🧠 feat: Bedrock Anthropic Reasoning & Update Endpoint Handling (#6163)

* feat: Add thinking and thinkingBudget parameters for Bedrock Anthropic models

* chore: Update @librechat/agents to version 2.1.8

* refactor: change region order in params

* refactor: Add maxTokens parameter to conversation preset schema

* refactor: Update agent client to use bedrockInputSchema and improve error handling for model parameters

* refactor: streamline/optimize llmConfig initialization and saving for bedrock

* fix: ensure config titleModel is used for all endpoints

* refactor: enhance OpenAIClient and agent initialization to support endpoint checks for OpenRouter

* chore: bump @google/generative-ai

api/app/clients/GoogleClient.js

@@ -827,7 +827,8 @@ class GoogleClient extends BaseClient {
     let reply = '';
     const { abortController } = options;
 
-    const model = this.modelOptions.modelName ?? this.modelOptions.model ?? '';
+    const model =
+      this.options.titleModel ?? this.modelOptions.modelName ?? this.modelOptions.model ?? '';
     const safetySettings = getSafetySettings(model);
     if (!EXCLUDED_GENAI_MODELS.test(model) && !this.project_id) {
       logger.debug('Identified titling model as GenAI version');

api/app/clients/OpenAIClient.js

@@ -112,7 +112,12 @@ class OpenAIClient extends BaseClient {
     const { OPENAI_FORCE_PROMPT } = process.env ?? {};
     const { reverseProxyUrl: reverseProxy } = this.options;
 
-    if (!this.useOpenRouter && reverseProxy && reverseProxy.includes(KnownEndpoints.openrouter)) {
+    if (
+      !this.useOpenRouter &&
+      ((reverseProxy && reverseProxy.includes(KnownEndpoints.openrouter)) ||
+        (this.options.endpoint &&
+          this.options.endpoint.toLowerCase().includes(KnownEndpoints.openrouter)))
+    ) {
       this.useOpenRouter = true;
     }
 

api/models/schema/defaults.js

@@ -56,6 +56,10 @@ const conversationPreset = {
     type: Number,
     required: false,
   },
+  maxTokens: {
+    type: Number,
+    required: false,
+  },
   presence_penalty: {
     type: Number,
     required: false,

api/package.json

@@ -36,7 +36,7 @@
   "dependencies": {
     "@anthropic-ai/sdk": "^0.37.0",
     "@azure/search-documents": "^12.0.0",
-    "@google/generative-ai": "^0.21.0",
+    "@google/generative-ai": "^0.23.0",
     "@googleapis/youtube": "^20.0.0",
     "@keyv/mongo": "^2.1.8",
     "@keyv/redis": "^2.8.1",
@@ -45,7 +45,7 @@
     "@langchain/google-genai": "^0.1.9",
     "@langchain/google-vertexai": "^0.2.0",
     "@langchain/textsplitters": "^0.1.0",
-    "@librechat/agents": "^2.1.7",
+    "@librechat/agents": "^2.1.8",
     "@waylaidwanderer/fetch-event-source": "^3.0.1",
     "axios": "1.7.8",
     "bcryptjs": "^2.4.3",

api/server/controllers/agents/client.js

@@ -17,7 +17,7 @@ const {
   KnownEndpoints,
   anthropicSchema,
   isAgentsEndpoint,
-  bedrockOutputParser,
+  bedrockInputSchema,
   removeNullishValues,
 } = require('librechat-data-provider');
 const {
@@ -30,6 +30,7 @@ const {
 const { spendTokens, spendStructuredTokens } = require('~/models/spendTokens');
 const { getBufferString, HumanMessage } = require('@langchain/core/messages');
 const { encodeAndFormat } = require('~/server/services/Files/images/encode');
+const { getCustomEndpointConfig } = require('~/server/services/Config');
 const Tokenizer = require('~/server/services/Tokenizer');
 const BaseClient = require('~/app/clients/BaseClient');
 const { createRun } = require('./run');
@@ -39,10 +40,10 @@ const { logger } = require('~/config');
 /** @typedef {import('@langchain/core/runnables').RunnableConfig} RunnableConfig */
 
