Fix github action failure and unlock Gemini limit #7971
Conversation
Set HARM levels of gemini to 'OFF'
|
@ous50 is attempting to deploy a commit to the LobeHub Team on Vercel. A member of the Team first needs to authorize it. |
dosubot
bot
added
the
size:S
|
👍 @ous50 Thank you for raising your pull request and contributing to our Community |
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
There was a problem hiding this comment.
PR Summary
This PR makes two significant changes to LobeChat that require careful review:
- 🔒 Modifies Google AI's safety settings by setting
HarmBlockThreshold.OFFfor all Gemini models, effectively disabling content safety filters which could have security implications - 🔑 Updates GitHub Actions sync workflow to use Personal Access Token (PAT) with specific permissions (
Actions,Commit statuses,Contents,Pull requests,Workflows) instead of defaultGITHUB_TOKEN ⚠️ The removal of safety thresholds for Gemini models requires thorough security assessment before merging- 🔄 The sync workflow changes require users to manually generate PAT with correct scopes
- 🛡️ Consider keeping some safety thresholds enabled or adding user warnings about unfiltered content
💡 (1/5) You can manually trigger the bot by mentioning @greptileai in a comment!
2 file(s) reviewed, no comment(s)
Edit PR Review Bot Settings | Greptile
.github/workflows/sync.yml
Outdated
| @@ -32,7 +32,8 @@ jobs: | |||
| upstream_sync_repo: lobehub/lobe-chat | |||
| upstream_sync_branch: main | |||
| target_sync_branch: main | |||
| target_repo_token: ${{ secrets.GITHUB_TOKEN }} # automatically generated, no need to set | |||
| target_repo_token: ${{ secrets.PAT_FOR_SYNC }} # automatically generated, no need to set | |||
There was a problem hiding this comment.
This is not ideal. please add GITHUB_TOKEN as fallback
| // if (modelsOffSafetySettings.has(model)) { | ||
| // return 'OFF' as HarmBlockThreshold; // https://discuss.ai.google.dev/t/59352 | ||
| // } | ||
| return HarmBlockThreshold.OFF; |
There was a problem hiding this comment.
Where's the document about this update?
There was a problem hiding this comment.
https://ai.google.dev/api/generate-content#harmblockthreshold
FYI, the lowest level supported by HARM_CIVIC_INTEGRITY category is BLOCK_NONE when I tested it. This is not mentioned in the documentation.
There was a problem hiding this comment.
https://ai.google.dev/api/generate-content#v1beta.HarmCategory
You can use the OFF value, but there is no evidence to show any difference in results between BLOCK_NONE and OFF.
The probability of being truncated by the safety filters now is actually very low, far less stringent than the model's built-in safety measures.
There was a problem hiding this comment.
but there is no evidence to show any difference in results between
BLOCK_NONEandOFF.
I have tested on Cherry Studio, it works perfectly. At the some time, it also reduces the probability of Gemini poping out unintended words (such as words in Hindi or Korean in a large paragraph of Chinese or Japanese).
There was a problem hiding this comment.
Could you please test all models, including older models like Gemma and Gemini 1.5, to avoid potential compatibility issues?
Add the GITHUB_TOKEN back as a fallback for users to execute.
💻 变更类型 | Change Type
🔀 变更说明 | Description of Change
Actions,Commit statuses,Contents,Pull requestsandWorkflowspermissions.📝 补充信息 | Additional Information