🔧 fix: Google Gemma Support & OpenAI Reasoning Instructions (#7196)

* 🔄 chore: Update @langchain/google-vertexai to version 0.2.5 in package.json and package-lock.json

* chore: temp remove agents

* 🔄 chore: Update @langchain/google-genai to version 0.2.5 in package.json and package-lock.json

* 🔄 chore: Update @langchain/community to version 0.3.42 in package.json and package-lock.json

* 🔄 chore: Add license information for @langchain/textsplitters in package-lock.json

* 🔄 chore: Update @langchain/core to version 0.3.51 in package.json and package-lock.json

* 🔄 chore: Update openai dependency to version 4.96.2 in package.json and package-lock.json

* chore: @librechat/agents to v2.4.30

* fix: streaming condition in ModelEndHandler to account for boundModel `disableStreaming` setting

* fix: update regex for noSystemModel and refactor message handling in AgentClient

* feat: Google Gemma models

* chore: remove unnecessary empty JSX fragment in PopoverButtons component

api/app/clients/GoogleClient.js

@@ -140,8 +140,7 @@ class GoogleClient extends BaseClient {
     this.options.attachments?.then((attachments) => this.checkVisionRequest(attachments));
 
     /** @type {boolean} Whether using a "GenerativeAI" Model */
-    this.isGenerativeModel =
-      this.modelOptions.model.includes('gemini') || this.modelOptions.model.includes('learnlm');
+    this.isGenerativeModel = /gemini|learnlm|gemma/.test(this.modelOptions.model);
 
     this.maxContextTokens =
       this.options.maxContextTokens ??

api/models/tx.js

@@ -111,6 +111,10 @@ const tokenValues = Object.assign(
     /* cohere doesn't have rates for the older command models,
   so this was from https://artificialanalysis.ai/models/command-light/providers */
     command: { prompt: 0.38, completion: 0.38 },
+    gemma: { prompt: 0, completion: 0 }, // https://ai.google.dev/pricing
+    'gemma-2': { prompt: 0, completion: 0 }, // https://ai.google.dev/pricing
+    'gemma-3': { prompt: 0, completion: 0 }, // https://ai.google.dev/pricing
+    'gemma-3-27b': { prompt: 0, completion: 0 }, // https://ai.google.dev/pricing
     'gemini-2.0-flash-lite': { prompt: 0.075, completion: 0.3 },
     'gemini-2.0-flash': { prompt: 0.1, completion: 0.4 },
     'gemini-2.0': { prompt: 0, completion: 0 }, // https://ai.google.dev/pricing

api/package.json

@@ -43,12 +43,12 @@
     "@google/generative-ai": "^0.23.0",
     "@googleapis/youtube": "^20.0.0",
     "@keyv/redis": "^4.3.3",
-    "@langchain/community": "^0.3.39",
-    "@langchain/core": "^0.3.43",
-    "@langchain/google-genai": "^0.2.2",
-    "@langchain/google-vertexai": "^0.2.3",
+    "@langchain/community": "^0.3.42",
+    "@langchain/core": "^0.3.51",
+    "@langchain/google-genai": "^0.2.5",
+    "@langchain/google-vertexai": "^0.2.5",
     "@langchain/textsplitters": "^0.1.0",
-    "@librechat/agents": "^2.4.22",
+    "@librechat/agents": "^2.4.30",
     "@librechat/data-schemas": "*",
     "@waylaidwanderer/fetch-event-source": "^3.0.1",
     "axios": "^1.8.2",
@@ -90,7 +90,7 @@
     "nanoid": "^3.3.7",
     "nodemailer": "^6.9.15",
     "ollama": "^0.5.0",
-    "openai": "^4.47.1",
+    "openai": "^4.96.2",
     "openai-chat-tokens": "^0.2.8",
     "openid-client": "^5.4.2",
     "passport": "^0.6.0",

api/server/controllers/agents/callbacks.js

@@ -61,7 +61,10 @@ class ModelEndHandler {
       }
 
       this.collectedUsage.push(usage);
-      if (!graph.clientOptions?.disableStreaming) {
+      const streamingDisabled = !!(
+        graph.clientOptions?.disableStreaming || graph?.boundModel?.disableStreaming
+      );
+      if (!streamingDisabled) {
         return;
       }
       if (!data.output.content) {

api/server/controllers/agents/client.js

@@ -58,7 +58,7 @@ const payloadParser = ({ req, agent, endpoint }) => {
 
 const legacyContentEndpoints = new Set([KnownEndpoints.groq, KnownEndpoints.deepseek]);
 
-const noSystemModelRegex = [/\b(o\d)\b/gi];
+const noSystemModelRegex = [/\b(o1)\b/gi];
 
 // const { processMemory, memoryInstructions } = require('~/server/services/Endpoints/agents/memory');
 // const { getFormattedMemories } = require('~/models/Memory');
@@ -728,12 +728,14 @@ class AgentClient extends BaseClient {
         }
 
         if (noSystemMessages === true && systemContent?.length) {
-          let latestMessage = _messages.pop().content;
+          const latestMessageContent = _messages.pop().content;
           if (typeof latestMessage !== 'string') {
-            latestMessage = latestMessage[0].text;
+            latestMessageContent[0].text = [systemContent, latestMessageContent[0].text].join('\n');
+            _messages.push(new HumanMessage({ content: latestMessageContent }));
+          } else {
+            const text = [systemContent, latestMessageContent].join('\n');
+            _messages.push(new HumanMessage(text));
           }
-          latestMessage = [systemContent, latestMessage].join('\n');
-          _messages.push(new HumanMessage(latestMessage));
         }
 
         let messages = _messages;

api/utils/tokens.js

@@ -60,6 +60,10 @@ const cohereModels = {
 
 const googleModels = {
   /* Max I/O is combined so we subtract the amount from max response tokens for actual total */
+  gemma: 8196,
+  'gemma-2': 32768,
+  'gemma-3': 32768,
+  'gemma-3-27b': 131072,
   gemini: 30720, // -2048 from max
   'gemini-pro-vision': 12288,
   'gemini-exp': 2000000,