mirror of
https://github.com/danny-avila/LibreChat.git
synced 2025-12-16 16:30:15 +01:00
🧠 feat: Thinking Budget, Include Thoughts, and Dynamic Thinking for Gemini 2.5 (#8055)
* feat: support thinking budget parameter for Gemini 2.5 series (#6949, #7542) https://ai.google.dev/gemini-api/docs/thinking#set-budget * refactor: update thinking budget minimum value to -1 for dynamic thinking - see: https://ai.google.dev/gemini-api/docs/thinking#set-budget * chore: bump @librechat/agents to v2.4.43 * refactor: rename LLMConfigOptions to OpenAIConfigOptions for clarity and consistency - Updated type definitions and references in initialize.ts, llm.ts, and openai.ts to reflect the new naming convention. - Ensured that the OpenAI configuration options are consistently used across the relevant files. * refactor: port Google LLM methods to TypeScript Package * chore: update @librechat/agents version to 2.4.43 in package-lock.json and package.json * refactor: update thinking budget description for clarity and adjust placeholder in parameter settings * refactor: enhance googleSettings default value for thinking budget to support dynamic adjustment * chore: update @librechat/agents to v2.4.44 for Vertex Dynamic Thinking workaround * refactor: rename google config function, update `createRun` types, use `reasoning` as `reasoningKey` for Google * refactor: simplify placeholder handling in DynamicInput component * refactor: enhance thinking budget description for clarity and allow automatic decision by setting to "-1" * refactor: update text styling in OptionHover component for improved readability * chore: update @librechat/agents dependency to v2.4.46 in package.json and package-lock.json * chore: update @librechat/api version to 1.2.5 in package.json and package-lock.json * refactor: enhance `clientOptions` handling by filtering `omitTitleOptions`, add `json` field for Google models --------- Co-authored-by: ciffelia <15273128+ciffelia@users.noreply.github.com>
This commit is contained in:
Danny Avila
GitHub
parent
b169306096
commit
c87422a1e0
21 changed files with 212 additions and 108 deletions
|
|
@ -1,7 +1,7 @@
|
|||
const { google } = require('googleapis');
|
||||
const { Tokenizer } = require('@librechat/api');
|
||||
const { concat } = require('@langchain/core/utils/stream');
|
||||
const { ChatVertexAI } = require('@langchain/google-vertexai');
|
||||
const { Tokenizer, getSafetySettings } = require('@librechat/api');
|
||||
const { ChatGoogleGenerativeAI } = require('@langchain/google-genai');
|
||||
const { GoogleGenerativeAI: GenAI } = require('@google/generative-ai');
|
||||
const { HumanMessage, SystemMessage } = require('@langchain/core/messages');
|
||||
|
|
@ -12,13 +12,13 @@ const {
|
|||
endpointSettings,
|
||||
parseTextParts,
|
||||
EModelEndpoint,
|
||||
googleSettings,
|
||||
ContentTypes,
|
||||
VisionModes,
|
||||
ErrorTypes,
|
||||
Constants,
|
||||
AuthKeys,
|
||||
} = require('librechat-data-provider');
|
||||
const { getSafetySettings } = require('~/server/services/Endpoints/google/llm');
|
||||
const { encodeAndFormat } = require('~/server/services/Files/images');
|
||||
const { spendTokens } = require('~/models/spendTokens');
|
||||
const { getModelMaxTokens } = require('~/utils');
|
||||
|
|
@ -166,6 +166,16 @@ class GoogleClient extends BaseClient {
|
|||
);
|
||||
}
|
||||
|
||||
// Add thinking configuration
|
||||
this.modelOptions.thinkingConfig = {
|
||||
thinkingBudget:
|
||||
(this.modelOptions.thinking ?? googleSettings.thinking.default)
|
||||
? this.modelOptions.thinkingBudget
|
||||
: 0,
|
||||
};
|
||||
delete this.modelOptions.thinking;
|
||||
delete this.modelOptions.thinkingBudget;
|
||||
|
||||
this.sender =
|
||||
this.options.sender ??
