🔧 refactor: Improve Agent Context & Minor Fixes (#5349)

* refactor: Improve Context for Agents

* 🔧 fix: Safeguard against undefined properties in OpenAIClient response handling

* refactor: log error before re-throwing for original stack trace

* refactor: remove toolResource state from useFileHandling, allow svg files

* refactor: prevent verbose logs from axios errors when using actions

* refactor: add silent method recordTokenUsage in AgentClient

* refactor: streamline token count assignment in BaseClient

* refactor: enhance safety settings handling for Gemini 2.0 model

* fix: capabilities structure in MCPConnection

* refactor: simplify civic integrity threshold handling in GoogleClient and llm

* refactor: update token count retrieval method in BaseClient tests

* ci: fix test for svg

api/app/clients/BaseClient.js

@@ -4,13 +4,15 @@ const {
   supportsBalanceCheck,
   isAgentsEndpoint,
   isParamEndpoint,
+  EModelEndpoint,
   ErrorTypes,
   Constants,
   CacheKeys,
   Time,
 } = require('librechat-data-provider');
 const { getMessages, saveMessage, updateMessage, saveConvo } = require('~/models');
 const { addSpaceIfNeeded, isEnabled } = require('~/server/utils');
+const { truncateToolCallOutputs } = require('./prompts');
 const checkBalance = require('~/models/checkBalance');
 const { getFiles } = require('~/models/File');
 const { getLogStores } = require('~/cache');
@@ -95,7 +97,7 @@ class BaseClient {
    * @returns {number}
    */
   getTokenCountForResponse(responseMessage) {
-    logger.debug('`[BaseClient] recordTokenUsage` not implemented.', responseMessage);
+    logger.debug('[BaseClient] `recordTokenUsage` not implemented.', responseMessage);
   }
 
   /**
@@ -106,7 +108,7 @@ class BaseClient {
    * @returns {Promise<void>}
    */
   async recordTokenUsage({ promptTokens, completionTokens }) {
-    logger.debug('`[BaseClient] recordTokenUsage` not implemented.', {
+    logger.debug('[BaseClient] `recordTokenUsage` not implemented.', {
       promptTokens,
       completionTokens,
     });
@@ -287,6 +289,9 @@ class BaseClient {
   }
 
   async handleTokenCountMap(tokenCountMap) {
+    if (this.clientName === EModelEndpoint.agents) {
+      return;
+    }
     if (this.currentMessages.length === 0) {
       return;
     }
@@ -394,6 +399,21 @@ class BaseClient {
     _instructions && logger.debug('[BaseClient] instructions tokenCount: ' + tokenCount);
     let payload = this.addInstructions(formattedMessages, _instructions);
     let orderedWithInstructions = this.addInstructions(orderedMessages, instructions);
+    if (this.clientName === EModelEndpoint.agents) {
+      const { dbMessages, editedIndices } = truncateToolCallOutputs(
+        orderedWithInstructions,
+        this.maxContextTokens,
+        this.getTokenCountForMessage.bind(this),
+      );
+
+      if (editedIndices.length > 0) {
+        logger.debug('[BaseClient] Truncated tool call outputs:', editedIndices);
+        for (const index of editedIndices) {
+          payload[index].content = dbMessages[index].content;
+        }
+        orderedWithInstructions = dbMessages;
+      }
+    }
 
     let { context, remainingContextTokens, messagesToRefine, summaryIndex } =
       await this.getMessagesWithinTokenLimit(orderedWithInstructions);
@@ -625,7 +645,7 @@ class BaseClient {
         await this.updateUserMessageTokenCount({ usage, tokenCountMap, userMessage, opts });
       } else {
         responseMessage.tokenCount = this.getTokenCountForResponse(responseMessage);
-        completionTokens = this.getTokenCount(completion);
+        completionTokens = responseMessage.tokenCount;
       }
 
       await this.recordTokenUsage({ promptTokens, completionTokens, usage });

