@librechat/agents 3.3.13 → 3.4.0

package/dist/cjs/llm/bedrock/index.cjs

@@ -1,5 +1,6 @@
 const require_cache = require("../../messages/cache.cjs");
 const require_streamLimits = require("../streamLimits.cjs");
+const require_smoother = require("../stream/smoother.cjs");
 const require_message_inputs = require("./utils/message_inputs.cjs");
 const require_message_outputs = require("./utils/message_outputs.cjs");
 require("./utils/index.cjs");
@@ -32,33 +33,6 @@ let _aws_sdk_client_bedrock_runtime = require("@aws-sdk/client-bedrock-runtime")
 * are promoted from strings to array form with `index` so they merge correctly once
 * the accumulated content is already an array.
 */
-const MAX_STREAM_QUEUE_CHUNKS = 256;
-const MAX_STREAM_QUEUE_TEXT_CHARS = 8192;
-const STREAM_CHUNK_MIN_SIZE = 4;
-const STREAM_BOUNDARIES = new Set([
-	" ",
-	".",
-	",",
-	"!",
-	"?",
-	";",
-	":"
-]);
-function findStreamChunkBoundary(text, minSize) {
-	if (minSize >= text.length) return text.length;
-	for (let position = minSize; position < text.length; position++) if (STREAM_BOUNDARIES.has(text[position])) return position + 1;
-	return text.length;
-}
-function splitStreamToken(text) {
-	const chunks = [];
-	let currentIndex = 0;
-	while (currentIndex < text.length) {
-		const chunkSize = findStreamChunkBoundary(text.slice(currentIndex), STREAM_CHUNK_MIN_SIZE);
-		chunks.push(text.slice(currentIndex, currentIndex + chunkSize));
-		currentIndex += chunkSize;
-	}
-	return chunks;
-}
 /**
 * Resolves the text a delta contributes to the smoothing cadence, preferring a
 * text delta over a reasoning delta and ignoring non-string payloads.
@@ -68,30 +42,6 @@ function resolveVisibleText(text, reasoningText) {
 	if (typeof reasoningText === "string") return reasoningText;
 	return "";
 }
-function getCadencedStreamDelay({ targetDelay, lastVisibleContentAt, now }) {
-	if (targetDelay <= 0 || lastVisibleContentAt == null) return 0;
-	return Math.max(0, targetDelay - (now - lastVisibleContentAt));
-}
-async function waitForStreamDelay(delay, signal) {
-	if (delay <= 0 || isSignalAborted(signal)) return;
-	await new Promise((resolve) => {
-		const timeoutRef = {};
-		const onAbort = () => {
-			if (timeoutRef.current) clearTimeout(timeoutRef.current);
-			signal?.removeEventListener("abort", onAbort);
-			resolve();
-		};
-		timeoutRef.current = setTimeout(() => {
-			signal?.removeEventListener("abort", onAbort);
-			resolve();
-		}, delay);
-		signal?.addEventListener("abort", onAbort, { once: true });
-		if (isSignalAborted(signal)) onAbort();
-	});
-}
-function isSignalAborted(signal) {
-	return signal?.aborted === true;
-}
 var CustomChatBedrockConverse = class extends _langchain_aws.ChatBedrockConverse {
 	_lc_stream_delay;
 	/**
@@ -121,7 +71,7 @@ var CustomChatBedrockConverse = class extends _langchain_aws.ChatBedrockConverse
 		super(fields);
 		this.promptCache = fields?.promptCache;
 		this.promptCacheTtl = fields?.promptCacheTtl;
-		this._lc_stream_delay = Math.max(0, fields?._lc_stream_delay ?? 0);
+		this._lc_stream_delay = require_smoother.resolveStreamDelay(fields?._lc_stream_delay);
 		this.applicationInferenceProfile = fields?.applicationInferenceProfile;
 		this.serviceTier = fields?.serviceTier;
 		this.cacheModelId = fields?.model ?? this.model;
@@ -192,21 +142,12 @@ var CustomChatBedrockConverse = class extends _langchain_aws.ChatBedrockConverse
 		const streamAbortController = new AbortController();
 		const abortStream = () => streamAbortController.abort();
 		options.signal?.addEventListener("abort", abortStream, { once: true });
-		if (isSignalAborted(options.signal)) abortStream();
+		if (require_smoother.isSignalAborted(options.signal)) abortStream();
 		try {
 			const stream = (await this.client.send(command, { abortSignal: streamAbortController.signal })).stream;
 			if (!stream) return;
 			const seenBlockIndices = /* @__PURE__ */ new Set();
 			const toolUseBlockIndices = /* @__PURE__ */ new Set();
-			const queuedChunks = [];
-			const producerState = { done: false };
-			let queuedChunkIndex = 0;
-			let bufferedTextLength = 0;
-			let consumerClosed = false;
-			let notifyConsumer;
-			let notifyProducer;
-			let hasEmittedVisibleContent = false;
-			let lastVisibleContentAt;
 			/**
 			* Guardrails can reject an already-streamed toolUse block at
 			* `messageStop` (`guardrail_intervened`), after `contentBlockStop` has
@@ -214,202 +155,141 @@ var CustomChatBedrockConverse = class extends _langchain_aws.ChatBedrockConverse
 			* configured, so a later intervention can't race an eagerly started tool.