|
||||
getResponseSender({
|
||||
|
|
|
|||
|
|
@ -48,7 +48,7 @@
|
|||
"@langchain/google-genai": "^0.2.13",
|
||||
"@langchain/google-vertexai": "^0.2.13",
|
||||
"@langchain/textsplitters": "^0.1.0",
|
||||
"@librechat/agents": "^2.4.42",
|
||||
"@librechat/agents": "^2.4.46",
|
||||
"@librechat/api": "*",
|
||||
"@librechat/data-schemas": "*",
|
||||
"@node-saml/passport-saml": "^5.0.0",
|
||||
|
|
|
|||
|
|
@ -44,6 +44,17 @@ const BaseClient = require('~/app/clients/BaseClient');
|
|||
const { loadAgent } = require('~/models/Agent');
|
||||
const { getMCPManager } = require('~/config');
|
||||
|
||||
const omitTitleOptions = new Set([
|
||||
'stream',
|
||||
'thinking',
|
||||
'streaming',
|
||||
'clientOptions',
|
||||
'thinkingConfig',
|
||||
'thinkingBudget',
|
||||
'includeThoughts',
|
||||
'maxOutputTokens',
|
||||
]);
|
||||
|
||||
/**
|
||||
* @param {ServerRequest} req
|
||||
* @param {Agent} agent
|
||||
|
|
@ -1038,6 +1049,16 @@ class AgentClient extends BaseClient {
|
|||
delete clientOptions.maxTokens;
|
||||
}
|
||||
|
||||
clientOptions = Object.assign(
|
||||
Object.fromEntries(
|
||||
Object.entries(clientOptions).filter(([key]) => !omitTitleOptions.has(key)),
|
||||
),
|
||||
);
|
||||
|
||||
if (provider === Providers.GOOGLE) {
|
||||
clientOptions.json = true;
|
||||
}
|
||||
|
||||
try {
|
||||
const titleResult = await this.run.generateTitle({
|
||||
provider,
|
||||
|
|
|
|||
|
|
@ -1,7 +1,6 @@
|
|||
const { getGoogleConfig, isEnabled } = require('@librechat/api');
|
||||
const { EModelEndpoint, AuthKeys } = require('librechat-data-provider');
|
||||
const { getUserKey, checkUserKeyExpiry } = require('~/server/services/UserService');
|
||||
const { getLLMConfig } = require('~/server/services/Endpoints/google/llm');
|
||||
const { isEnabled } = require('~/server/utils');
|
||||
const { GoogleClient } = require('~/app');
|
||||
|
||||
const initializeClient = async ({ req, res, endpointOption, overrideModel, optionsOnly }) => {
|
||||
|
|
@ -65,7 +64,7 @@ const initializeClient = async ({ req, res, endpointOption, overrideModel, optio
|
|||
if (overrideModel) {
|
||||
clientOptions.modelOptions.model = overrideModel;
|
||||
}
|
||||
return getLLMConfig(credentials, clientOptions);
|
||||
return getGoogleConfig(credentials, clientOptions);
|
||||
}
|
||||
|
||||
const client = new GoogleClient(credentials, clientOptions);
|
||||
|
|
|
|||
|
|
@ -46,6 +46,10 @@ function DynamicInput({
|
|||
setInputValue(e, !isNaN(Number(e.target.value)));
|
||||
};
|
||||
|
||||
const placeholderText = placeholderCode
|
||||
? localize(placeholder as TranslationKeys) || placeholder
|
||||
: placeholder;
|
||||
|
||||
return (
|
||||
<div
|
||||
className={`flex flex-col items-center justify-start gap-6 ${
|
||||
|
|
@ -76,11 +80,7 @@ function DynamicInput({
|
|||
disabled={readonly}
|
||||
value={inputValue ?? defaultValue ?? ''}
|
||||
onChange={handleInputChange}
|
||||
placeholder={
|
||||
placeholderCode
|
||||
? localize(placeholder as TranslationKeys) || placeholder
|
||||
: placeholder
|
||||
}
|
||||
placeholder={placeholderText}
|
||||
className={cn(
|
||||
'flex h-10 max-h-10 w-full resize-none border-none bg-surface-secondary px-3 py-2',
|
||||
)}
|
||||
|
|
|
|||
|
|
@ -29,7 +29,7 @@ function OptionHover({
|
|||
<HoverCardPortal>
|
||||
<HoverCardContent side={side} className={`z-[999] w-80 ${className}`} sideOffset={sideOffset}>
|
||||
<div className="space-y-2">
|
||||
<p className="text-sm text-gray-600 dark:text-gray-300">{text}</p>
|
||||
<p className="whitespace-pre-wrap text-sm text-text-secondary">{text}</p>
|
||||
</div>
|
||||
</HoverCardContent>
|
||||
</HoverCardPortal>
|
||||
|
|
|
|||
|
|
@ -207,6 +207,8 @@
|
|||
"com_endpoint_google_temp": "Higher values = more random, while lower values = more focused and deterministic. We recommend altering this or Top P but not both.",
|
||||
