ChatBase.validateHistory fails for response with finishReason SAFTY

[joerg1985]

Environment details

• Programming language: Java
• OS: Windows 11
• Language runtime version: 21
• Package version: 1.30.0
• Model: gemini-2.5-flash

Steps to reproduce

1. Trigger a response with finishReason SAFETY while using the com.google.genai.Chat
2. The com.google.genai.ChatBase.validateHistory will fail with java.util.NoSuchElementException: No value present as the role of the response is empty. There is no chance for the custom logic to handle the response.

java.util.NoSuchElementException: No value present
java.base/java.util.Optional.get (Optional.java:143)
com.google.genai.ChatBase.validateHistory (ChatBase.java:130)
com.google.genai.ChatBase.recordHistory (ChatBase.java:55)
com.google.genai.ChatBase.updateHistoryNonStreaming (ChatBase.java:258)
com.google.genai.Chat.privateSendMessage (Chat.java:149)
com.google.genai.Chat.sendMessage (Chat.java:104)

[hemasekhar-p]

Hello @joerg1985 , Thank you for bringing this issue to our attention I am investigating this issue but i am unable to reproduce the NoSuchElementException locally, the model currently returns a standard refusal message with a valid 'model' role. Could you please help me with a Minimal Reproducible Example that consistently triggers this error and the specific prompt you are using? This will help me to match your environment exactly and trace the internal ChatBase state.

[joerg1985]

@hemasekhar-p thanks for taking care, i will try to provide as mutch details as possible here.
The issue does only happen very rarely, i run into it today only once, so here is what i found out.

My safty settings are:

 List<SafetySetting> safetySettings
                = List.of(
                        SafetySetting.builder()
                                .category(HarmCategory.Known.HARM_CATEGORY_HATE_SPEECH)
                                .threshold(HarmBlockThreshold.Known.BLOCK_LOW_AND_ABOVE)
                                .build(),
                        SafetySetting.builder()
                                .category(HarmCategory.Known.HARM_CATEGORY_SEXUALLY_EXPLICIT)
                                .threshold(HarmBlockThreshold.Known.BLOCK_LOW_AND_ABOVE)
                                .build(),
                        SafetySetting.builder()
                                .category(HarmCategory.Known.HARM_CATEGORY_DANGEROUS_CONTENT)
                                .threshold(HarmBlockThreshold.Known.BLOCK_LOW_AND_ABOVE)
                                .build()
                );

I could only see this when a function call is completed and i return the outcome of a function to the model.
This is what does trigger the SAFETY, the dump of modelResponse.candidates().get().get(0).toJson():

{"content":{},"finishMessage":"The response is blocked due to safety","finishReason":"SAFETY","safetyRatings":[{"category":"HARM_CATEGORY_HATE_SPEECH","probability":"NEGLIGIBLE","probabilityScore":3.85059E-6,"severity":"HARM_SEVERITY_NEGLIGIBLE","severityScore":0.022735804},{"blocked":true,"category":"HARM_CATEGORY_DANGEROUS_CONTENT","probability":"LOW","probabilityScore":0.25504392,"severity":"HARM_SEVERITY_LOW","severityScore":0.16457349},{"category":"HARM_CATEGORY_HARASSMENT","probability":"NEGLIGIBLE","probabilityScore":2.1515683E-5,"severity":"HARM_SEVERITY_NEGLIGIBLE"},{"category":"HARM_CATEGORY_SEXUALLY_EXPLICIT","probability":"NEGLIGIBLE","probabilityScore":3.42794E-6,"severity":"HARM_SEVERITY_NEGLIGIBLE","severityScore":0.053882897}]}

Repeating the request does usually not trigger the SAFETY, but the ChatBase instance is burned.
The ChatBase.recordHistory does add the faulty entries to comprehensiveHistory before calling validateHistory (lines 52 to 55).
So it might be a good idea to first call validate and then update the internal state of the ChatBase?

[hemasekhar-p]

Hello @joerg1985 , Thank you for your information. I am still unable to repro the issue with your information. could you please provide me the complete code which you are trying. So that i can compare and confirm that we both are in the same path to repro the issue in local and it will be more useful to investigate further to resolve this issue?

[joerg1985]

@hemasekhar-p i am currently on xmas vacation, but i will try to provide the code in january.

[github-actions]

This issue has been marked as stale because it has been open for 7 days with no activity. It will be closed in 2 days if no further activity occurs.

[hemasekhar-p]

@joerg1985 ,Thank you for letting us know. We will keep this issue open and wait for your update in January.