api/app/clients/GoogleClient.js

@@ -886,32 +886,42 @@ class GoogleClient extends BaseClient {
   }
 
   getSafetySettings() {
+    const isGemini2 = this.modelOptions.model.includes('gemini-2.0');
+    const mapThreshold = (value) => {
+      if (isGemini2 && value === 'BLOCK_NONE') {
+        return 'OFF';
+      }
+      return value;
+    };
+
     return [
       {
         category: 'HARM_CATEGORY_SEXUALLY_EXPLICIT',
-        threshold:
+        threshold: mapThreshold(
           process.env.GOOGLE_SAFETY_SEXUALLY_EXPLICIT || 'HARM_BLOCK_THRESHOLD_UNSPECIFIED',
+        ),
       },
       {
         category: 'HARM_CATEGORY_HATE_SPEECH',
-        threshold: process.env.GOOGLE_SAFETY_HATE_SPEECH || 'HARM_BLOCK_THRESHOLD_UNSPECIFIED',
+        threshold: mapThreshold(
+          process.env.GOOGLE_SAFETY_HATE_SPEECH || 'HARM_BLOCK_THRESHOLD_UNSPECIFIED',
+        ),
       },
       {
         category: 'HARM_CATEGORY_HARASSMENT',
-        threshold: process.env.GOOGLE_SAFETY_HARASSMENT || 'HARM_BLOCK_THRESHOLD_UNSPECIFIED',
+        threshold: mapThreshold(
+          process.env.GOOGLE_SAFETY_HARASSMENT || 'HARM_BLOCK_THRESHOLD_UNSPECIFIED',
+        ),
       },
       {
         category: 'HARM_CATEGORY_DANGEROUS_CONTENT',
-        threshold:
+        threshold: mapThreshold(
           process.env.GOOGLE_SAFETY_DANGEROUS_CONTENT || 'HARM_BLOCK_THRESHOLD_UNSPECIFIED',
+        ),
       },
       {
         category: 'HARM_CATEGORY_CIVIC_INTEGRITY',
-        /**
-         * Note: this was added since `gemini-2.0-flash-thinking-exp-1219` does not
-         * accept 'HARM_BLOCK_THRESHOLD_UNSPECIFIED' for 'HARM_CATEGORY_CIVIC_INTEGRITY'
-         * */
-        threshold: process.env.GOOGLE_SAFETY_CIVIC_INTEGRITY || 'BLOCK_NONE',
+        threshold: mapThreshold(process.env.GOOGLE_SAFETY_CIVIC_INTEGRITY || 'BLOCK_NONE'),
       },
     ];
   }

api/app/clients/OpenAIClient.js

@@ -1293,7 +1293,7 @@ ${convo}
           });
 
         for await (const chunk of stream) {
-          const token = chunk.choices[0]?.delta?.content || '';
+          const token = chunk?.choices?.[0]?.delta?.content || '';
           intermediateReply.push(token);
           onProgress(token);
           if (abortController.signal.aborted) {

api/app/clients/prompts/index.js

@@ -4,7 +4,7 @@ const summaryPrompts = require('./summaryPrompts');
 const handleInputs = require('./handleInputs');
 const instructions = require('./instructions');
 const titlePrompts = require('./titlePrompts');
-const truncateText = require('./truncateText');
+const truncate = require('./truncate');
 const createVisionPrompt = require('./createVisionPrompt');
 const createContextHandlers = require('./createContextHandlers');
 
@@ -15,7 +15,7 @@ module.exports = {
   ...handleInputs,
   ...instructions,
   ...titlePrompts,
-  ...truncateText,
+  ...truncate,
   createVisionPrompt,
   createContextHandlers,
 };

api/app/clients/prompts/truncate.js

@@ -0,0 +1,115 @@
+const MAX_CHAR = 255;
+
+/**
+ * Truncates a given text to a specified maximum length, appending ellipsis and a notification
+ * if the original text exceeds the maximum length.
+ *
+ * @param {string} text - The text to be truncated.
+ * @param {number} [maxLength=MAX_CHAR] - The maximum length of the text after truncation. Defaults to MAX_CHAR.
+ * @returns {string} The truncated text if the original text length exceeds maxLength, otherwise returns the original text.
+ */
+function truncateText(text, maxLength = MAX_CHAR) {
+  if (text.length > maxLength) {
+    return `${text.slice(0, maxLength)}... [text truncated for brevity]`;
+  }
+  return text;
+}
+
+/**
+ * Truncates a given text to a specified maximum length by showing the first half and the last half of the text,
+ * separated by ellipsis. This method ensures the output does not exceed the maximum length, including the addition
+ * of ellipsis and notification if the original text exceeds the maximum length.
+ *
+ * @param {string} text - The text to be truncated.
+ * @param {number} [maxLength=MAX_CHAR] - The maximum length of the output text after truncation. Defaults to MAX_CHAR.
+ * @returns {string} The truncated text showing the first half and the last half, or the original text if it does not exceed maxLength.
+ */
+function smartTruncateText(text, maxLength = MAX_CHAR) {
+  const ellipsis = '...';
+  const notification = ' [text truncated for brevity]';
+  const halfMaxLength = Math.floor((maxLength - ellipsis.length - notification.length) / 2);
+
+  if (text.length > maxLength) {
+    const startLastHalf = text.length - halfMaxLength;
+    return `${text.slice(0, halfMaxLength)}${ellipsis}${text.slice(startLastHalf)}${notification}`;
+  }
+
+  return text;
+}
+
+/**
+ * @param {TMessage[]} _messages
+ * @param {number} maxContextTokens
+ * @param {function({role: string, content: TMessageContent[]}): number} getTokenCountForMessage
+ *
+ * @returns {{
+ *  dbMessages: TMessage[],
+ * editedIndices: number[]
+ * }}
+ */
+function truncateToolCallOutputs(_messages, maxContextTokens, getTokenCountForMessage) {
+  const THRESHOLD_PERCENTAGE = 0.5;
+  const targetTokenLimit = maxContextTokens * THRESHOLD_PERCENTAGE;
+
+  let currentTokenCount = 3;
+  const messages = [..._messages];
+  const processedMessages = [];
+  let currentIndex = messages.length;
+  const editedIndices = new Set();
+  while (messages.length > 0) {
+    currentIndex--;
+    const message = messages.pop();
+    currentTokenCount += message.tokenCount;
+    if (currentTokenCount < targetTokenLimit) {
+      processedMessages.push(message);
+      continue;
+    }
+
+    if (!message.content || !Array.isArray(message.content)) {
+      processedMessages.push(message);
+      continue;
+    }
+
+    const toolCallIndices = message.content
+      .map((item, index) => (item.type === 'tool_call' ? index : -1))
+      .filter((index) => index !== -1)
+      .reverse();
+
+    if (toolCallIndices.length === 0) {
+      processedMessages.push(message);
+      continue;
+    }
+
+    const newContent = [...message.content];
+
+    // Truncate all tool outputs since we're over threshold
+    for (const index of toolCallIndices) {
+      const toolCall = newContent[index].tool_call;
+      if (!toolCall || !toolCall.output) {
+        continue;
+      }
+
+      editedIndices.add(currentIndex);
+
+      newContent[index] = {
+        ...newContent[index],
+        tool_call: {
+          ...toolCall,
+          output: '[OUTPUT_OMITTED_FOR_BREVITY]',
+        },
+      };
+    }
+
+    const truncatedMessage = {
+      ...message,
+      content: newContent,
+      tokenCount: getTokenCountForMessage({ role: 'assistant', content: newContent }),
+    };
+
+    processedMessages.push(truncatedMessage);
+  }
+
+  return { dbMessages: processedMessages.reverse(), editedIndices: Array.from(editedIndices) };
+}
+
+module.exports = { truncateText, smartTruncateText, truncateToolCallOutputs };

api/app/clients/specs/BaseClient.test.js

@@ -615,9 +615,9 @@ describe('BaseClient', () => {
     test('getTokenCount for response is called with the correct arguments', async () => {
       const tokenCountMap = {}; // Mock tokenCountMap
       TestClient.buildMessages.mockReturnValue({ prompt: [], tokenCountMap });
-      TestClient.getTokenCount = jest.fn();
+      TestClient.getTokenCountForResponse = jest.fn();
       const response = await TestClient.sendMessage('Hello, world!', {});
-      expect(TestClient.getTokenCount).toHaveBeenCalledWith(response.text);
+      expect(TestClient.getTokenCountForResponse).toHaveBeenCalledWith(response);
     });
 
     test('returns an object with the correct shape', async () => {

api/app/clients/tools/util/handleOpenAIErrors.js

@@ -23,6 +23,8 @@ async function handleOpenAIErrors(err, errorCallback, context = 'stream') {
     logger.warn(`[OpenAIClient.chatCompletion][${context}] Unhandled error type`);
   }
 
+  logger.error(err);
+
   if (errorCallback) {
     errorCallback(err);
   }