"com_endpoint_google_topk": "Top-k changes how the model selects tokens for output. A top-k of 1 means the selected token is the most probable among all tokens in the model's vocabulary (also called greedy decoding), while a top-k of 3 means that the next token is selected from among the 3 most probable tokens (using temperature).",
|
||||
"com_endpoint_google_topp": "Top-p changes how the model selects tokens for output. Tokens are selected from most K (see topK parameter) probable to least until the sum of their probabilities equals the top-p value.",
|
||||
"com_endpoint_google_thinking": "Enables or disables reasoning. This setting is only supported by certain models (2.5 series). For older models, this setting may have no effect.",
|
||||
"com_endpoint_google_thinking_budget": "Guides the number of thinking tokens the model uses. The actual amount may exceed or fall below this value depending on the prompt.\n\nThis setting is only supported by certain models (2.5 series). Gemini 2.5 Pro supports 128-32,768 tokens. Gemini 2.5 Flash supports 0-24,576 tokens. Gemini 2.5 Flash Lite supports 512-24,576 tokens.\n\nLeave blank or set to \"-1\" to let the model automatically decide when and how much to think. By default, Gemini 2.5 Flash Lite does not think.",
|
||||
"com_endpoint_instructions_assistants": "Override Instructions",
|
||||
"com_endpoint_instructions_assistants_placeholder": "Overrides the instructions of the assistant. This is useful for modifying the behavior on a per-run basis.",
|
||||
"com_endpoint_max_output_tokens": "Max Output Tokens",
|
||||
|
|
@ -582,6 +584,7 @@
|
|||
"com_ui_auth_url": "Authorization URL",
|
||||
"com_ui_authentication": "Authentication",
|
||||
"com_ui_authentication_type": "Authentication Type",
|
||||
"com_ui_auto": "Auto",
|
||||
"com_ui_available_tools": "Available Tools",
|
||||
"com_ui_avatar": "Avatar",
|
||||
"com_ui_azure": "Azure",
|
||||
|
|
|
|||
64
package-lock.json
generated
64
package-lock.json
generated
|
|
@ -64,7 +64,7 @@
|
|||
"@langchain/google-genai": "^0.2.13",
|
||||
"@langchain/google-vertexai": "^0.2.13",
|
||||
"@langchain/textsplitters": "^0.1.0",
|
||||
"@librechat/agents": "^2.4.42",
|
||||
"@librechat/agents": "^2.4.46",
|
||||
"@librechat/api": "*",
|
||||
"@librechat/data-schemas": "*",
|
||||
"@node-saml/passport-saml": "^5.0.0",
|
||||
|
|
@ -1351,33 +1351,6 @@
|
|||
}
|
||||
}
|
||||
},
|
||||
"api/node_modules/@librechat/agents": {
|
||||
"version": "2.4.42",
|
||||
"resolved": "https://registry.npmjs.org/@librechat/agents/-/agents-2.4.42.tgz",
|
||||
"integrity": "sha512-52ux2PeEAV79yr6/h6GN3omlpqX6H0FYl6qwjJ6gT04MMko/imnLd3bQrX0gm3i0KL5ygHbRjQeonONKjJayHw==",
|
||||
"license": "MIT",
|
||||
"dependencies": {
|
||||
"@langchain/anthropic": "^0.3.23",
|
||||
"@langchain/aws": "^0.1.11",
|
||||
"@langchain/community": "^0.3.47",
|
||||
"@langchain/core": "^0.3.60",
|
||||
"@langchain/deepseek": "^0.0.2",
|
||||
"@langchain/google-genai": "^0.2.13",
|
||||
"@langchain/google-vertexai": "^0.2.13",
|
||||
"@langchain/langgraph": "^0.3.4",
|
||||
"@langchain/mistralai": "^0.2.1",
|
||||
"@langchain/ollama": "^0.2.3",
|
||||
"@langchain/openai": "^0.5.14",
|
||||
"@langchain/xai": "^0.0.3",
|
||||
"cheerio": "^1.0.0",
|
||||
"dotenv": "^16.4.7",
|
||||
"https-proxy-agent": "^7.0.6",
|
||||
"nanoid": "^3.3.7"
|
||||
},
|
||||
"engines": {
|
||||
"node": ">=14.0.0"
|
||||
}
|
||||
},
|
||||
"api/node_modules/@smithy/abort-controller": {
|
||||
"version": "4.0.2",
|
||||
"resolved": "https://registry.npmjs.org/@smithy/abort-controller/-/abort-controller-4.0.2.tgz",
|
||||
|
|
@ -19153,9 +19126,9 @@
|
|||
}
|
||||
},
|
||||
"node_modules/@langchain/langgraph": {
|
||||
"version": "0.3.4",
|
||||
"resolved": "https://registry.npmjs.org/@langchain/langgraph/-/langgraph-0.3.4.tgz",
|
||||
"integrity": "sha512-Vuja8Qtu3Zjx7k4fK7Cnw+p8gtvIRPciWp9btPhAs3aUo6aBgOJOZVcK5Ii3mHfEHK/aQmRElR0x/u/YwykOrg==",
|
||||
"version": "0.3.5",
|
||||
"resolved": "https://registry.npmjs.org/@langchain/langgraph/-/langgraph-0.3.5.tgz",
|
||||