 const providerParsers = {
-  [EModelEndpoint.openAI]: openAISchema,
-  [EModelEndpoint.azureOpenAI]: openAISchema,
-  [EModelEndpoint.anthropic]: anthropicSchema,
-  [EModelEndpoint.bedrock]: bedrockOutputParser,
+  [EModelEndpoint.openAI]: openAISchema.parse,
+  [EModelEndpoint.azureOpenAI]: openAISchema.parse,
+  [EModelEndpoint.anthropic]: anthropicSchema.parse,
+  [EModelEndpoint.bedrock]: bedrockInputSchema.parse,
 };
 
 const legacyContentEndpoints = new Set([KnownEndpoints.groq, KnownEndpoints.deepseek]);
@@ -187,7 +188,14 @@ class AgentClient extends BaseClient {
         : {};
 
     if (parseOptions) {
-      runOptions = parseOptions(this.options.agent.model_parameters);
+      try {
+        runOptions = parseOptions(this.options.agent.model_parameters);
+      } catch (error) {
+        logger.error(
+          '[api/server/controllers/agents/client.js #getSaveOptions] Error parsing options',
+          error,
+        );
+      }
     }
 
     return removeNullishValues(
@@ -824,13 +832,16 @@ class AgentClient extends BaseClient {
     const clientOptions = {
       maxTokens: 75,
     };
-    const providerConfig = this.options.req.app.locals[this.options.agent.provider];
+    let endpointConfig = this.options.req.app.locals[this.options.agent.endpoint];
+    if (!endpointConfig) {
+      endpointConfig = await getCustomEndpointConfig(this.options.agent.endpoint);
+    }
     if (
-      providerConfig &&
-      providerConfig.titleModel &&
-      providerConfig.titleModel !== Constants.CURRENT_MODEL
+      endpointConfig &&
+      endpointConfig.titleModel &&
+      endpointConfig.titleModel !== Constants.CURRENT_MODEL
     ) {
-      clientOptions.model = providerConfig.titleModel;
+      clientOptions.model = endpointConfig.titleModel;
     }
     try {
       const titleResult = await this.run.generateTitle({

api/server/controllers/agents/run.js

@@ -45,7 +45,10 @@ async function createRun({
 
   /** @type {'reasoning_content' | 'reasoning'} */
   let reasoningKey;
-  if (llmConfig.configuration?.baseURL?.includes(KnownEndpoints.openrouter)) {
+  if (
+    llmConfig.configuration?.baseURL?.includes(KnownEndpoints.openrouter) ||
+    (agent.endpoint && agent.endpoint.toLowerCase().includes(KnownEndpoints.openrouter))
+  ) {
     reasoningKey = 'reasoning';
   }
   if (/o1(?!-(?:mini|preview)).*$/.test(llmConfig.model)) {

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

@@ -101,6 +101,7 @@ const initializeAgentOptions = async ({
   });
 
   const provider = agent.provider;
+  agent.endpoint = provider;
   let getOptions = providerConfigMap[provider];
   if (!getOptions && providerConfigMap[provider.toLowerCase()] != null) {
     agent.provider = provider.toLowerCase();
@@ -112,9 +113,7 @@ const initializeAgentOptions = async ({
     }
     getOptions = initCustom;
     agent.provider = Providers.OPENAI;
-    agent.endpoint = provider.toLowerCase();
   }
-
   const model_parameters = Object.assign(
     {},
     agent.model_parameters ?? { model: agent.model },

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

@@ -27,6 +27,7 @@ const initializeClient = async ({ req, res, endpointOption, overrideModel, optio
 
   if (anthropicConfig) {
     clientOptions.streamRate = anthropicConfig.streamRate;
+    clientOptions.titleModel = anthropicConfig.titleModel;
   }
 