 			*/
 			const sealToolUseOnStop = options.guardrailConfig == null && this.guardrailConfig == null;
-			const notifyConsumerForChunk = () => {
-				notifyConsumer?.();
-				notifyConsumer = void 0;
-			};
-			const notifyProducerForSpace = () => {
-				notifyProducer?.();
-				notifyProducer = void 0;
-			};
-			const hasQueuedChunks = () => queuedChunkIndex < queuedChunks.length;
-			const getQueuedChunkCount = () => queuedChunks.length - queuedChunkIndex;
-			const isQueueAtCapacity = () => getQueuedChunkCount() >= MAX_STREAM_QUEUE_CHUNKS || bufferedTextLength >= MAX_STREAM_QUEUE_TEXT_CHARS;
-			const waitForNextChunk = async () => {
-				if (hasQueuedChunks() || producerState.done || producerState.error != null) return;
-				await new Promise((resolve) => {
-					notifyConsumer = resolve;
-				});
-			};
-			const waitForQueueSpace = async () => {
-				while (isQueueAtCapacity() && !consumerClosed && !isSignalAborted(options.signal)) await new Promise((resolve) => {
-					const signal = options.signal;
-					const onAbort = () => {
-						signal?.removeEventListener("abort", onAbort);
-						resolve();
-					};
-					const onSpace = () => {
-						signal?.removeEventListener("abort", onAbort);
-						resolve();
-					};
-					notifyProducer = onSpace;
-					signal?.addEventListener("abort", onAbort, { once: true });
-					if (isSignalAborted(signal)) onAbort();
-				});
-			};
-			const dequeue = () => {
-				if (!hasQueuedChunks()) return;
-				const queuedChunk = queuedChunks[queuedChunkIndex];
-				queuedChunkIndex++;
-				if (queuedChunkIndex > 128 && queuedChunkIndex * 2 >= queuedChunks.length) {
-					queuedChunks.splice(0, queuedChunkIndex);
-					queuedChunkIndex = 0;
-				}
-				return queuedChunk;
-			};
-			const enqueue = async (queuedChunk) => {
-				await waitForQueueSpace();
-				if (consumerClosed || isSignalAborted(options.signal)) {
-					abortStream();
-					throw new Error("AbortError: User aborted the request.");
-				}
-				queuedChunks.push(queuedChunk);
-				if (queuedChunk.smooth) bufferedTextLength += queuedChunk.textLength;
-				notifyConsumerForChunk();
-			};
-			const enqueueChunk = async ({ chunk, callbackChunk, callbackToken = "", smooth = false, textLength = 0 }) => {
-				await enqueue({
-					chunk,
-					callbackChunk,
-					callbackToken,
-					smooth,
-					textLength: smooth ? textLength : 0
-				});
-			};
-			const enqueueDelta = async (contentBlockDelta) => {
+			/**
+			* Builds the emission for one piece of a delta, reproducing the exact
+			* per-piece pipeline (sliced delta → chunk → enrichment → stream-limit
+			* link). `indicesSnapshot` is the arrival-time copy of the seen block
+			* indices, so lazily emitted pieces observe the same enrichment
+			* decisions the delta saw when it arrived.
+			*/
+			const buildDeltaEmission = (contentBlockDelta, token, indicesSnapshot) => {
 				const delta = contentBlockDelta.delta;
-				if (delta == null) throw new Error("No delta found in content block.");
-				const idx = contentBlockDelta.contentBlockIndex;
-				if (idx != null) seenBlockIndices.add(idx);
-				const text = delta.text;
-				const reasoningContent = delta.reasoningContent;
+				const text = delta?.text;
+				const reasoningContent = delta?.reasoningContent;
 				const reasoningText = reasoningContent?.text;
-				const visibleText = resolveVisibleText(text, reasoningText);
-				const smooth = this._lc_stream_delay > 0 && visibleText !== "";
-				const tokenChunks = smooth ? splitStreamToken(visibleText) : [visibleText];
-				for (const token of tokenChunks) {
-					let splitDelta = contentBlockDelta;
-					if (typeof text === "string") splitDelta = {
-						...contentBlockDelta,
-						delta: { text: token }
-					};
-					else if (typeof reasoningText === "string" && reasoningContent != null) splitDelta = {
-						...contentBlockDelta,
-						delta: { reasoningContent: {
-							...reasoningContent,
-							text: token
-						} }
-					};
-					const deltaChunk = require_message_outputs.handleConverseStreamContentBlockDelta(splitDelta);
-					const enrichedChunk = this.enrichChunk(deltaChunk, seenBlockIndices);
-					if (enrichedChunk !== deltaChunk)
+				let splitDelta = contentBlockDelta;
+				if (typeof text === "string") splitDelta = {
+					...contentBlockDelta,
+					delta: { text: token }
+				};
+				else if (typeof reasoningText === "string" && reasoningContent != null) splitDelta = {
+					...contentBlockDelta,
+					delta: { reasoningContent: {
+						...reasoningContent,
+						text: token
+					} }
+				};
+				const deltaChunk = require_message_outputs.handleConverseStreamContentBlockDelta(splitDelta);
+				const enrichedChunk = this.enrichChunk(deltaChunk, indicesSnapshot);
+				if (enrichedChunk !== deltaChunk)
  /** The callback copy is the same emission as the enriched yield;
-					* without the link, stream-limit accounting charges both message
-					* objects and Bedrock tool arguments falsely trip near half the
-					* cap. Linked on the messages, which are what the accounting
-					* observes. */
-					require_streamLimits.linkStreamLimitCanonical(deltaChunk.message, enrichedChunk.message);
-					await enqueueChunk({
-						chunk: enrichedChunk,
-						callbackChunk: deltaChunk,
-						callbackToken: deltaChunk.text,
-						smooth,
-						textLength: token.length
-					});
-				}
+				* without the link, stream-limit accounting charges both message
+				* objects and Bedrock tool arguments falsely trip near half the
+				* cap. Linked on the messages, which are what the accounting
+				* observes. */
+				require_streamLimits.linkStreamLimitCanonical(deltaChunk.message, enrichedChunk.message);