"integrity": "sha512-7astlgnp6BdMQJqmr+cbDgR10FYWNCaDLnbfEDHpqhKCCajU59m5snOdl4Vtu5UM6V2k3lgatNqWoflBtxhIyg==",
|
||||
"license": "MIT",
|
||||
"dependencies": {
|
||||
"@langchain/langgraph-checkpoint": "~0.0.18",
|
||||
|
|
@ -19463,11 +19436,10 @@
|
|||
}
|
||||
},
|
||||
"node_modules/@librechat/agents": {
|
||||
"version": "2.4.41",
|
||||
"resolved": "https://registry.npmjs.org/@librechat/agents/-/agents-2.4.41.tgz",
|
||||
"integrity": "sha512-kYmdk5WVRp0qZxTx6BuGCs4l0Ir9iBLLx4ZY4/1wxr80al5/vq3P8wbgGdKMeO2qTu4ZaT4RyWRQYWBg5HDkUQ==",
|
||||
"version": "2.4.46",
|
||||
"resolved": "https://registry.npmjs.org/@librechat/agents/-/agents-2.4.46.tgz",
|
||||
"integrity": "sha512-zR27U19/WGF3HN64oBbiaFgjjWHaF7BjYzRFWzQKEkk+iEzCe59IpuEZUizQ54YcY02nhhh6S3MNUjhAJwMYVA==",
|
||||
"license": "MIT",
|
||||
"peer": true,
|
||||
"dependencies": {
|
||||
"@langchain/anthropic": "^0.3.23",
|
||||
"@langchain/aws": "^0.1.11",
|
||||
|
|
@ -19495,7 +19467,6 @@
|
|||
"resolved": "https://registry.npmjs.org/@langchain/community/-/community-0.3.47.tgz",
|
||||
"integrity": "sha512-Vo42kAfkXpTFSevhEkeqqE55az8NyQgDktCbitXYuhipNbFYx08XVvqEDkFkB20MM/Z7u+cvLb+DxCqnKuH0CQ==",
|
||||
"license": "MIT",
|
||||
"peer": true,
|
||||
"dependencies": {
|
||||
"@langchain/openai": ">=0.2.0 <0.6.0",
|
||||
"@langchain/weaviate": "^0.2.0",
|
||||
|
|
@ -20017,11 +19988,10 @@
|
|||
}
|
||||
},
|
||||
"node_modules/@librechat/agents/node_modules/@langchain/openai": {
|
||||
"version": "0.5.14",
|
||||
"resolved": "https://registry.npmjs.org/@langchain/openai/-/openai-0.5.14.tgz",
|
||||
"integrity": "sha512-0GEj5K/qi1MRuZ4nE7NvyI4jTG+RSewLZqsExUwRukWdeqmkPNHGrogTa5ZDt7eaJxAaY7EgLC5ZnvCM3L1oug==",
|
||||
"version": "0.5.15",
|
||||
"resolved": "https://registry.npmjs.org/@langchain/openai/-/openai-0.5.15.tgz",
|
||||
"integrity": "sha512-ANadEHyAj5sufQpz+SOPpKbyoMcTLhnh8/d+afbSPUqWsIMPpEFX3HoSY3nrBPG6l4NQQNG5P5oHb4SdC8+YIg==",
|
||||
"license": "MIT",
|
||||
"peer": true,
|
||||
"dependencies": {
|
||||
"js-tiktoken": "^1.0.12",
|
||||
"openai": "^5.3.0",
|
||||
|
|
@ -20039,7 +20009,6 @@
|
|||
"resolved": "https://registry.npmjs.org/agent-base/-/agent-base-7.1.3.tgz",
|
||||
"integrity": "sha512-jRR5wdylq8CkOe6hei19GGZnxM6rBGwFl3Bg0YItGDimvjGtAvdZk4Pu6Cl4u4Igsws4a1fd1Vq3ezrhn4KmFw==",
|
||||
"license": "MIT",
|
||||
"peer": true,
|
||||
"engines": {
|
||||
"node": ">= 14"
|
||||
}
|
||||
|
|
@ -20049,7 +20018,6 @@
|
|||
"resolved": "https://registry.npmjs.org/https-proxy-agent/-/https-proxy-agent-7.0.6.tgz",
|
||||
"integrity": "sha512-vK9P5/iUfdl95AI+JVyUuIcVtd4ofvtrOr3HNtM2yxC9bnMbEdp3x01OhQNnjb8IJYi38VlTE3mBXwcfvywuSw==",
|
||||
"license": "MIT",
|
||||
"peer": true,
|
||||
"dependencies": {
|
||||
"agent-base": "^7.1.2",
|
||||
"debug": "4"
|
||||
|
|
@ -20059,11 +20027,10 @@
|
|||
}
|
||||
},
|
||||
"node_modules/@librechat/agents/node_modules/openai": {
|
||||
"version": "5.5.1",
|
||||
"resolved": "https://registry.npmjs.org/openai/-/openai-5.5.1.tgz",
|
||||
"integrity": "sha512-5i19097mGotHA1eFsM6Tjd/tJ8uo9sa5Ysv4Q6bKJ2vtN6rc0MzMrUefXnLXYAJcmMQrC1Efhj0AvfIkXrQamw==",
|
||||
"version": "5.7.0",
|
||||
"resolved": "https://registry.npmjs.org/openai/-/openai-5.7.0.tgz",
|
||||
"integrity": "sha512-zXWawZl6J/P5Wz57/nKzVT3kJQZvogfuyuNVCdEp4/XU2UNrjL7SsuNpWAyLZbo6HVymwmnfno9toVzBhelygA==",
|
||||
"license": "Apache-2.0",
|
||||
"peer": true,
|
||||
"bin": {
|
||||
"openai": "bin/cli"
|
||||
},
|
||||
|
|
@ -20089,7 +20056,6 @@
|
|||
"https://github.com/sponsors/ctavan"
|
||||
],
|
||||
"license": "MIT",
|
||||
"peer": true,
|
||||
"bin": {
|
||||
"uuid": "dist/bin/uuid"
|
||||
}
|
||||
|
|
@ -46568,7 +46534,7 @@
|
|||
},
|
||||
"packages/api": {
|
||||
"name": "@librechat/api",
|
||||
"version": "1.2.4",
|
||||
"version": "1.2.5",
|
||||
"license": "ISC",
|
||||
"devDependencies": {
|
||||
"@babel/preset-env": "^7.21.5",
|
||||
|
|
@ -46600,7 +46566,7 @@
|
|||
"typescript": "^5.0.4"
|
||||
},
|
||||
"peerDependencies": {
|
||||
"@librechat/agents": "^2.4.41",
|
||||
"@librechat/agents": "^2.4.46",
