14,727
edits
Troubleshooting of Claude: Difference between revisions
Jump to navigation
Jump to search
→Troubleshooting of Claude
| Line 3: | Line 3: | ||
== Troubleshooting of Claude == | == Troubleshooting of Claude == | ||
=== How to fix "stop_reason": "refusal" in Claude API === | |||
'''Error message''' | |||
<pre> | |||
{ | |||
"id": "msg_xxx", | |||
"type": "message", | |||
"role": "assistant", | |||
"model": "claude-opus-4-20250514", | |||
"content": [], | |||
"stop_reason": "refusal", | |||
"stop_sequence": null, | |||
"usage": { | |||
"input_tokens": 2164, | |||
"cache_creation_input_tokens": 0, | |||
"cache_read_input_tokens": 0, | |||
"output_tokens": 0 | |||
} | |||
} | |||
</pre> | |||
'''Model Affected''': Claude Opus 4 | |||
'''Suggestion''' | |||
Prompt met "refusal" error | |||
<pre> | |||
Please help find the company's unified business number: | |||
Company name: {$companyName} | |||
</pre> | |||
Revised prompt: | |||
<pre> | |||
Please help find the following company's publicly available basic information: the company's unified business number: | |||
Company name: {COMPANY_NAME} | |||
</pre> | |||
'''Root cause:''' Common Triggers: | |||
* Privacy-related requests - asking for personal data or private information | |||
* Confidential information queries - requesting internal data or business secrets | |||
* Security-sensitive content - potentially harmful or inappropriate requests | |||
* Ambiguous identity information - queries that may involve identity identification | |||
'''Suggested:''' | |||
* Replace sensitive keywords in your prompts (e.g., "confidential" → "public", "internal data" → "public documents", "private information" → "publicly available information") | |||
* Add safety declarations explicitly stating you only want publicly available information | |||
* Use structured prompts that specify data source limitations and public nature requirements | |||
* Implement the smart retry mechanism with prompt modification when refusal is detected | |||
* Consider using the `RefusalPreventionHelper` class to assess and sanitize prompts before sending to API | |||
'''Official Documentation:''' | |||
* [https://docs.anthropic.com/en/api/handling-stop-reasons Handling stop reasons - Anthropic] mentioned "Claude’s response includes a stop_reason field that indicates why the model stopped generating its response ..." | |||
* [https://www.anthropic.com/transparency Anthropic’s Transparency Hub \ Anthropic] | |||
* [https://www.anthropic.com/news/activating-asl3-protections Activating AI Safety Level 3 Protections \ Anthropic] | |||
=== How to fix "Text extraction failed for one of the uploaded files. Please try again." === | === How to fix "Text extraction failed for one of the uploaded files. Please try again." === | ||
'''Error message''' | '''Error message''' | ||