   /** @type {undefined | TBaseEndpoint} */

api/server/services/Endpoints/bedrock/build.js

@@ -1,6 +1,5 @@
-const { removeNullishValues, bedrockInputParser } = require('librechat-data-provider');
+const { removeNullishValues } = require('librechat-data-provider');
 const generateArtifactsPrompt = require('~/app/clients/prompts/artifacts');
-const { logger } = require('~/config');
 
 const buildOptions = (endpoint, parsedBody) => {
   const {
@@ -15,12 +14,6 @@ const buildOptions = (endpoint, parsedBody) => {
     artifacts,
     ...model_parameters
   } = parsedBody;
-  let parsedParams = model_parameters;
-  try {
-    parsedParams = bedrockInputParser.parse(model_parameters);
-  } catch (error) {
-    logger.warn('Failed to parse bedrock input', error);
-  }
   const endpointOption = removeNullishValues({
     endpoint,
     name,
@@ -31,7 +24,7 @@ const buildOptions = (endpoint, parsedBody) => {
     spec,
     promptPrefix,
     maxContextTokens,
-    model_parameters: parsedParams,
+    model_parameters,
   });
 
   if (typeof artifacts === 'string') {

api/server/services/Endpoints/bedrock/options.js

@@ -1,14 +1,16 @@
 const { HttpsProxyAgent } = require('https-proxy-agent');
 const {
-  EModelEndpoint,
-  Constants,
   AuthType,
+  Constants,
+  EModelEndpoint,
+  bedrockInputParser,
+  bedrockOutputParser,
   removeNullishValues,
 } = require('librechat-data-provider');
 const { getUserKey, checkUserKeyExpiry } = require('~/server/services/UserService');
 const { sleep } = require('~/server/utils');
 
-const getOptions = async ({ req, endpointOption }) => {
+const getOptions = async ({ req, overrideModel, endpointOption }) => {
   const {
     BEDROCK_AWS_SECRET_ACCESS_KEY,
     BEDROCK_AWS_ACCESS_KEY_ID,
@@ -62,39 +64,44 @@ const getOptions = async ({ req, endpointOption }) => {
 
   /** @type {BedrockClientOptions} */
   const requestOptions = {
-    model: endpointOption.model,
+    model: overrideModel ?? endpointOption.model,
     region: BEDROCK_AWS_DEFAULT_REGION,
-    streaming: true,
-    streamUsage: true,
-    callbacks: [
-      {
-        handleLLMNewToken: async () => {
-          if (!streamRate) {
-            return;
-          }
-          await sleep(streamRate);
-        },
-      },
-    ],
   };
 
-  if (credentials) {
-    requestOptions.credentials = credentials;
-  }
-
-  if (BEDROCK_REVERSE_PROXY) {
-    requestOptions.endpointHost = BEDROCK_REVERSE_PROXY;
-  }
-
   const configOptions = {};
   if (PROXY) {
     /** NOTE: NOT SUPPORTED BY BEDROCK */
     configOptions.httpAgent = new HttpsProxyAgent(PROXY);
   }
 
+  const llmConfig = bedrockOutputParser(
+    bedrockInputParser.parse(
+      removeNullishValues(Object.assign(requestOptions, endpointOption.model_parameters)),
+    ),
+  );
+
+  if (credentials) {
+    llmConfig.credentials = credentials;
+  }
+
+  if (BEDROCK_REVERSE_PROXY) {
+    llmConfig.endpointHost = BEDROCK_REVERSE_PROXY;
+  }
+
+  llmConfig.callbacks = [
+    {
+      handleLLMNewToken: async () => {
+        if (!streamRate) {
+          return;
+        }
+        await sleep(streamRate);
+      },
+    },
+  ];
+
   return {
     /** @type {BedrockClientOptions} */
-    llmConfig: removeNullishValues(Object.assign(requestOptions, endpointOption.model_parameters)),
+    llmConfig,
     configOptions,
   };
 };