|
||||
"@librechat/data-schemas": "*",
|
||||
"@modelcontextprotocol/sdk": "^1.12.3",
|
||||
"axios": "^1.8.2",
|
||||
|
|
|
|||
|
|
@ -1,6 +1,6 @@
|
|||
{
|
||||
"name": "@librechat/api",
|
||||
"version": "1.2.4",
|
||||
"version": "1.2.5",
|
||||
"type": "commonjs",
|
||||
"description": "MCP services for LibreChat",
|
||||
"main": "dist/index.js",
|
||||
|
|
@ -69,7 +69,7 @@
|
|||
"registry": "https://registry.npmjs.org/"
|
||||
},
|
||||
"peerDependencies": {
|
||||
"@librechat/agents": "^2.4.41",
|
||||
"@librechat/agents": "^2.4.46",
|
||||
"@librechat/data-schemas": "*",
|
||||
"@modelcontextprotocol/sdk": "^1.12.3",
|
||||
"axios": "^1.8.2",
|
||||
|
|
|
|||
|
|
@ -46,7 +46,10 @@ export async function createRun({
|
|||
customHandlers?: Record<GraphEvents, EventHandler>;
|
||||
}): Promise<Run<IState>> {
|
||||
const provider =
|
||||
providerEndpointMap[agent.provider as keyof typeof providerEndpointMap] ?? agent.provider;
|
||||
(providerEndpointMap[
|
||||
agent.provider as keyof typeof providerEndpointMap
|
||||
] as unknown as Providers) ?? agent.provider;
|
||||
|
||||
const llmConfig: t.RunLLMConfig = Object.assign(
|
||||
{
|
||||
provider,
|
||||
|
|
@ -66,7 +69,9 @@ export async function createRun({
|
|||
}
|
||||
|
||||
let reasoningKey: 'reasoning_content' | 'reasoning' | undefined;
|
||||
if (
|
||||
if (provider === Providers.GOOGLE) {
|
||||
reasoningKey = 'reasoning';
|
||||
} else if (
|
||||
llmConfig.configuration?.baseURL?.includes(KnownEndpoints.openrouter) ||
|
||||
(agent.endpoint && agent.endpoint.toLowerCase().includes(KnownEndpoints.openrouter))
|
||||
) {
|
||||
|
|
|
|||
1
packages/api/src/endpoints/google/index.ts
Normal file
1
packages/api/src/endpoints/google/index.ts
Normal file
|
|
@ -0,0 +1 @@
|
|||
export * from './llm';
|
||||
|
|
@ -1,13 +1,15 @@
|
|||
const { Providers } = require('@librechat/agents');
|
||||
const { AuthKeys } = require('librechat-data-provider');
|
||||
const { isEnabled } = require('~/server/utils');
|
||||
import { Providers } from '@librechat/agents';
|
||||
import { googleSettings, AuthKeys } from 'librechat-data-provider';
|
||||
import type { GoogleClientOptions, VertexAIClientOptions } from '@librechat/agents';
|
||||
import type * as t from '~/types';
|
||||
import { isEnabled } from '~/utils';
|
||||
|
||||
function getThresholdMapping(model) {
|
||||
function getThresholdMapping(model: string) {
|
||||
const gemini1Pattern = /gemini-(1\.0|1\.5|pro$|1\.0-pro|1\.5-pro|1\.5-flash-001)/;
|
||||
const restrictedPattern = /(gemini-(1\.5-flash-8b|2\.0|exp)|learnlm)/;
|
||||
|
||||
if (gemini1Pattern.test(model)) {
|
||||
return (value) => {
|
||||
return (value: string) => {
|
||||
if (value === 'OFF') {
|
||||
return 'BLOCK_NONE';
|
||||
}
|
||||
|
|
@ -16,7 +18,7 @@ function getThresholdMapping(model) {
|
|||
}
|
||||
|
||||
if (restrictedPattern.test(model)) {
|
||||
return (value) => {
|
||||
return (value: string) => {
|
||||
if (value === 'OFF' || value === 'HARM_BLOCK_THRESHOLD_UNSPECIFIED') {
|
||||
return 'BLOCK_NONE';
|
||||
}
|
||||
|
|
@ -24,19 +26,16 @@ function getThresholdMapping(model) {
|
|||
};
|
||||
}
|
||||
|
||||
return (value) => value;
|
||||
return (value: string) => value;
|
||||
}
|
||||
|
||||
/**
|
||||
*
|
||||
* @param {string} model
|
||||
* @returns {Array<{category: string, threshold: string}> | undefined}
|
||||
*/
|
||||
function getSafetySettings(model) {
|
||||
export function getSafetySettings(
|
||||
model?: string,
|
||||
): Array<{ category: string; threshold: string }> | undefined {
|
||||
if (isEnabled(process.env.GOOGLE_EXCLUDE_SAFETY_SETTINGS)) {
|
||||
return undefined;
|
||||
}
|
||||
const mapThreshold = getThresholdMapping(model);
|
||||
const mapThreshold = getThresholdMapping(model ?? '');
|
||||
|
||||
return [
|
||||
{
|
||||
|
|
@ -74,24 +73,27 @@ function getSafetySettings(model) {
|
|||
* Replicates core logic from GoogleClient's constructor and setOptions, plus client determination.
