🧠 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>

api/app/clients/GoogleClient.js

@@ -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({

api/package.json

@@ -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",

api/server/controllers/agents/client.js

@@ -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,

api/server/services/Endpoints/google/initialize.js

@@ -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);

api/server/services/Endpoints/google/llm.js

@@ -1,176 +0,0 @@
-const { Providers } = require('@librechat/agents');
-const { AuthKeys } = require('librechat-data-provider');
-const { isEnabled } = require('~/server/utils');
-
-function getThresholdMapping(model) {
-  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) => {
-      if (value === 'OFF') {
-        return 'BLOCK_NONE';
-      }
-      return value;
-    };
-  }
-
-  if (restrictedPattern.test(model)) {
-    return (value) => {
-      if (value === 'OFF' || value === 'HARM_BLOCK_THRESHOLD_UNSPECIFIED') {
-        return 'BLOCK_NONE';
-      }
-      return value;
-    };
-  }
-
-  return (value) => value;
-}
-
-/**
- *
- * @param {string} model
- * @returns {Array<{category: string, threshold: string}> | undefined}
- */
-function getSafetySettings(model) {
-  if (isEnabled(process.env.GOOGLE_EXCLUDE_SAFETY_SETTINGS)) {
-    return undefined;
-  }
-  const mapThreshold = getThresholdMapping(model);
-
-  return [
-    {
-      category: 'HARM_CATEGORY_SEXUALLY_EXPLICIT',
-      threshold: mapThreshold(
-        process.env.GOOGLE_SAFETY_SEXUALLY_EXPLICIT || 'HARM_BLOCK_THRESHOLD_UNSPECIFIED',
-      ),
-    },
-    {
-      category: 'HARM_CATEGORY_HATE_SPEECH',
-      threshold: mapThreshold(
-        process.env.GOOGLE_SAFETY_HATE_SPEECH || 'HARM_BLOCK_THRESHOLD_UNSPECIFIED',
-      ),
-    },
-    {
-      category: 'HARM_CATEGORY_HARASSMENT',
-      threshold: mapThreshold(
-        process.env.GOOGLE_SAFETY_HARASSMENT || 'HARM_BLOCK_THRESHOLD_UNSPECIFIED',
-      ),
-    },
-    {
-      category: 'HARM_CATEGORY_DANGEROUS_CONTENT',
-      threshold: mapThreshold(
-        process.env.GOOGLE_SAFETY_DANGEROUS_CONTENT || 'HARM_BLOCK_THRESHOLD_UNSPECIFIED',
-      ),
-    },
-    {
-      category: 'HARM_CATEGORY_CIVIC_INTEGRITY',
-      threshold: mapThreshold(process.env.GOOGLE_SAFETY_CIVIC_INTEGRITY || 'BLOCK_NONE'),
-    },
-  ];
-}
-
-/**
- * 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
- */
-
-function getLLMConfig(credentials, options = {}) {
-  // 1. Parse credentials
-  let creds = {};
-  if (typeof credentials === 'string') {
-    try {
-      creds = JSON.parse(credentials);
-    } catch (err) {
-      throw new Error(`Error parsing string credentials: ${err.message}`);
-    }
-  } 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 ?? {});
-
-  const project_id = serviceKey?.project_id ?? null;
-  const apiKey = creds[AuthKeys.GOOGLE_API_KEY] ?? null;
-
-  const reverseProxyUrl = options.reverseProxyUrl;
-  const authHeader = options.authHeader;
-
-  /** @type {GoogleClientOptions | VertexAIClientOptions} */
-  let llmConfig = {
-    ...(options.modelOptions || {}),
-    maxRetries: 2,
-  };
-
-  /** Used only for Safety Settings */
-  llmConfig.safetySettings = getSafetySettings(llmConfig.model);
-
-  let provider;
-
-  if (project_id) {
-    provider = Providers.VERTEXAI;
-  } else {
-    provider = Providers.GOOGLE;
-  }
-
-  // If we have a GCP project => Vertex AI
-  if (project_id && provider === Providers.VERTEXAI) {
-    /** @type {VertexAIClientOptions['authOptions']} */
-    llmConfig.authOptions = {
-      credentials: { ...serviceKey },
-      projectId: project_id,
-    };
-    llmConfig.location = process.env.GOOGLE_LOC || 'us-central1';
-  } else if (apiKey && provider === Providers.GOOGLE) {
-    llmConfig.apiKey = apiKey;
-  }
-
-  /*
-  let legacyOptions = {};
-  // Filter out any "examples" that are empty
-  legacyOptions.examples = (legacyOptions.examples ?? [])
-    .filter(Boolean)
-    .filter((obj) => obj?.input?.content !== '' && obj?.output?.content !== '');
-
-  // If user has "examples" from legacyOptions, push them onto llmConfig
-  if (legacyOptions.examples?.length) {
-    llmConfig.examples = legacyOptions.examples.map((ex) => {
-      const { input, output } = ex;
-      if (!input?.content || !output?.content) {return undefined;}
-      return {
-        input: new HumanMessage(input.content),
-        output: new AIMessage(output.content),
-      };
-    }).filter(Boolean);
-  }
-  */
-
-  if (reverseProxyUrl) {
-    llmConfig.baseUrl = reverseProxyUrl;
-  }
-
-  if (authHeader) {
-    llmConfig.customHeaders = {
-      Authorization: `Bearer ${apiKey}`,
-    };
-  }
-
-  // Return the final shape
-  return {
-    /** @type {Providers.GOOGLE | Providers.VERTEXAI} */
-    provider,
-    /** @type {GoogleClientOptions | VertexAIClientOptions} */
-    llmConfig,
-  };
-}
-
-module.exports = {
-  getLLMConfig,
-  getSafetySettings,
-};