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

@@ -141,7 +141,7 @@ const initializeClient = async ({ req, res, endpointOption, optionsOnly, overrid
         },
         clientOptions,
       );
-      const options = getLLMConfig(apiKey, clientOptions);
+      const options = getLLMConfig(apiKey, clientOptions, endpoint);
       if (!customOptions.streamRate) {
         return options;
       }

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

@@ -5,12 +5,7 @@ const { isEnabled } = require('~/server/utils');
 const { GoogleClient } = require('~/app');
 
 const initializeClient = async ({ req, res, endpointOption, overrideModel, optionsOnly }) => {
-  const {
-    GOOGLE_KEY,
-    GOOGLE_REVERSE_PROXY,
-    GOOGLE_AUTH_HEADER,
-    PROXY,
-  } = process.env;
+  const { GOOGLE_KEY, GOOGLE_REVERSE_PROXY, GOOGLE_AUTH_HEADER, PROXY } = process.env;
   const isUserProvided = GOOGLE_KEY === 'user_provided';
   const { key: expiresAt } = req.body;
 
@@ -43,6 +38,7 @@ const initializeClient = async ({ req, res, endpointOption, overrideModel, optio
 
   if (googleConfig) {
     clientOptions.streamRate = googleConfig.streamRate;
+    clientOptions.titleModel = googleConfig.titleModel;
   }
 
   if (allConfig) {

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

@@ -113,6 +113,7 @@ const initializeClient = async ({
 
   if (!isAzureOpenAI && openAIConfig) {
     clientOptions.streamRate = openAIConfig.streamRate;
+    clientOptions.titleModel = openAIConfig.titleModel;
   }
 
   /** @type {undefined | TBaseEndpoint} */

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

@@ -23,9 +23,10 @@ const { isEnabled } = require('~/server/utils');
  * @param {boolean} [options.streaming] - Whether to use streaming mode.
  * @param {Object} [options.addParams] - Additional parameters to add to the model options.
  * @param {string[]} [options.dropParams] - Parameters to remove from the model options.
+ * @param {string|null} [endpoint=null] - The endpoint name
  * @returns {Object} Configuration options for creating an LLM instance.
  */
-function getLLMConfig(apiKey, options = {}) {
+function getLLMConfig(apiKey, options = {}, endpoint = null) {
   const {
     modelOptions = {},
     reverseProxyUrl,
@@ -58,7 +59,10 @@ function getLLMConfig(apiKey, options = {}) {
   let useOpenRouter;
   /** @type {OpenAIClientOptions['configuration']} */
   const configOptions = {};
-  if (reverseProxyUrl && reverseProxyUrl.includes(KnownEndpoints.openrouter)) {
+  if (
+    (reverseProxyUrl && reverseProxyUrl.includes(KnownEndpoints.openrouter)) ||
+    (endpoint && endpoint.toLowerCase().includes(KnownEndpoints.openrouter))
+  ) {
     useOpenRouter = true;
     llmConfig.include_reasoning = true;
     configOptions.baseURL = reverseProxyUrl;

client/src/components/SidePanel/Parameters/settings.ts

@@ -553,8 +553,10 @@ const bedrockAnthropic: SettingsConfiguration = [
   bedrock.topP,
   bedrock.topK,
   baseDefinitions.stop,
-  bedrock.region,
   librechat.resendFiles,
+  bedrock.region,
+  anthropic.thinking,
+  anthropic.thinkingBudget,
 ];
 
 const bedrockMistral: SettingsConfiguration = [
@@ -564,8 +566,8 @@ const bedrockMistral: SettingsConfiguration = [
   bedrock.maxTokens,
   mistral.temperature,
   mistral.topP,
-  bedrock.region,
   librechat.resendFiles,
+  bedrock.region,
 ];
 