|
||||
* Returns an object with the provider label and the final options that would be passed to createLLM.
|
||||
*
|
||||
* @param {string | object} credentials - Either a JSON string or an object containing Google keys
|
||||
* @param {object} [options={}] - The same shape as the "GoogleClient" constructor options
|
||||
* @param credentials - Either a JSON string or an object containing Google keys
|
||||
* @param options - The same shape as the "GoogleClient" constructor options
|
||||
*/
|
||||
|
||||
function getLLMConfig(credentials, options = {}) {
|
||||
// 1. Parse credentials
|
||||
let creds = {};
|
||||
export function getGoogleConfig(
|
||||
credentials: string | t.GoogleCredentials | undefined,
|
||||
options: t.GoogleConfigOptions = {},
|
||||
) {
|
||||
let creds: t.GoogleCredentials = {};
|
||||
if (typeof credentials === 'string') {
|
||||
try {
|
||||
creds = JSON.parse(credentials);
|
||||
} catch (err) {
|
||||
throw new Error(`Error parsing string credentials: ${err.message}`);
|
||||
} catch (err: unknown) {
|
||||
throw new Error(
|
||||
`Error parsing string credentials: ${err instanceof Error ? err.message : 'Unknown error'}`,
|
||||
);
|
||||
}
|
||||
} else if (credentials && typeof credentials === 'object') {
|
||||
creds = credentials;
|
||||
}
|
||||
|
||||
// Extract from credentials
|
||||
const serviceKeyRaw = creds[AuthKeys.GOOGLE_SERVICE_KEY] ?? {};
|
||||
const serviceKey =
|
||||
typeof serviceKeyRaw === 'string' ? JSON.parse(serviceKeyRaw) : (serviceKeyRaw ?? {});
|
||||
|
|
@ -102,9 +104,15 @@ function getLLMConfig(credentials, options = {}) {
|
|||
const reverseProxyUrl = options.reverseProxyUrl;
|
||||
const authHeader = options.authHeader;
|
||||
|
||||
/** @type {GoogleClientOptions | VertexAIClientOptions} */
|
||||
let llmConfig = {
|
||||
...(options.modelOptions || {}),
|
||||
const {
|
||||
thinking = googleSettings.thinking.default,
|
||||
thinkingBudget = googleSettings.thinkingBudget.default,
|
||||
...modelOptions
|
||||
} = options.modelOptions || {};
|
||||
|
||||
const llmConfig: GoogleClientOptions | VertexAIClientOptions = {
|
||||
...(modelOptions || {}),
|
||||
model: modelOptions?.model ?? '',
|
||||
maxRetries: 2,
|
||||
};
|
||||
|
||||
|
|
@ -121,16 +129,30 @@ function getLLMConfig(credentials, options = {}) {
|
|||
|
||||
// If we have a GCP project => Vertex AI
|
||||
if (project_id && provider === Providers.VERTEXAI) {
|
||||
/** @type {VertexAIClientOptions['authOptions']} */
|
||||
llmConfig.authOptions = {
|
||||
(llmConfig as VertexAIClientOptions).authOptions = {
|
||||
credentials: { ...serviceKey },
|
||||
projectId: project_id,
|
||||
};
|
||||
llmConfig.location = process.env.GOOGLE_LOC || 'us-central1';
|
||||
(llmConfig as VertexAIClientOptions).location = process.env.GOOGLE_LOC || 'us-central1';
|
||||
} else if (apiKey && provider === Providers.GOOGLE) {
|
||||
llmConfig.apiKey = apiKey;
|
||||
}
|
||||
|
||||
const shouldEnableThinking =
|
||||
thinking && thinkingBudget != null && (thinkingBudget > 0 || thinkingBudget === -1);
|
||||
|
||||
if (shouldEnableThinking && provider === Providers.GOOGLE) {
|
||||
(llmConfig as GoogleClientOptions).thinkingConfig = {
|
||||
thinkingBudget: thinking ? thinkingBudget : googleSettings.thinkingBudget.default,
|
||||
includeThoughts: Boolean(thinking),
|
||||
};
|
||||
} else if (shouldEnableThinking && provider === Providers.VERTEXAI) {
|
||||
(llmConfig as VertexAIClientOptions).thinkingBudget = thinking
|
||||
? thinkingBudget
|
||||
: googleSettings.thinkingBudget.default;
|
||||
(llmConfig as VertexAIClientOptions).includeThoughts = Boolean(thinking);
|
||||
}
|
||||
|
||||
/*
|
||||
