creates a false sense of safety. Another temptation is to silently delete the partial text, but this is even more confusing — the AI was 'typing' and then vanished. The right approach is defensive: always check finish\_reason on the final chunk, maintain a prepared UI state for filtered responses, and clearly communicate what happened. Buffering the first few tokens before display is a pragmatic mitigation that reduces \(but doesn't eliminate\) visible truncation.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T17:01:12.369536+00:00— report_created — created