Troubleshoot creation and output issues

Available in Classic and VPC

The following describes issues that may occur while using CLOVA Studio and how to resolve them.

Inference process not displayed

The inference/thought process is not visible when using the HCX-007 model.

Cause

The playground does not display inference content (thought process, thinking/reasoning trace).

Solution

Check the inference content (thought process, thinking/reasoning trace) of the HCX-007 model in the result.message.thinkingContent field of the Chat Completions v3 API response, or in the choices[0].message.thinking_content field of the OpenAI-compatible API (/chat/completions) response.

Mismatch between entered characters and token count

The number of tokens used is significantly higher than the number of characters entered.

Cause

The token count is calculated not by the number of characters, but by the number of word pieces (tokens) obtained by segmenting the sentence according to language-specific tokenization rules. Additionally, in the Chat Completions API, the entire message including system, user, assistant etc. is counted toward the input token count. In particular, setting a long system prompt can significantly increase the number of tokens used.

Solution

Use the Token calculator to check the token count for the entire request syntax in advance.

HCX-007 tool call error

Tool calls work normally with HCX-005, but an error occurs with HCX-007.

Cause

The HCX-007 is a hybrid inference model. Function calling and similar features cannot be used when using the inference function. The default value of the parameter thinking.effort, which sets whether inference is performed and the depth of the thought process, is 'low'. Therefore, if not set separately, the inference function is used.

Solution

If function calling is required with HCX-007, additional parameters must be defined to disable the inference function.

model_hcx007 = ChatClovaX(model="HCX-007", api_key={api_key}, reasoning_effort="none")

For detailed information, see Inference, Function calling.

Masking of specific information (phone numbers, URLs)

Some information, such as phone numbers and URLs, is automatically masked with * when transmitted.
Masking cannot be disabled.

Cause

According to Article 10, Clause 2 (Rights and Obligations, etc.) of the CLOVA Studio Terms of Service, personal information must not be entered when using CLOVA Studio.

Solution

In CLOVA Studio, personally identifiable information such as phone numbers, emails, and URLs is automatically masked to protect privacy. This feature cannot be disabled due to system security and privacy policies. For testing or service validation that requires personal information, use sample data (e.g., 010-XXXX-XXXX, test@example.com) or de-identified data.