let legacyOptions = {};
|
||||
// Filter out any "examples" that are empty
|
||||
|
|
@ -152,11 +174,11 @@ function getLLMConfig(credentials, options = {}) {
|
|||
*/
|
||||
|
||||
if (reverseProxyUrl) {
|
||||
llmConfig.baseUrl = reverseProxyUrl;
|
||||
(llmConfig as GoogleClientOptions).baseUrl = reverseProxyUrl;
|
||||
}
|
||||
|
||||
if (authHeader) {
|
||||
llmConfig.customHeaders = {
|
||||
(llmConfig as GoogleClientOptions).customHeaders = {
|
||||
Authorization: `Bearer ${apiKey}`,
|
||||
};
|
||||
}
|
||||
|
|
@ -169,8 +191,3 @@ function getLLMConfig(credentials, options = {}) {
|
|||
llmConfig,
|
||||
};
|
||||
}
|
||||
|
||||
module.exports = {
|
||||
getLLMConfig,
|
||||
getSafetySettings,
|
||||
};
|
||||
|
|
@ -1 +1,2 @@
|
|||
export * from './google';
|
||||
export * from './openai';
|
||||
|
|
|
|||
|
|
@ -1,9 +1,9 @@
|
|||
import { ErrorTypes, EModelEndpoint, mapModelToAzureConfig } from 'librechat-data-provider';
|
||||
import type {
|
||||
LLMConfigOptions,
|
||||
UserKeyValues,
|
||||
InitializeOpenAIOptionsParams,
|
||||
OpenAIOptionsResult,
|
||||
OpenAIConfigOptions,
|
||||
InitializeOpenAIOptionsParams,
|
||||
} from '~/types';
|
||||
import { createHandleLLMNewToken } from '~/utils/generators';
|
||||
import { getAzureCredentials } from '~/utils/azure';
|
||||
|
|
@ -64,7 +64,7 @@ export const initializeOpenAI = async ({
|
|||
? userValues?.baseURL
|
||||
: baseURLOptions[endpoint as keyof typeof baseURLOptions];
|
||||
|
||||
const clientOptions: LLMConfigOptions = {
|
||||
const clientOptions: OpenAIConfigOptions = {
|
||||
proxy: PROXY ?? undefined,
|
||||
reverseProxyUrl: baseURL || undefined,
|
||||
streaming: true,
|
||||
|
|
@ -135,7 +135,7 @@ export const initializeOpenAI = async ({
|
|||
user: req.user.id,
|
||||
};
|
||||
|
||||
const finalClientOptions: LLMConfigOptions = {
|
||||
const finalClientOptions: OpenAIConfigOptions = {
|
||||
...clientOptions,
|
||||
modelOptions,
|
||||
};
|
||||
|
|
|
|||
|
|
@ -13,7 +13,7 @@ import { isEnabled } from '~/utils/common';
|
|||
*/
|
||||
export function getOpenAIConfig(
|
||||
apiKey: string,
|
||||
options: t.LLMConfigOptions = {},
|
||||
options: t.OpenAIConfigOptions = {},
|
||||
endpoint?: string | null,
|
||||
): t.LLMConfigResult {
|
||||
const {
|
||||
|
|
|
|||
24
packages/api/src/types/google.ts
Normal file
24
packages/api/src/types/google.ts
Normal file
|
|
@ -0,0 +1,24 @@
|
|||
import { z } from 'zod';
|
||||
import { AuthKeys, googleBaseSchema } from 'librechat-data-provider';
|
||||
|
||||
export type GoogleParameters = z.infer<typeof googleBaseSchema>;
|
||||
|
||||
export type GoogleCredentials = {
|
||||
[AuthKeys.GOOGLE_SERVICE_KEY]?: string;
|
||||
[AuthKeys.GOOGLE_API_KEY]?: string;
|
||||
};
|
||||
|
||||
/**
|
||||
* Configuration options for the getLLMConfig function
|
||||
*/
|
||||
export interface GoogleConfigOptions {
|
||||
modelOptions?: Partial<GoogleParameters>;
|
||||
reverseProxyUrl?: string;
|
||||
defaultQuery?: Record<string, string | undefined>;
|
||||
headers?: Record<string, string>;
|
||||
proxy?: string;
|
||||
streaming?: boolean;
|
||||
authHeader?: boolean;
|
||||
addParams?: Record<string, unknown>;
|
||||
dropParams?: string[];
|
||||
}
|
||||
|
|
@ -1,5 +1,6 @@
|
|||
export * from './azure';
|
||||
export * from './events';
|
||||
export * from './google';
|
||||
export * from './mistral';
|
||||
export * from './openai';
|
||||
export * from './run';
|
||||
|
|
|
|||
|
|
@ -9,7 +9,7 @@ export type OpenAIParameters = z.infer<typeof openAISchema>;
|
|||
/**
|
||||
* Configuration options for the getLLMConfig function
|
||||
*/
|
||||
export interface LLMConfigOptions {
|
||||