 const bedrockCohere: SettingsConfiguration = [
@@ -575,8 +577,8 @@ const bedrockCohere: SettingsConfiguration = [
   bedrock.maxTokens,
   cohere.temperature,
   cohere.topP,
-  bedrock.region,
   librechat.resendFiles,
+  bedrock.region,
 ];
 
 const bedrockGeneral: SettingsConfiguration = [
@@ -585,8 +587,8 @@ const bedrockGeneral: SettingsConfiguration = [
   librechat.maxContextTokens,
   meta.temperature,
   meta.topP,
-  bedrock.region,
   librechat.resendFiles,
+  bedrock.region,
 ];
 
 const bedrockAnthropicCol1: SettingsConfiguration = [
@@ -602,8 +604,10 @@ const bedrockAnthropicCol2: SettingsConfiguration = [
   bedrock.temperature,
   bedrock.topP,
   bedrock.topK,
-  bedrock.region,
   librechat.resendFiles,
+  bedrock.region,
+  anthropic.thinking,
+  anthropic.thinkingBudget,
 ];
 
 const bedrockMistralCol1: SettingsConfiguration = [
@@ -617,8 +621,8 @@ const bedrockMistralCol2: SettingsConfiguration = [
   bedrock.maxTokens,
   mistral.temperature,
   mistral.topP,
-  bedrock.region,
   librechat.resendFiles,
+  bedrock.region,
 ];
 
 const bedrockCohereCol1: SettingsConfiguration = [
@@ -632,8 +636,8 @@ const bedrockCohereCol2: SettingsConfiguration = [
   bedrock.maxTokens,
   cohere.temperature,
   cohere.topP,
-  bedrock.region,
   librechat.resendFiles,
+  bedrock.region,
 ];
 
 const bedrockGeneralCol1: SettingsConfiguration = [
@@ -646,8 +650,8 @@ const bedrockGeneralCol2: SettingsConfiguration = [
   librechat.maxContextTokens,
   meta.temperature,
   meta.topP,
-  bedrock.region,
   librechat.resendFiles,
+  bedrock.region,
 ];
 
 export const settings: Record<string, SettingsConfiguration | undefined> = {

packages/data-provider/src/bedrock.ts

@@ -1,6 +1,20 @@
 import { z } from 'zod';
 import * as s from './schemas';
 
+type ThinkingConfig = {
+  type: 'enabled';
+  budget_tokens: number;
+};
+type AnthropicReasoning = {
+  thinking?: ThinkingConfig | boolean;
+  thinkingBudget?: number;
+};
+
+type AnthropicInput = BedrockConverseInput & {
+  additionalModelRequestFields: BedrockConverseInput['additionalModelRequestFields'] &
+    AnthropicReasoning;
+};
+
 export const bedrockInputSchema = s.tConversationSchema
   .pick({
     /* LibreChat params; optionType: 'conversation' */
@@ -21,11 +35,24 @@ export const bedrockInputSchema = s.tConversationSchema
     temperature: true,
     topP: true,
     stop: true,
+    thinking: true,
+    thinkingBudget: true,
     /* Catch-all fields */
     topK: true,
     additionalModelRequestFields: true,
   })
-  .transform((obj) => s.removeNullishValues(obj))
+  .transform((obj) => {
+    if ((obj as AnthropicInput).additionalModelRequestFields?.thinking != null) {
+      const _obj = obj as AnthropicInput;
+      obj.thinking = !!_obj.additionalModelRequestFields.thinking;
+      obj.thinkingBudget =
+        typeof _obj.additionalModelRequestFields.thinking === 'object'
+          ? (_obj.additionalModelRequestFields.thinking as ThinkingConfig)?.budget_tokens
+          : undefined;
+      delete obj.additionalModelRequestFields;
+    }
+    return s.removeNullishValues(obj);
+  })
   .catch(() => ({}));
 
 export type BedrockConverseInput = z.infer<typeof bedrockInputSchema>;
@@ -49,6 +76,8 @@ export const bedrockInputParser = s.tConversationSchema
     temperature: true,
     topP: true,
     stop: true,
+    thinking: true,
+    thinkingBudget: true,
     /* Catch-all fields */
     topK: true,
     additionalModelRequestFields: true,
@@ -87,6 +116,27 @@ export const bedrockInputParser = s.tConversationSchema
       }
     });
 
+    /** Default thinking and thinkingBudget for 'anthropic.claude-3-7-sonnet' models, if not defined */
+    if (
+      typeof typedData.model === 'string' &&
+      typedData.model.includes('anthropic.claude-3-7-sonnet')
+    ) {
+      if (additionalFields.thinking === undefined) {
+        additionalFields.thinking = true;
+      } else if (additionalFields.thinking === false) {
+        delete additionalFields.thinking;
+        delete additionalFields.thinkingBudget;
+      }
+
+      if (additionalFields.thinking === true && additionalFields.thinkingBudget === undefined) {
+        additionalFields.thinkingBudget = 2000;
+      }
+      additionalFields.anthropic_beta = ['output-128k-2025-02-19'];
+    } else if (additionalFields.thinking != null || additionalFields.thinkingBudget != null) {
+      delete additionalFields.thinking;
+      delete additionalFields.thinkingBudget;
+    }
+
     if (Object.keys(additionalFields).length > 0) {
       typedData.additionalModelRequestFields = {
         ...((typedData.additionalModelRequestFields as Record<string, unknown> | undefined) || {}),
@@ -104,9 +154,34 @@ export const bedrockInputParser = s.tConversationSchema
   })
   .catch(() => ({}));
 
+/**
+ * Configures the "thinking" parameter based on given input and thinking options.
+ *
+ * @param data - The parsed Bedrock request options object
+ * @returns The object with thinking configured appropriately
+ */
+function configureThinking(data: AnthropicInput): AnthropicInput {
+  const updatedData = { ...data };
+  if (updatedData.additionalModelRequestFields?.thinking === true) {
+    updatedData.maxTokens = updatedData.maxTokens ?? updatedData.maxOutputTokens ?? 8192;
+    delete updatedData.maxOutputTokens;
+    const thinkingConfig: AnthropicReasoning['thinking'] = {
+      type: 'enabled',
+      budget_tokens: updatedData.additionalModelRequestFields.thinkingBudget ?? 2000,
+    };
+
+    if (thinkingConfig.budget_tokens > updatedData.maxTokens) {
+      thinkingConfig.budget_tokens = Math.floor(updatedData.maxTokens * 0.9);
+    }
+    updatedData.additionalModelRequestFields.thinking = thinkingConfig;
+    delete updatedData.additionalModelRequestFields.thinkingBudget;
+  }
+  return updatedData;
+}
+
 export const bedrockOutputParser = (data: Record<string, unknown>) => {
   const knownKeys = [...Object.keys(s.tConversationSchema.shape), 'topK', 'top_k'];
-  const result: Record<string, unknown> = {};
+  let result: Record<string, unknown> = {};
 
   // Extract known fields from the root level
   Object.entries(data).forEach(([key, value]) => {
@@ -125,6 +200,8 @@ export const bedrockOutputParser = (data: Record<string, unknown>) => {
         if (knownKeys.includes(key)) {
           if (key === 'top_k') {
             result['topK'] = value;
+          } else if (key === 'thinking' || key === 'thinkingBudget') {
+            return;
           } else {
             result[key] = value;
           }
@@ -140,8 +217,11 @@ export const bedrockOutputParser = (data: Record<string, unknown>) => {
     result.maxTokens = result.maxOutputTokens;
   }
 
-  // Remove additionalModelRequestFields from the result
-  delete result.additionalModelRequestFields;
+  result = configureThinking(result as AnthropicInput);
+  // Remove additionalModelRequestFields from the result if it doesn't thinking config
+  if ((result as AnthropicInput).additionalModelRequestFields?.thinking == null) {
+    delete result.additionalModelRequestFields;
+  }
 
   return result;
 };