🛠️fix: apply safety settings to Gemini API (#3533)

- Introduce getSafetySettings method for centralized safety settings
- Apply safety settings uniformly in sendCompletion for all models, including Gemini
- Ensure consistent safety settings application in getCompletion
- Update .env.example to clarify safety settings usage for both APIs
This commit is contained in:
Oliver Faust 2024-08-04 20:08:57 +02:00 committed by GitHub
parent 62854c91d3
commit 433d8f832a
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 56 additions and 59 deletions

View file

@ -117,7 +117,7 @@ BINGAI_TOKEN=user_provided
GOOGLE_KEY=user_provided
# GOOGLE_REVERSE_PROXY=
# Gemini API
# Gemini API (AI Studio)
# GOOGLE_MODELS=gemini-1.5-flash-latest,gemini-1.0-pro,gemini-1.0-pro-001,gemini-1.0-pro-latest,gemini-1.0-pro-vision-latest,gemini-1.5-pro-latest,gemini-pro,gemini-pro-vision
# Vertex AI
@ -125,20 +125,24 @@ GOOGLE_KEY=user_provided
# GOOGLE_TITLE_MODEL=gemini-pro
# Google Gemini Safety Settings
# NOTE (Vertex AI): You do not have access to the BLOCK_NONE setting by default.
# To use this restricted HarmBlockThreshold setting, you will need to either:
# Google Safety Settings
# NOTE: These settings apply to both Vertex AI and Gemini API (AI Studio)
#
# (a) Get access through an allowlist via your Google account team
# (b) Switch your account type to monthly invoiced billing following this instruction:
# https://cloud.google.com/billing/docs/how-to/invoiced-billing
# For Vertex AI:
# To use the BLOCK_NONE setting, you need either:
# (a) Access through an allowlist via your Google account team, or
# (b) Switch to monthly invoiced billing: https://cloud.google.com/billing/docs/how-to/invoiced-billing
#
# For Gemini API (AI Studio):
# BLOCK_NONE is available by default, no special account requirements.
#
# Available options: BLOCK_NONE, BLOCK_ONLY_HIGH, BLOCK_MEDIUM_AND_ABOVE, BLOCK_LOW_AND_ABOVE
#
# GOOGLE_SAFETY_SEXUALLY_EXPLICIT=BLOCK_ONLY_HIGH
# GOOGLE_SAFETY_HATE_SPEECH=BLOCK_ONLY_HIGH
# GOOGLE_SAFETY_HARASSMENT=BLOCK_ONLY_HIGH
# GOOGLE_SAFETY_DANGEROUS_CONTENT=BLOCK_ONLY_HIGH
#============#
# OpenAI #
#============#

View file

@ -677,7 +677,6 @@ class GoogleClient extends BaseClient {
const modelName = clientOptions.modelName ?? clientOptions.model ?? '';
if (modelName?.includes('1.5') && !this.project_id) {
/** @type {GenerativeModel} */
const client = model;
const requestOptions = {
contents: _payload,
@ -693,8 +692,7 @@ class GoogleClient extends BaseClient {
};
}
const safetySettings = _payload.safetySettings;
requestOptions.safetySettings = safetySettings;
requestOptions.safetySettings = _payload.safetySettings;
const delay = modelName.includes('flash') ? 8 : 14;
const result = await client.generateContentStream(requestOptions);
@ -709,11 +707,10 @@ class GoogleClient extends BaseClient {
return reply;
}
const safetySettings = _payload.safetySettings;
const stream = await model.stream(messages, {
signal: abortController.signal,
timeout: 7000,
safetySettings: safetySettings,
safetySettings: _payload.safetySettings,
});
let delay = this.options.streamRate || 8;
@ -871,38 +868,34 @@ class GoogleClient extends BaseClient {
}
async sendCompletion(payload, opts = {}) {
const modelName = payload.parameters?.model;
if (modelName && modelName.toLowerCase().includes('gemini')) {
const safetySettings = [
{
category: 'HARM_CATEGORY_SEXUALLY_EXPLICIT',
threshold:
process.env.GOOGLE_SAFETY_SEXUALLY_EXPLICIT || 'HARM_BLOCK_THRESHOLD_UNSPECIFIED',
},
{
category: 'HARM_CATEGORY_HATE_SPEECH',
threshold: process.env.GOOGLE_SAFETY_HATE_SPEECH || 'HARM_BLOCK_THRESHOLD_UNSPECIFIED',
},
{
category: 'HARM_CATEGORY_HARASSMENT',
threshold: process.env.GOOGLE_SAFETY_HARASSMENT || 'HARM_BLOCK_THRESHOLD_UNSPECIFIED',
},
{
category: 'HARM_CATEGORY_DANGEROUS_CONTENT',
threshold:
process.env.GOOGLE_SAFETY_DANGEROUS_CONTENT || 'HARM_BLOCK_THRESHOLD_UNSPECIFIED',
},
];
payload.safetySettings = safetySettings;
}
payload.safetySettings = this.getSafetySettings();
let reply = '';
reply = await this.getCompletion(payload, opts);
return reply.trim();
}
getSafetySettings() {
return [
{
category: 'HARM_CATEGORY_SEXUALLY_EXPLICIT',
threshold: process.env.GOOGLE_SAFETY_SEXUALLY_EXPLICIT || 'HARM_BLOCK_THRESHOLD_UNSPECIFIED',
},
{
category: 'HARM_CATEGORY_HATE_SPEECH',
threshold: process.env.GOOGLE_SAFETY_HATE_SPEECH || 'HARM_BLOCK_THRESHOLD_UNSPECIFIED',
},
{
category: 'HARM_CATEGORY_HARASSMENT',
threshold: process.env.GOOGLE_SAFETY_HARASSMENT || 'HARM_BLOCK_THRESHOLD_UNSPECIFIED',
},
{
category: 'HARM_CATEGORY_DANGEROUS_CONTENT',
threshold: process.env.GOOGLE_SAFETY_DANGEROUS_CONTENT || 'HARM_BLOCK_THRESHOLD_UNSPECIFIED',
},
];
}
/* TO-DO: Handle tokens with Google tokenization NOTE: these are required */
static getTokenizer(encoding, isModelName = false, extendSpecialTokens = {}) {
if (tokenizersCache[encoding]) {