export interface OpenAIConfigOptions {
|
||||
modelOptions?: Partial<OpenAIParameters>;
|
||||
reverseProxyUrl?: string;
|
||||
defaultQuery?: Record<string, string | undefined>;
|
||||
|
|
|
|||
|
|
@ -1,8 +1,9 @@
|
|||
import type { AgentModelParameters, EModelEndpoint } from 'librechat-data-provider';
|
||||
import type { Providers } from '@librechat/agents';
|
||||
import type { AgentModelParameters } from 'librechat-data-provider';
|
||||
import type { OpenAIConfiguration } from './openai';
|
||||
|
||||
export type RunLLMConfig = {
|
||||
provider: EModelEndpoint;
|
||||
provider: Providers;
|
||||
streaming: boolean;
|
||||
streamUsage: boolean;
|
||||
usage?: boolean;
|
||||
|
|
|
|||
|
|
@ -450,6 +450,37 @@ const google: Record<string, SettingDefinition> = {
|
|||
optionType: 'model',
|
||||
columnSpan: 2,
|
||||
},
|
||||
thinking: {
|
||||
key: 'thinking',
|
||||
label: 'com_endpoint_thinking',
|
||||
labelCode: true,
|
||||
description: 'com_endpoint_google_thinking',
|
||||
descriptionCode: true,
|
||||
type: 'boolean',
|
||||
default: googleSettings.thinking.default,
|
||||
component: 'switch',
|
||||
optionType: 'conversation',
|
||||
showDefault: false,
|
||||
columnSpan: 2,
|
||||
},
|
||||
thinkingBudget: {
|
||||
key: 'thinkingBudget',
|
||||
label: 'com_endpoint_thinking_budget',
|
||||
labelCode: true,
|
||||
description: 'com_endpoint_google_thinking_budget',
|
||||
descriptionCode: true,
|
||||
placeholder: 'com_ui_auto',
|
||||
placeholderCode: true,
|
||||
type: 'number',
|
||||
component: 'input',
|
||||
range: {
|
||||
min: googleSettings.thinkingBudget.min,
|
||||
max: googleSettings.thinkingBudget.max,
|
||||
step: googleSettings.thinkingBudget.step,
|
||||
},
|
||||
optionType: 'conversation',
|
||||
columnSpan: 2,
|
||||
},
|
||||
};
|
||||
|
||||
const googleConfig: SettingsConfiguration = [
|
||||
|
|
@ -461,6 +492,8 @@ const googleConfig: SettingsConfiguration = [
|
|||
google.topP,
|
||||
google.topK,
|
||||
librechat.resendFiles,
|
||||
google.thinking,
|
||||
google.thinkingBudget,
|
||||
];
|
||||
|
||||
const googleCol1: SettingsConfiguration = [
|
||||
|
|
@ -476,6 +509,8 @@ const googleCol2: SettingsConfiguration = [
|
|||
google.topP,
|
||||
google.topK,
|
||||
librechat.resendFiles,
|
||||
google.thinking,
|
||||
google.thinkingBudget,
|
||||
];
|
||||
|
||||
const openAI: SettingsConfiguration = [
|
||||
|
|
|
|||
|
|
@ -255,6 +255,18 @@ export const googleSettings = {
|
|||
step: 1 as const,
|
||||
default: 40 as const,
|
||||
},
|
||||
thinking: {
|
||||
default: true as const,
|
||||
},
|
||||
thinkingBudget: {
|
||||
min: -1 as const,
|
||||
max: 32768 as const,
|
||||
step: 1 as const,
|
||||
/** `-1` = Dynamic Thinking, meaning the model will adjust
|
||||
* the budget based on the complexity of the request.
|
||||
*/
|
||||
default: -1 as const,
|
||||
},
|
||||
};
|
||||
|
||||
const ANTHROPIC_MAX_OUTPUT = 128000 as const;
|
||||
|
|
@ -785,6 +797,8 @@ export const googleBaseSchema = tConversationSchema.pick({
|
|||
artifacts: true,
|
||||
topP: true,
|
||||
topK: true,
|
||||
thinking: true,
|
||||
thinkingBudget: true,
|
||||
iconURL: true,
|
||||
greeting: true,
|
||||
spec: true,
|
||||
|
|
@ -810,6 +824,12 @@ export const googleGenConfigSchema = z
|
|||
presencePenalty: coerceNumber.optional(),
|
||||
frequencyPenalty: coerceNumber.optional(),
|
||||
stopSequences: z.array(z.string()).optional(),
|
||||
thinkingConfig: z
|
||||
.object({
|
||||
includeThoughts: z.boolean().optional(),
|
||||
thinkingBudget: coerceNumber.optional(),
|
||||
})
|
||||
.optional(),
|
||||
})
|
||||
.strip()
|
||||
.optional();
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue