🧠 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/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;