+				return {
+					chunk: enrichedChunk,
+					callbackChunk: deltaChunk,
+					callbackToken: deltaChunk.text
+				};
 			};
-			const producer = (async () => {
-				try {
-					for await (const event of stream) {
-						if (isSignalAborted(options.signal)) {
-							abortStream();
-							throw new Error("AbortError: User aborted the request.");
-						}
-						if (event.contentBlockStart != null) {
-							const startChunk = require_message_outputs.handleConverseStreamContentBlockStart(event.contentBlockStart);
-							if (startChunk != null) {
-								const idx = event.contentBlockStart.contentBlockIndex;
-								if (idx != null) {
-									seenBlockIndices.add(idx);
-									if (event.contentBlockStart.start?.toolUse != null) toolUseBlockIndices.add(idx);
-								}
-								const enrichedStart = this.enrichChunk(startChunk, seenBlockIndices);
-								if (enrichedStart !== startChunk) require_streamLimits.linkStreamLimitCanonical(startChunk.message, enrichedStart.message);
-								await enqueueChunk({
+			const enrichChunk = this.enrichChunk.bind(this);
+			const smoothed = require_smoother.smoothStream({
+				source: (async function* () {
+					for await (const event of stream) if (event.contentBlockStart != null) {
+						const startChunk = require_message_outputs.handleConverseStreamContentBlockStart(event.contentBlockStart);
+						if (startChunk != null) {
+							const idx = event.contentBlockStart.contentBlockIndex;
+							if (idx != null) {
+								seenBlockIndices.add(idx);
+								if (event.contentBlockStart.start?.toolUse != null) toolUseBlockIndices.add(idx);
+							}
+							const enrichedStart = enrichChunk(startChunk, seenBlockIndices);
+							if (enrichedStart !== startChunk) require_streamLimits.linkStreamLimitCanonical(startChunk.message, enrichedStart.message);
+							yield {
+								text: "",
+								smooth: false,
+								emit: () => ({
 									chunk: enrichedStart,
 									callbackChunk: startChunk,
 									callbackToken: startChunk.text
-								});
-							}
-						} else if (event.contentBlockDelta != null) await enqueueDelta(event.contentBlockDelta);
-						else if (event.metadata != null) await enqueueChunk({ chunk: require_message_outputs.handleConverseStreamMetadata(event.metadata, { streamUsage }) });
-						else if (event.contentBlockStop != null) {
-							const stopIdx = event.contentBlockStop.contentBlockIndex;
-							if (stopIdx != null) {
-								seenBlockIndices.add(stopIdx);
-								if (sealToolUseOnStop && toolUseBlockIndices.has(stopIdx)) {
-									const sealChunk = require_message_outputs.createConverseToolUseStopChunk(stopIdx);
-									await enqueueChunk({
+								})
+							};
+						}
+					} else if (event.contentBlockDelta != null) {
+						const contentBlockDelta = event.contentBlockDelta;
+						const delta = contentBlockDelta.delta;
+						if (delta == null) throw new Error("No delta found in content block.");
+						const idx = contentBlockDelta.contentBlockIndex;
+						if (idx != null) seenBlockIndices.add(idx);
+						const visibleText = resolveVisibleText(delta.text, delta.reasoningContent?.text);
+						const indicesSnapshot = new Set(seenBlockIndices);
+						if (visibleText === "") {
+							yield {
+								text: "",
+								smooth: false,
+								emit: () => buildDeltaEmission(contentBlockDelta, visibleText, indicesSnapshot)
+							};
+							continue;
+						}
+						yield {
+							text: visibleText,
+							smooth: true,
+							emit: (piece) => buildDeltaEmission(contentBlockDelta, piece.text, indicesSnapshot)
+						};
+					} else if (event.metadata != null) {
+						const metadataChunk = require_message_outputs.handleConverseStreamMetadata(event.metadata, { streamUsage });
+						yield {
+							text: "",
+							smooth: false,
+							emit: () => ({
+								chunk: metadataChunk,
+								callbackToken: ""
+							})
+						};
+					} else if (event.contentBlockStop != null) {
+						const stopIdx = event.contentBlockStop.contentBlockIndex;
+						if (stopIdx != null) {
+							seenBlockIndices.add(stopIdx);
+							if (sealToolUseOnStop && toolUseBlockIndices.has(stopIdx)) {
+								const sealChunk = require_message_outputs.createConverseToolUseStopChunk(stopIdx);
+								yield {
+									text: "",
+									smooth: false,
+									emit: () => ({
 										chunk: sealChunk,
 										callbackChunk: sealChunk,
 										callbackToken: sealChunk.text
-									});
-								}
+									})
+								};
 							}
-						} else await enqueueChunk({ chunk: new _langchain_core_outputs.ChatGenerationChunk({
+						}
+					} else {
+						const eventChunk = new _langchain_core_outputs.ChatGenerationChunk({
 							text: "",
 							message: new _langchain_core_messages.AIMessageChunk({
 								content: "",
 								response_metadata: { ...event }
 							})
-						}) });
-					}
-				} catch (error) {
-					producerState.error = error;
-				} finally {
-					producerState.done = true;
-					notifyConsumerForChunk();
-				}
-			})();
-			try {
-				let keepStreaming = true;
-				while (keepStreaming) {
-					if (isSignalAborted(options.signal)) {
-						abortStream();
-						throw new Error("AbortError: User aborted the request.");
-					}
-					await waitForNextChunk();
-					const queuedChunk = dequeue();
-					if (!queuedChunk) {
-						if (producerState.error != null) throw producerState.error;
-						if (producerState.done) keepStreaming = false;
-						continue;
+						});
+						yield {
+							text: "",
+							smooth: false,
+							emit: () => ({
+								chunk: eventChunk,
+								callbackToken: ""
