@librechat/agents 2.2.1 → 2.2.3

package/dist/esm/utils/tokens.mjs

@@ -0,0 +1,62 @@
+import { Tiktoken } from 'js-tiktoken/lite';
+import { ContentTypes } from '../common/enum.mjs';
+
+function getTokenCountForMessage(message, getTokenCount) {
+    let tokensPerMessage = 3;
+    const processValue = (value) => {
+        if (Array.isArray(value)) {
+            for (let item of value) {
+                if (!item ||
+                    !item.type ||
+                    item.type === ContentTypes.ERROR ||
+                    item.type === ContentTypes.IMAGE_URL) {
+                    continue;
+                }
+                if (item.type === ContentTypes.TOOL_CALL && item.tool_call != null) {
+                    const toolName = item.tool_call?.name || '';
+                    if (toolName != null && toolName && typeof toolName === 'string') {
+                        numTokens += getTokenCount(toolName);
+                    }
+                    const args = item.tool_call?.args || '';
+                    if (args != null && args && typeof args === 'string') {
+                        numTokens += getTokenCount(args);
+                    }
+                    const output = item.tool_call?.output || '';
+                    if (output != null && output && typeof output === 'string') {
+                        numTokens += getTokenCount(output);
+                    }
+                    continue;
+                }
+                const nestedValue = item[item.type];
+                if (!nestedValue) {
+                    continue;
+                }
+                processValue(nestedValue);
+            }
+        }
+        else if (typeof value === 'string') {
+            numTokens += getTokenCount(value);
+        }
+        else if (typeof value === 'number') {
+            numTokens += getTokenCount(value.toString());
+        }
+        else if (typeof value === 'boolean') {
+            numTokens += getTokenCount(value.toString());
+        }
+    };
+    let numTokens = tokensPerMessage;
+    processValue(message.content);
+    return numTokens;
+}
+const createTokenCounter = async () => {
+    const res = await fetch(`https://tiktoken.pages.dev/js/o200k_base.json`);
+    const o200k_base = await res.json();
+    const countTokens = (text) => {
+        const enc = new Tiktoken(o200k_base);
+        return enc.encode(text).length;
+    };
+    return (message) => getTokenCountForMessage(message, countTokens);
+};
+
+export { createTokenCounter };
+//# sourceMappingURL=tokens.mjs.map

package/dist/esm/utils/tokens.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"tokens.mjs","sources":["../../../src/utils/tokens.ts"],"sourcesContent":["import { Tiktoken } from \"js-tiktoken/lite\";\nimport type { BaseMessage } from \"@langchain/core/messages\";\nimport { ContentTypes } from \"@/common/enum\";\n\nfunction getTokenCountForMessage(message: BaseMessage, getTokenCount: (text: string) => number): number {\n  let tokensPerMessage = 3;\n\n  const processValue = (value: unknown) => {\n    if (Array.isArray(value)) {\n      for (let item of value) {\n        if (\n          !item ||\n          !item.type ||\n          item.type === ContentTypes.ERROR ||\n          item.type === ContentTypes.IMAGE_URL\n        ) {\n          continue;\n        }\n\n        if (item.type === ContentTypes.TOOL_CALL && item.tool_call != null) {\n          const toolName = item.tool_call?.name || '';\n          if (toolName != null && toolName && typeof toolName === 'string') {\n            numTokens += getTokenCount(toolName);\n          }\n\n          const args = item.tool_call?.args || '';\n          if (args != null && args && typeof args === 'string') {\n            numTokens += getTokenCount(args);\n          }\n\n          const output = item.tool_call?.output || '';\n          if (output != null && output && typeof output === 'string') {\n            numTokens += getTokenCount(output);\n          }\n          continue;\n        }\n\n        const nestedValue = item[item.type];\n\n        if (!nestedValue) {\n          continue;\n        }\n\n        processValue(nestedValue);\n      }\n    } else if (typeof value === 'string') {\n      numTokens += getTokenCount(value);\n    } else if (typeof value === 'number') {\n      numTokens += getTokenCount(value.toString());\n    } else if (typeof value === 'boolean') {\n      numTokens += getTokenCount(value.toString());\n    }\n  };\n\n  let numTokens = tokensPerMessage;\n  processValue(message.content);\n  return numTokens;\n}\n\nexport const createTokenCounter = async () => {\n  const res = await fetch(`https://tiktoken.pages.dev/js/o200k_base.json`);\n  const o200k_base = await res.json();\n\n  const countTokens = (text: string) => {\n    const enc = new Tiktoken(o200k_base);\n    return enc.encode(text).length;\n  }\n  \n  return (message: BaseMessage) => getTokenCountForMessage(message, countTokens);\n}"],"names":[],"mappings":";;;AAIA,SAAS,uBAAuB,CAAC,OAAoB,EAAE,aAAuC,EAAA;IAC5F,IAAI,gBAAgB,GAAG,CAAC;AAExB,IAAA,MAAM,YAAY,GAAG,CAAC,KAAc,KAAI;AACtC,QAAA,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,EAAE;AACxB,YAAA,KAAK,IAAI,IAAI,IAAI,KAAK,EAAE;AACtB,gBAAA,IACE,CAAC,IAAI;oBACL,CAAC,IAAI,CAAC,IAAI;AACV,oBAAA,IAAI,CAAC,IAAI,KAAK,YAAY,CAAC,KAAK;AAChC,oBAAA,IAAI,CAAC,IAAI,KAAK,YAAY,CAAC,SAAS,EACpC;oBACA;;AAGF,gBAAA,IAAI,IAAI,CAAC,IAAI,KAAK,YAAY,CAAC,SAAS,IAAI,IAAI,CAAC,SAAS,IAAI,IAAI,EAAE;oBAClE,MAAM,QAAQ,GAAG,IAAI,CAAC,SAAS,EAAE,IAAI,IAAI,EAAE;oBAC3C,IAAI,QAAQ,IAAI,IAAI,IAAI,QAAQ,IAAI,OAAO,QAAQ,KAAK,QAAQ,EAAE;AAChE,wBAAA,SAAS,IAAI,aAAa,CAAC,QAAQ,CAAC;;oBAGtC,MAAM,IAAI,GAAG,IAAI,CAAC,SAAS,EAAE,IAAI,IAAI,EAAE;oBACvC,IAAI,IAAI,IAAI,IAAI,IAAI,IAAI,IAAI,OAAO,IAAI,KAAK,QAAQ,EAAE;AACpD,wBAAA,SAAS,IAAI,aAAa,CAAC,IAAI,CAAC;;oBAGlC,MAAM,MAAM,GAAG,IAAI,CAAC,SAAS,EAAE,MAAM,IAAI,EAAE;oBAC3C,IAAI,MAAM,IAAI,IAAI,IAAI,MAAM,IAAI,OAAO,MAAM,KAAK,QAAQ,EAAE;AAC1D,wBAAA,SAAS,IAAI,aAAa,CAAC,MAAM,CAAC;;oBAEpC;;gBAGF,MAAM,WAAW,GAAG,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC;gBAEnC,IAAI,CAAC,WAAW,EAAE;oBAChB;;gBAGF,YAAY,CAAC,WAAW,CAAC;;;AAEtB,aAAA,IAAI,OAAO,KAAK,KAAK,QAAQ,EAAE;AACpC,YAAA,SAAS,IAAI,aAAa,CAAC,KAAK,CAAC;;AAC5B,aAAA,IAAI,OAAO,KAAK,KAAK,QAAQ,EAAE;YACpC,SAAS,IAAI,aAAa,CAAC,KAAK,CAAC,QAAQ,EAAE,CAAC;;AACvC,aAAA,IAAI,OAAO,KAAK,KAAK,SAAS,EAAE;YACrC,SAAS,IAAI,aAAa,CAAC,KAAK,CAAC,QAAQ,EAAE,CAAC;;AAEhD,KAAC;IAED,IAAI,SAAS,GAAG,gBAAgB;AAChC,IAAA,YAAY,CAAC,OAAO,CAAC,OAAO,CAAC;AAC7B,IAAA,OAAO,SAAS;AAClB;AAEa,MAAA,kBAAkB,GAAG,YAAW;AAC3C,IAAA,MAAM,GAAG,GAAG,MAAM,KAAK,CAAC,CAAA,6CAAA,CAA+C,CAAC;AACxE,IAAA,MAAM,UAAU,GAAG,MAAM,GAAG,CAAC,IAAI,EAAE;AAEnC,IAAA,MAAM,WAAW,GAAG,CAAC,IAAY,KAAI;AACnC,QAAA,MAAM,GAAG,GAAG,IAAI,QAAQ,CAAC,UAAU,CAAC;QACpC,OAAO,GAAG,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,MAAM;AAChC,KAAC;IAED,OAAO,CAAC,OAAoB,KAAK,uBAAuB,CAAC,OAAO,EAAE,WAAW,CAAC;AAChF;;;;"}

package/dist/types/graphs/Graph.d.ts

@@ -2,10 +2,11 @@ import { ToolNode } from '@langchain/langgraph/prebuilt';
 import { START } from '@langchain/langgraph';
 import { Runnable, RunnableConfig } from '@langchain/core/runnables';
 import { SystemMessage } from '@langchain/core/messages';
-import type { BaseMessage } from '@langchain/core/messages';
+import type { BaseMessage, UsageMetadata } from '@langchain/core/messages';
 import type * as t from '@/types';
 import { Providers, GraphNodeKeys, Callback, ContentTypes } from '@/common';
 import { ToolNode as CustomToolNode } from '@/tools/ToolNode';
+import { createPruneMessages } from '@/messages';
 import { HandlerRegistry } from '@/events';
 export type GraphNode = GraphNodeKeys | typeof START;
 export type ClientCallback<T extends unknown[]> = (graph: StandardGraph, ...args: T) => void;
@@ -49,8 +50,13 @@ export declare abstract class Graph<T extends t.BaseGraphState = t.BaseGraphStat
     stepKeyIds: Map<string, string[]>;
     contentIndexMap: Map<string, number>;
     toolCallStepIds: Map<string, string>;
+    currentUsage: Partial<UsageMetadata> | undefined;
+    indexTokenCountMap: Record<string, number>;
+    maxContextTokens: number | undefined;
+    pruneMessages?: ReturnType<typeof createPruneMessages>;
     /** The amount of time that should pass before another consecutive API call */
     streamBuffer: number | undefined;
+    tokenCounter?: t.TokenCounter;
     signal?: AbortSignal;
 }
 export declare class StandardGraph extends Graph<t.BaseGraphState, GraphNode> {
@@ -87,6 +93,7 @@ export declare class StandardGraph extends Graph<t.BaseGraphState, GraphNode> {
         clientOptions?: t.ClientOptions;
         omitOriginalOptions?: string[];
     }): t.ChatModelInstance;
+    storeUsageMetadata(finalMessage?: BaseMessage): void;
     createCallModel(): (state: t.BaseGraphState, config?: RunnableConfig) => Promise<Partial<t.BaseGraphState>>;
     createWorkflow(): t.CompiledWorkflow<t.BaseGraphState>;
     /**

package/dist/types/messages/format.d.ts

@@ -0,0 +1,120 @@
+import { ToolMessage, BaseMessage } from '@langchain/core/messages';
+import { HumanMessage, AIMessage, SystemMessage } from '@langchain/core/messages';
+import { MessageContentImageUrl } from '@langchain/core/messages';
+import type { MessageContentComplex } from '@/types';
+import { Providers, ContentTypes } from '@/common';
+interface VisionMessageParams {
+    message: {
+        role: string;
+        content: string;
+        name?: string;
+        [key: string]: any;
+    };
+    image_urls: MessageContentImageUrl[];
+    endpoint?: Providers;
+}
+/**
+ * Formats a message to OpenAI Vision API payload format.
+ *
+ * @param {VisionMessageParams} params - The parameters for formatting.
+ * @returns {Object} - The formatted message.
+ */
+export declare const formatVisionMessage: ({ message, image_urls, endpoint }: VisionMessageParams) => {
+    role: string;
+    content: MessageContentComplex[];
+    name?: string;
+    [key: string]: any;
+};
+interface MessageInput {
+    role?: string;
+    _name?: string;
+    sender?: string;
+    text?: string;
+    content?: string | MessageContentComplex[];
+    image_urls?: MessageContentImageUrl[];
+    lc_id?: string[];
+    [key: string]: any;
+}
+interface FormatMessageParams {
+    message: MessageInput;
+    userName?: string;
+    assistantName?: string;
+    endpoint?: Providers;
+    langChain?: boolean;
+}
+interface FormattedMessage {
+    role: string;
+    content: string | MessageContentComplex[];
+    name?: string;
+    [key: string]: any;
+}
+/**
+ * Formats a message to OpenAI payload format based on the provided options.
+ *
+ * @param {FormatMessageParams} params - The parameters for formatting.
+ * @returns {FormattedMessage | HumanMessage | AIMessage | SystemMessage} - The formatted message.
+ */
+export declare const formatMessage: ({ message, userName, assistantName, endpoint, langChain }: FormatMessageParams) => FormattedMessage | HumanMessage | AIMessage | SystemMessage;
+/**
+ * Formats an array of messages for LangChain.
+ *
+ * @param {Array<MessageInput>} messages - The array of messages to format.
+ * @param {Omit<FormatMessageParams, 'message' | 'langChain'>} formatOptions - The options for formatting each message.
+ * @returns {Array<HumanMessage | AIMessage | SystemMessage>} - The array of formatted LangChain messages.
+ */
+export declare const formatLangChainMessages: (messages: Array<MessageInput>, formatOptions: Omit<FormatMessageParams, "message" | "langChain">) => Array<HumanMessage | AIMessage | SystemMessage>;
+interface LangChainMessage {
+    lc_kwargs?: {
+        additional_kwargs?: Record<string, any>;
+        [key: string]: any;
+    };
+    kwargs?: {
+        additional_kwargs?: Record<string, any>;
+        [key: string]: any;
+    };
+    [key: string]: any;
+}
+/**
+ * Formats a LangChain message object by merging properties from `lc_kwargs` or `kwargs` and `additional_kwargs`.
+ *
+ * @param {LangChainMessage} message - The message object to format.
+ * @returns {Record<string, any>} The formatted LangChain message.
+ */
+export declare const formatFromLangChain: (message: LangChainMessage) => Record<string, any>;
+interface TMessage {
+    role?: string;
+    content?: string | Array<{
+        type: ContentTypes;
+        text?: string;
+        tool_call_ids?: string[];
+        [key: string]: any;
+    }>;
+    [key: string]: any;
+}
+/**
+ * Formats an array of messages for LangChain, handling tool calls and creating ToolMessage instances.
+ *
+ * @param {Array<Partial<TMessage>>} payload - The array of messages to format.
+ * @param {Record<number, number>} [indexTokenCountMap] - Optional map of message indices to token counts.
+ * @returns {Object} - Object containing formatted messages and updated indexTokenCountMap if provided.
+ */
+export declare const formatAgentMessages: (payload: Array<Partial<TMessage>>, indexTokenCountMap?: Record<number, number>) => {
+    messages: Array<HumanMessage | AIMessage | SystemMessage | ToolMessage>;
+    indexTokenCountMap?: Record<number, number>;
+};
+/**
+ * Formats an array of messages for LangChain, making sure all content fields are strings
+ * @param {Array<HumanMessage | AIMessage | SystemMessage | ToolMessage>} payload - The array of messages to format.
+ * @returns {Array<HumanMessage | AIMessage | SystemMessage | ToolMessage>} - The array of formatted LangChain messages, including ToolMessages for tool calls.
+ */
+export declare const formatContentStrings: (payload: Array<BaseMessage>) => Array<BaseMessage>;
+/**
+ * Adds a value at key 0 for system messages and shifts all key indices by one in an indexTokenCountMap.
+ * This is useful when adding a system message at the beginning of a conversation.
+ *
+ * @param indexTokenCountMap - The original map of message indices to token counts
+ * @param instructionsTokenCount - The token count for the system message to add at index 0
+ * @returns A new map with the system message at index 0 and all other indices shifted by 1
+ */
+export declare function shiftIndexTokenCountMap(indexTokenCountMap: Record<number, number>, instructionsTokenCount: number): Record<number, number>;
+export {};

package/dist/types/messages/index.d.ts

@@ -0,0 +1,3 @@
+export * from './core';
+export * from './prune';
+export * from './format';

package/dist/types/messages/prune.d.ts

@@ -0,0 +1,16 @@
+import type { BaseMessage, UsageMetadata } from '@langchain/core/messages';
+import type { TokenCounter } from '@/types/run';
+export type PruneMessagesFactoryParams = {
+    maxTokens: number;
+    startIndex: number;
+    tokenCounter: TokenCounter;
+    indexTokenCountMap: Record<string, number>;
+};
+export type PruneMessagesParams = {
+    messages: BaseMessage[];
+    usageMetadata?: Partial<UsageMetadata>;
+};
+export declare function createPruneMessages(factoryParams: PruneMessagesFactoryParams): (params: PruneMessagesParams) => {
+    context: BaseMessage[];
+    indexTokenCountMap: Record<string, number>;
+};

package/dist/types/types/run.d.ts

@@ -50,7 +50,11 @@ export type RunConfig = {
     returnContent?: boolean;
 };
 export type ProvidedCallbacks = (BaseCallbackHandler | CallbackHandlerMethods)[] | undefined;
+export type TokenCounter = (message: BaseMessage) => number;
 export type EventStreamOptions = {
     callbacks?: graph.ClientCallbacks;
     keepContent?: boolean;
+    maxContextTokens?: number;
+    tokenCounter?: TokenCounter;
+    indexTokenCountMap?: Record<string, number>;
 };

package/dist/types/utils/tokens.d.ts

@@ -0,0 +1,2 @@
+import type { BaseMessage } from "@langchain/core/messages";
+export declare const createTokenCounter: () => Promise<(message: BaseMessage) => number>;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@librechat/agents",
-  "version": "2.2.1",
+  "version": "2.2.3",
   "main": "./dist/cjs/main.cjs",
   "module": "./dist/esm/main.mjs",
   "types": "./dist/types/index.d.ts",

package/src/graphs/Graph.ts

@@ -8,12 +8,13 @@ import { ChatOpenAI, AzureChatOpenAI } from '@langchain/openai';
 import { Runnable, RunnableConfig } from '@langchain/core/runnables';
 import { dispatchCustomEvent } from '@langchain/core/callbacks/dispatch';
 import { AIMessageChunk, ToolMessage, SystemMessage } from '@langchain/core/messages';
-import type { BaseMessage, BaseMessageFields } from '@langchain/core/messages';
+import type { BaseMessage, BaseMessageFields, UsageMetadata } from '@langchain/core/messages';
 import type * as t from '@/types';
 import { Providers, GraphEvents, GraphNodeKeys, StepTypes, Callback, ContentTypes } from '@/common';
 import { getChatModelClass, manualToolStreamProviders } from '@/llm/providers';
 import { ToolNode as CustomToolNode, toolsCondition } from '@/tools/ToolNode';
 import {
+  createPruneMessages,
   modifyDeltaProperties,
   formatArtifactPayload,
   convertMessagesToContent,
@@ -74,8 +75,13 @@ export abstract class Graph<
   stepKeyIds: Map<string, string[]> = new Map<string, string[]>();
   contentIndexMap: Map<string, number> = new Map();
   toolCallStepIds: Map<string, string> = new Map();
+  currentUsage: Partial<UsageMetadata> | undefined;
+  indexTokenCountMap: Record<string, number> = {};
+  maxContextTokens: number | undefined;
+  pruneMessages?: ReturnType<typeof createPruneMessages>;
   /** The amount of time that should pass before another consecutive API call */
   streamBuffer: number | undefined;
+  tokenCounter?: t.TokenCounter;
   signal?: AbortSignal;
 }
 
@@ -166,6 +172,10 @@ export class StandardGraph extends Graph<
     this.currentTokenType = resetIfNotEmpty(this.currentTokenType, ContentTypes.TEXT);
     this.lastToken = resetIfNotEmpty(this.lastToken, undefined);
     this.tokenTypeSwitch = resetIfNotEmpty(this.tokenTypeSwitch, undefined);
+    this.indexTokenCountMap = resetIfNotEmpty(this.indexTokenCountMap, {});
+    this.currentUsage = resetIfNotEmpty(this.currentUsage, undefined);
+    this.tokenCounter = resetIfNotEmpty(this.tokenCounter, undefined);
+    this.maxContextTokens = resetIfNotEmpty(this.maxContextTokens, undefined);
   }
 
   /* Run Step Processing */
@@ -326,6 +336,12 @@ export class StandardGraph extends Graph<
     return new ChatModelClass(options);
   }
 
+  storeUsageMetadata(finalMessage?: BaseMessage): void {
+    if (finalMessage && 'usage_metadata' in finalMessage && finalMessage.usage_metadata) {
+      this.currentUsage = finalMessage.usage_metadata as Partial<UsageMetadata>;
+    }
+  }
+
   createCallModel() {
     return async (state: t.BaseGraphState, config?: RunnableConfig): Promise<Partial<t.BaseGraphState>> => {
       const { provider = '' } = (config?.configurable as t.GraphConfig | undefined) ?? {} ;
@@ -338,9 +354,27 @@ export class StandardGraph extends Graph<
       this.config = config;
       const { messages } = state;
 
-      const finalMessages = messages;
-      const lastMessageX = finalMessages[finalMessages.length - 2];
-      const lastMessageY = finalMessages[finalMessages.length - 1];
+      let messagesToUse = messages;
+      if (!this.pruneMessages && this.tokenCounter && this.maxContextTokens && this.indexTokenCountMap[0] != null) {
+        this.pruneMessages = createPruneMessages({
+          indexTokenCountMap: this.indexTokenCountMap,
+          maxTokens: this.maxContextTokens,
+          tokenCounter: this.tokenCounter,
+          startIndex: this.startIndex,
+        });
+      }
+      if (this.pruneMessages) {
+        const { context, indexTokenCountMap } = this.pruneMessages({
+          messages,
+          usageMetadata: this.currentUsage,
+        });
+        this.indexTokenCountMap = indexTokenCountMap;
+        messagesToUse = context;
+      }
+
+      const finalMessages = messagesToUse;
+      const lastMessageX = finalMessages.length >= 2 ? finalMessages[finalMessages.length - 2] : null;
+      const lastMessageY = finalMessages.length >= 1 ? finalMessages[finalMessages.length - 1] : null;
 
       if (
         provider === Providers.BEDROCK
@@ -372,6 +406,7 @@ export class StandardGraph extends Graph<
 
       this.lastStreamCall = Date.now();
 
+      let result: Partial<t.BaseGraphState>;
       if ((this.tools?.length ?? 0) > 0 && manualToolStreamProviders.has(provider)) {
         const stream = await this.boundModel.stream(finalMessages, config);
         let finalChunk: AIMessageChunk | undefined;
@@ -385,19 +420,22 @@ export class StandardGraph extends Graph<
         }
 
         finalChunk = modifyDeltaProperties(this.provider, finalChunk);
-        return { messages: [finalChunk as AIMessageChunk] };
-      }
-
-      const finalMessage = (await this.boundModel.invoke(finalMessages, config)) as AIMessageChunk;
-      if ((finalMessage.tool_calls?.length ?? 0) > 0) {
-        finalMessage.tool_calls = finalMessage.tool_calls?.filter((tool_call) => {
-          if (!tool_call.name) {
-            return false;
-          }
-          return true;
-        });
+        result = { messages: [finalChunk as AIMessageChunk] };
+      } else {
+        const finalMessage = (await this.boundModel.invoke(finalMessages, config)) as AIMessageChunk;
+        if ((finalMessage.tool_calls?.length ?? 0) > 0) {
+          finalMessage.tool_calls = finalMessage.tool_calls?.filter((tool_call) => {
+            if (!tool_call.name) {
+              return false;
+            }
+            return true;
+          });
+        }
+        result = { messages: [finalMessage] };
       }
-      return { messages: [finalMessage] };
+      
+      this.storeUsageMetadata(result?.messages?.[0]);
+      return result;
     };
   }