+							})
+						};
 					}
-					if (queuedChunk.smooth) {
-						bufferedTextLength = Math.max(0, bufferedTextLength - queuedChunk.textLength);
-						notifyProducerForSpace();
-						await waitForStreamDelay(getCadencedStreamDelay({
-							targetDelay: hasEmittedVisibleContent ? this._lc_stream_delay : 0,
-							lastVisibleContentAt,
-							now: Date.now()
-						}), options.signal);
-						if (isSignalAborted(options.signal)) {
-							abortStream();
-							throw new Error("AbortError: User aborted the request.");
-						}
-						hasEmittedVisibleContent = true;
-						lastVisibleContentAt = Date.now();
-					} else notifyProducerForSpace();
-					yield queuedChunk.chunk;
-					if (queuedChunk.callbackChunk != null) await runManager?.handleLLMNewToken(queuedChunk.callbackToken, void 0, void 0, void 0, void 0, { chunk: queuedChunk.callbackChunk });
-				}
-			} finally {
-				consumerClosed = true;
-				if (!producerState.done) {
-					abortStream();
-					notifyProducerForSpace();
-				}
-				await producer;
+				})(),
+				delayMs: this._lc_stream_delay,
+				signal: options.signal,
+				abortUpstream: abortStream
+			});
+			for await (const emitted of smoothed) {
+				yield emitted.chunk;
+				if (emitted.callbackChunk != null) await runManager?.handleLLMNewToken(emitted.callbackToken, void 0, void 0, void 0, void 0, { chunk: emitted.callbackChunk });
 			}
 		} finally {
 			options.signal?.removeEventListener("abort", abortStream);

package/dist/cjs/llm/bedrock/index.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"index.cjs","names":["ChatBedrockConverse","supportsBedrockToolCache","insertBedrockToolCachePoint","resolveBedrockPromptCacheTtl","convertToConverseMessages","ConverseStreamCommand","handleConverseStreamContentBlockDelta","handleConverseStreamContentBlockStart","handleConverseStreamMetadata","createConverseToolUseStopChunk","ChatGenerationChunk","AIMessageChunk"],"sources":["../../../../src/llm/bedrock/index.ts"],"sourcesContent":["/**\n * Optimized ChatBedrockConverse wrapper that fixes content block merging for\n * streaming responses and adds support for latest @langchain/aws features:\n *\n * - Application Inference Profiles (PR #9129)\n * - Service Tiers (Priority/Standard/Flex) (PR #9785) - requires AWS SDK 3.966.0+\n *\n * Bedrock's `@langchain/aws` library does not include an `index` property on content\n * blocks (unlike Anthropic/OpenAI), which causes LangChain's `_mergeLists` to append\n * each streaming chunk as a separate array entry instead of merging by index.\n *\n * This wrapper takes full ownership of the stream by directly interfacing with the\n * AWS SDK client (`this.client`) and using custom handlers from `./utils/` that\n * include `contentBlockIndex` in response_metadata for every delta type. It then\n * promotes `contentBlockIndex` to an `index` property on each content block\n * (mirroring Anthropic's pattern) and strips it from metadata to avoid\n * `_mergeDicts` conflicts.\n *\n * When multiple content block types are present (e.g. reasoning + text), text deltas\n * are promoted from strings to array form with `index` so they merge correctly once\n * the accumulated content is already an array.\n */\n\nimport { ChatBedrockConverse } from '@langchain/aws';\nimport { AIMessageChunk } from '@langchain/core/messages';\nimport { ChatGenerationChunk, ChatResult } from '@langchain/core/outputs';\nimport {\n  ConverseStreamCommand,\n  type ConverseStreamOutput,\n  type GuardrailConfiguration,\n  type GuardrailStreamConfiguration,\n} from '@aws-sdk/client-bedrock-runtime';\nimport type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';\nimport type { BaseMessage, ResponseMetadata } from '@langchain/core/messages';\nimport type { ChatBedrockConverseInput } from '@langchain/aws';\nimport type { ContentBlockDeltaEvent } from './types';\nimport {\n  convertToConverseMessages,\n  createConverseToolUseStopChunk,\n  handleConverseStreamContentBlockStart,\n  handleConverseStreamContentBlockDelta,\n  handleConverseStreamMetadata,\n} from './utils';\nimport {\n  resolveBedrockPromptCacheTtl,\n  supportsBedrockToolCache,\n  type PromptCacheTtl,\n} from '@/messages/cache';\nimport { linkStreamLimitCanonical } from '@/llm/streamLimits';\nimport { applyCachePointsToConversePayload } from './cachePoints';\nimport { insertBedrockToolCachePoint } from './toolCache';\n\n/**\n * Service tier type for Bedrock invocations.\n * Requires AWS SDK >= 3.966.0 to actually work.\n * @see https://docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html\n */\nexport type ServiceTierType = 'priority' | 'default' | 'flex' | 'reserved';\n\nexport type CustomGuardrailConfiguration = GuardrailConfiguration &\n  Pick<GuardrailStreamConfiguration, 'streamProcessingMode'>;\n\nconst MAX_STREAM_QUEUE_CHUNKS = 256;\nconst MAX_STREAM_QUEUE_TEXT_CHARS = 8192;\nconst STREAM_CHUNK_MIN_SIZE = 4;\nconst STREAM_BOUNDARIES = new Set([' ', '.', ',', '!', '?', ';', ':']);\n\ntype QueuedGenerationChunk = {\n  chunk: ChatGenerationChunk;\n  callbackChunk?: ChatGenerationChunk;\n  callbackToken: string;\n  smooth: boolean;\n  textLength: number;\n};\n\nfunction findStreamChunkBoundary(text: string, minSize: number): number {\n  if (minSize >= text.length) {\n    return text.length;\n  }\n\n  for (let position = minSize; position < text.length; position++) {\n    if (STREAM_BOUNDARIES.has(text[position])) {\n      return position + 1;\n    }\n  }\n\n  return text.length;\n}\n\nfunction splitStreamToken(text: string): string[] {\n  const chunks: string[] = [];\n  let currentIndex = 0;\n\n  while (currentIndex < text.length) {\n    const remainingText = text.slice(currentIndex);\n    const chunkSize = findStreamChunkBoundary(\n      remainingText,\n      STREAM_CHUNK_MIN_SIZE\n    );\n    chunks.push(text.slice(currentIndex, currentIndex + chunkSize));\n    currentIndex += chunkSize;\n  }\n\n  return chunks;\n}\n\n/**\n * Resolves the text a delta contributes to the smoothing cadence, preferring a\n * text delta over a reasoning delta and ignoring non-string payloads.\n */\nfunction resolveVisibleText(text?: string, reasoningText?: string): string {\n  if (typeof text === 'string') {\n    return text;\n  }\n\n  if (typeof reasoningText === 'string') {\n    return reasoningText;\n  }\n\n  return '';\n}\n\nfunction getCadencedStreamDelay({\n  targetDelay,\n  lastVisibleContentAt,\n  now,\n}: {\n  targetDelay: number;\n  lastVisibleContentAt?: number;\n  now: number;\n}): number {\n  if (targetDelay <= 0 || lastVisibleContentAt == null) {\n    return 0;\n  }\n  return Math.max(0, targetDelay - (now - lastVisibleContentAt));\n}\n\nasync function waitForStreamDelay(\n  delay: number,\n  signal?: AbortSignal\n): Promise<void> {\n  if (delay <= 0 || isSignalAborted(signal)) {\n    return;\n  }\n  await new Promise<void>((resolve) => {\n    const timeoutRef: { current?: ReturnType<typeof setTimeout> } = {};\n    const onAbort = (): void => {\n      if (timeoutRef.current) {\n        clearTimeout(timeoutRef.current);\n      }\n      signal?.removeEventListener('abort', onAbort);\n      resolve();\n    };\n    timeoutRef.current = setTimeout(() => {\n      signal?.removeEventListener('abort', onAbort);\n      resolve();\n    }, delay);\n    signal?.addEventListener('abort', onAbort, { once: true });\n    if (isSignalAborted(signal)) {\n      onAbort();\n    }\n  });\n}\n\nfunction isSignalAborted(signal?: AbortSignal): boolean {\n  return signal?.aborted === true;\n}\n\n/**\n * Extended input interface with additional features:\n * - applicationInferenceProfile: Use an inference profile ARN instead of model ID\n * - serviceTier: Specify service tier (Priority, Standard, Flex, Reserved)\n */\nexport interface CustomChatBedrockConverseInput\n  extends ChatBedrockConverseInput {\n  /**\n   * Enables Bedrock prompt cache checkpoints for message and tool prefixes.\n   */\n  promptCache?: boolean;\n\n  /**\n   * Prompt-cache checkpoint TTL. Defaults to `'1h'` (extended cache) when\n   * `promptCache` is enabled; set `'5m'` for the legacy 5-minute behavior.\n   * Bedrock models that don't support the 1-hour TTL downgrade to 5m\n   * server-side (verified on Sonnet/Opus 4.6), so the default is safe to leave\n   * on; use `'5m'` for any model that rejects it.\n   */\n  promptCacheTtl?: PromptCacheTtl;\n\n  /**\n   * Minimum delay in milliseconds between visible streamed content deltas.\n   */\n  _lc_stream_delay?: number;\n\n  /**\n   * Guardrail configuration for Converse and ConverseStream invocations.\n   * `streamProcessingMode` is only used by ConverseStream.\n   */\n  guardrailConfig?: CustomGuardrailConfiguration;\n\n  /**\n   * Application Inference Profile ARN to use for the model.\n   * For example, \"arn:aws:bedrock:eu-west-1:123456789102:application-inference-profile/fm16bt65tzgx\"\n   * When provided, this ARN will be used for the actual inference calls instead of the model ID.\n   * Must still provide `model` as normal modelId to benefit from all the metadata.\n   * @see https://docs.aws.amazon.com/bedrock/latest/userguide/inference-profiles-create.html\n   */\n  applicationInferenceProfile?: string;\n\n  /**\n   * Service tier for model invocation.\n   * Specifies the processing tier type used for serving the request.\n   * Supported values are 'priority', 'default', 'flex', and 'reserved'.\n   *\n   * - 'priority': Prioritized processing for lower latency\n   * - 'default': Standard processing tier\n   * - 'flex': Flexible processing tier with lower cost\n   * - 'reserved': Reserved capacity for consistent performance\n   *\n   * If not provided, AWS uses the default tier.\n   * Note: Requires AWS SDK >= 3.966.0 to work.\n   * @see https://docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html\n   */\n  serviceTier?: ServiceTierType;\n}\n\n/**\n * Extended call options with serviceTier override support.\n */\nexport interface CustomChatBedrockConverseCallOptions {\n  serviceTier?: ServiceTierType;\n  guardrailConfig?: CustomGuardrailConfiguration;\n}\n\nexport class CustomChatBedrockConverse extends ChatBedrockConverse {\n  _lc_stream_delay: number;\n\n  /**\n   * Whether to insert Bedrock prompt cache checkpoints when available.\n   */\n  promptCache?: boolean;\n\n  /**\n   * Prompt-cache checkpoint TTL (`'5m'` legacy or `'1h'` extended cache).\n   */\n  promptCacheTtl?: PromptCacheTtl;\n\n  /**\n   * Application Inference Profile ARN to use instead of model ID.\n   */\n  applicationInferenceProfile?: string;\n\n  /**\n   * Service tier for model invocation.\n   */\n  serviceTier?: ServiceTierType;\n\n  /**\n   * The configured model id, captured at construction so it survives the\n   * temporary `this.model` swap to an application-inference-profile ARN during\n   * generation. Used to gate the Bedrock tool cache point to Claude models\n   * (see {@link supportsBedrockToolCache}).\n   */\n  private readonly cacheModelId: string;\n\n  constructor(fields?: CustomChatBedrockConverseInput) {\n    super(fields);\n    this.promptCache = fields?.promptCache;\n    this.promptCacheTtl = fields?.promptCacheTtl;\n    this._lc_stream_delay = Math.max(0, fields?._lc_stream_delay ?? 0);\n    this.applicationInferenceProfile = fields?.applicationInferenceProfile;\n    this.serviceTier = fields?.serviceTier;\n    // `super(fields)` initializes `this.model` to LangChain's default Claude\n    // model when `fields.model` is omitted, so fall back to it rather than ''\n    // (which would treat the default Claude model as tool-cache-unsupported).\n    this.cacheModelId = fields?.model ?? this.model;\n  }\n\n  static lc_name(): string {\n    return 'LibreChatBedrockConverse';\n  }\n\n  /**\n   * Get the model ID to use for API calls.\n   * Returns applicationInferenceProfile if set, otherwise returns this.model.\n   */\n  protected getModelId(): string {\n    return this.applicationInferenceProfile ?? this.model;\n  }\n\n  /**\n   * Override invocationParams to add serviceTier support.\n   */\n  override invocationParams(\n    options?: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions\n  ): ReturnType<ChatBedrockConverse['invocationParams']> & {\n    serviceTier?: { type: ServiceTierType };\n  } {\n    const baseParams = super.invocationParams(options);\n    const toolConfig =\n      this.promptCache === true && supportsBedrockToolCache(this.cacheModelId)\n        ? insertBedrockToolCachePoint(\n          baseParams.toolConfig,\n          true,\n          resolveBedrockPromptCacheTtl(this.promptCacheTtl, this.cacheModelId)\n        )\n        : baseParams.toolConfig;\n\n    /** Service tier from options or fall back to class-level setting */\n    const serviceTierType = options?.serviceTier ?? this.serviceTier;\n\n    return {\n      ...baseParams,\n      toolConfig,\n      serviceTier: serviceTierType ? { type: serviceTierType } : undefined,\n    };\n  }\n\n  /**\n   * Override _generateNonStreaming to use applicationInferenceProfile as modelId.\n   * Uses the same model-swapping pattern as streaming for consistency.\n   */\n  override async _generateNonStreaming(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions,\n    runManager?: CallbackManagerForLLMRun\n  ): Promise<ChatResult> {\n    const originalModel = this.model;\n    if (\n      this.applicationInferenceProfile != null &&\n      this.applicationInferenceProfile !== ''\n    ) {\n      this.model = this.applicationInferenceProfile;\n    }\n\n    try {\n      return await super._generateNonStreaming(messages, options, runManager);\n    } finally {\n      this.model = originalModel;\n    }\n  }\n\n  /**\n   * Own the stream end-to-end so we have direct access to every\n   * `contentBlockDelta.contentBlockIndex` from the AWS SDK.\n   *\n   * This replaces the parent's implementation which strips contentBlockIndex\n   * from text and reasoning deltas, making it impossible to merge correctly.\n   */\n  override async *_streamResponseChunks(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions,\n    runManager?: CallbackManagerForLLMRun\n  ): AsyncGenerator<ChatGenerationChunk> {\n    const { converseMessages, converseSystem } =\n      convertToConverseMessages(messages);\n    const params = this.invocationParams(options);\n\n    let { streamUsage } = this;\n    if ((options as Record<string, unknown>).streamUsage !== undefined) {\n      streamUsage = (options as Record<string, unknown>).streamUsage as boolean;\n    }\n\n    const modelId = this.getModelId();\n\n    applyCachePointsToConversePayload({\n      cacheControl: options.cache_control,\n      system: converseSystem,\n      messages: converseMessages,\n      params,\n      modelId,\n    });\n\n    const command = new ConverseStreamCommand({\n      modelId,\n      messages: converseMessages,\n      system: converseSystem,\n      ...(params as Record<string, unknown>),\n    });\n\n    const streamAbortController = new AbortController();\n    const abortStream = (): void => streamAbortController.abort();\n    options.signal?.addEventListener('abort', abortStream, { once: true });\n    if (isSignalAborted(options.signal)) {\n      abortStream();\n    }\n\n    try {\n      const response = await this.client.send(command, {\n        abortSignal: streamAbortController.signal,\n      });\n\n      const stream: AsyncIterable<ConverseStreamOutput> | undefined =\n        response.stream;\n      if (!stream) {\n        return;\n      }\n\n      const seenBlockIndices = new Set<number>();\n      const toolUseBlockIndices = new Set<number>();\n      const queuedChunks: QueuedGenerationChunk[] = [];\n      const producerState: { done: boolean; error?: unknown } = { done: false };\n      let queuedChunkIndex = 0;\n      let bufferedTextLength = 0;\n      let consumerClosed = false;\n      let notifyConsumer: (() => void) | undefined;\n      let notifyProducer: (() => void) | undefined;\n      let hasEmittedVisibleContent = false;\n      let lastVisibleContentAt: number | undefined;\n\n      /**\n       * Guardrails can reject an already-streamed toolUse block at\n       * `messageStop` (`guardrail_intervened`), after `contentBlockStop` has\n       * passed. Only emit eager-execution seals when no guardrails are\n       * configured, so a later intervention can't race an eagerly started tool.\n       */\n      const sealToolUseOnStop =\n        options.guardrailConfig == null && this.guardrailConfig == null;\n\n      const notifyConsumerForChunk = (): void => {\n        notifyConsumer?.();\n        notifyConsumer = undefined;\n      };\n\n      const notifyProducerForSpace = (): void => {\n        notifyProducer?.();\n        notifyProducer = undefined;\n      };\n\n      const hasQueuedChunks = (): boolean =>\n        queuedChunkIndex < queuedChunks.length;\n\n      const getQueuedChunkCount = (): number =>\n        queuedChunks.length - queuedChunkIndex;\n\n      const isQueueAtCapacity = (): boolean =>\n        getQueuedChunkCount() >= MAX_STREAM_QUEUE_CHUNKS ||\n        bufferedTextLength >= MAX_STREAM_QUEUE_TEXT_CHARS;\n\n      const waitForNextChunk = async (): Promise<void> => {\n        if (\n          hasQueuedChunks() ||\n          producerState.done ||\n          producerState.error != null\n        ) {\n          return;\n        }\n        await new Promise<void>((resolve) => {\n          notifyConsumer = resolve;\n        });\n      };\n\n      const waitForQueueSpace = async (): Promise<void> => {\n        while (\n          isQueueAtCapacity() &&\n          !consumerClosed &&\n          !isSignalAborted(options.signal)\n        ) {\n          await new Promise<void>((resolve) => {\n            const signal = options.signal;\n            const onAbort = (): void => {\n              signal?.removeEventListener('abort', onAbort);\n              resolve();\n            };\n            const onSpace = (): void => {\n              signal?.removeEventListener('abort', onAbort);\n              resolve();\n            };\n            notifyProducer = onSpace;\n            signal?.addEventListener('abort', onAbort, { once: true });\n            if (isSignalAborted(signal)) {\n              onAbort();\n            }\n          });\n        }\n      };\n\n      const dequeue = (): QueuedGenerationChunk | undefined => {\n        if (!hasQueuedChunks()) {\n          return undefined;\n        }\n        const queuedChunk = queuedChunks[queuedChunkIndex];\n        queuedChunkIndex++;\n        if (\n          queuedChunkIndex > 128 &&\n          queuedChunkIndex * 2 >= queuedChunks.length\n        ) {\n          queuedChunks.splice(0, queuedChunkIndex);\n          queuedChunkIndex = 0;\n        }\n        return queuedChunk;\n      };\n\n      const enqueue = async (\n        queuedChunk: QueuedGenerationChunk\n      ): Promise<void> => {\n        await waitForQueueSpace();\n        if (consumerClosed || isSignalAborted(options.signal)) {\n          abortStream();\n          throw new Error('AbortError: User aborted the request.');\n        }\n        queuedChunks.push(queuedChunk);\n        if (queuedChunk.smooth) {\n          bufferedTextLength += queuedChunk.textLength;\n        }\n        notifyConsumerForChunk();\n      };\n\n      const enqueueChunk = async ({\n        chunk,\n        callbackChunk,\n        callbackToken = '',\n        smooth = false,\n        textLength = 0,\n      }: {\n        chunk: ChatGenerationChunk;\n        callbackChunk?: ChatGenerationChunk;\n        callbackToken?: string;\n        smooth?: boolean;\n        textLength?: number;\n      }): Promise<void> => {\n        await enqueue({\n          chunk,\n          callbackChunk,\n          callbackToken,\n          smooth,\n          textLength: smooth ? textLength : 0,\n        });\n      };\n\n      const enqueueDelta = async (\n        contentBlockDelta: ContentBlockDeltaEvent\n      ): Promise<void> => {\n        const delta = contentBlockDelta.delta;\n        if (delta == null) {\n          throw new Error('No delta found in content block.');\n        }\n\n        const idx = contentBlockDelta.contentBlockIndex;\n        if (idx != null) {\n          seenBlockIndices.add(idx);\n        }\n\n        const text = delta.text;\n        const reasoningContent = delta.reasoningContent;\n        const reasoningText = reasoningContent?.text;\n        const visibleText = resolveVisibleText(text, reasoningText);\n        const smooth = this._lc_stream_delay > 0 && visibleText !== '';\n        const tokenChunks = smooth\n          ? splitStreamToken(visibleText)\n          : [visibleText];\n\n        for (const token of tokenChunks) {\n          let splitDelta = contentBlockDelta;\n          if (typeof text === 'string') {\n            splitDelta = {\n              ...contentBlockDelta,\n              delta: { text: token },\n            };\n          } else if (\n            typeof reasoningText === 'string' &&\n            reasoningContent != null\n          ) {\n            splitDelta = {\n              ...contentBlockDelta,\n              delta: {\n                reasoningContent: {\n                  ...reasoningContent,\n                  text: token,\n                },\n              },\n            };\n          }\n\n          const deltaChunk = handleConverseStreamContentBlockDelta(splitDelta);\n          const enrichedChunk = this.enrichChunk(deltaChunk, seenBlockIndices);\n          if (enrichedChunk !== deltaChunk) {\n            /** The callback copy is the same emission as the enriched yield;\n             * without the link, stream-limit accounting charges both message\n             * objects and Bedrock tool arguments falsely trip near half the\n             * cap. Linked on the messages, which are what the accounting\n             * observes. */\n            linkStreamLimitCanonical(deltaChunk.message, enrichedChunk.message);\n          }\n          await enqueueChunk({\n            chunk: enrichedChunk,\n            callbackChunk: deltaChunk,\n            callbackToken: deltaChunk.text,\n            smooth,\n            textLength: token.length,\n          });\n        }\n      };\n\n      const producer = (async (): Promise<void> => {\n        try {\n          for await (const event of stream) {\n            if (isSignalAborted(options.signal)) {\n              abortStream();\n              throw new Error('AbortError: User aborted the request.');\n            }\n\n            if (event.contentBlockStart != null) {\n              const startChunk = handleConverseStreamContentBlockStart(\n                event.contentBlockStart\n              );\n              if (startChunk != null) {\n                const idx = event.contentBlockStart.contentBlockIndex;\n                if (idx != null) {\n                  seenBlockIndices.add(idx);\n                  if (event.contentBlockStart.start?.toolUse != null) {\n                    toolUseBlockIndices.add(idx);\n                  }\n                }\n                const enrichedStart = this.enrichChunk(\n                  startChunk,\n                  seenBlockIndices\n                );\n                if (enrichedStart !== startChunk) {\n                  linkStreamLimitCanonical(\n                    startChunk.message,\n                    enrichedStart.message\n                  );\n                }\n                await enqueueChunk({\n                  chunk: enrichedStart,\n                  callbackChunk: startChunk,\n                  callbackToken: startChunk.text,\n                });\n              }\n            } else if (event.contentBlockDelta != null) {\n              await enqueueDelta(event.contentBlockDelta);\n            } else if (event.metadata != null) {\n              await enqueueChunk({\n                chunk: handleConverseStreamMetadata(event.metadata, {\n                  streamUsage,\n                }),\n              });\n            } else if (event.contentBlockStop != null) {\n              const stopIdx = event.contentBlockStop.contentBlockIndex;\n              if (stopIdx != null) {\n                seenBlockIndices.add(stopIdx);\n                if (sealToolUseOnStop && toolUseBlockIndices.has(stopIdx)) {\n                  const sealChunk = createConverseToolUseStopChunk(stopIdx);\n                  await enqueueChunk({\n                    chunk: sealChunk,\n                    callbackChunk: sealChunk,\n                    callbackToken: sealChunk.text,\n                  });\n                }\n              }\n            } else {\n              await enqueueChunk({\n                chunk: new ChatGenerationChunk({\n                  text: '',\n                  message: new AIMessageChunk({\n                    content: '',\n                    response_metadata: { ...event } as ResponseMetadata,\n                  }),\n                }),\n              });\n            }\n          }\n        } catch (error) {\n          producerState.error = error;\n        } finally {\n          producerState.done = true;\n          notifyConsumerForChunk();\n        }\n      })();\n\n      try {\n        let keepStreaming = true;\n        while (keepStreaming) {\n          if (isSignalAborted(options.signal)) {\n            abortStream();\n            throw new Error('AbortError: User aborted the request.');\n          }\n\n          await waitForNextChunk();\n          const queuedChunk = dequeue();\n\n          if (!queuedChunk) {\n            if (producerState.error != null) {\n              throw producerState.error;\n            }\n            if (producerState.done) {\n              keepStreaming = false;\n            }\n            continue;\n          }\n\n          if (queuedChunk.smooth) {\n            bufferedTextLength = Math.max(\n              0,\n              bufferedTextLength - queuedChunk.textLength\n            );\n            notifyProducerForSpace();\n            await waitForStreamDelay(\n              getCadencedStreamDelay({\n                targetDelay: hasEmittedVisibleContent\n                  ? this._lc_stream_delay\n                  : 0,\n                lastVisibleContentAt,\n                now: Date.now(),\n              }),\n              options.signal\n            );\n            if (isSignalAborted(options.signal)) {\n              abortStream();\n              throw new Error('AbortError: User aborted the request.');\n            }\n            hasEmittedVisibleContent = true;\n            lastVisibleContentAt = Date.now();\n          } else {\n            notifyProducerForSpace();\n          }\n\n          yield queuedChunk.chunk;\n\n          if (queuedChunk.callbackChunk != null) {\n            await runManager?.handleLLMNewToken(\n              queuedChunk.callbackToken,\n              undefined,\n              undefined,\n              undefined,\n              undefined,\n              { chunk: queuedChunk.callbackChunk }\n            );\n          }\n        }\n      } finally {\n        consumerClosed = true;\n        if (!producerState.done) {\n          abortStream();\n          notifyProducerForSpace();\n        }\n        await producer;\n      }\n    } finally {\n      options.signal?.removeEventListener('abort', abortStream);\n      if (!streamAbortController.signal.aborted) {\n        streamAbortController.abort();\n      }\n    }\n  }\n\n  /**\n   * Inject `index` on content blocks for proper merge behaviour, then strip\n   * `contentBlockIndex` from response_metadata to prevent `_mergeDicts` conflicts.\n   *\n   * Text string content is promoted to array form only when the stream contains\n   * multiple content block indices (e.g. reasoning at index 0, text at index 1),\n   * ensuring text merges correctly with the already-array accumulated content.\n   */\n  private enrichChunk(\n    chunk: ChatGenerationChunk,\n    seenBlockIndices: Set<number>\n  ): ChatGenerationChunk {\n    const message = chunk.message;\n    if (!(message instanceof AIMessageChunk)) {\n      return chunk;\n    }\n\n    const metadata = message.response_metadata as Record<string, unknown>;\n    const blockIndex = this.extractContentBlockIndex(metadata);\n    const hasMetadataIndex = blockIndex != null;\n\n    let content: AIMessageChunk['content'] = message.content;\n    let contentModified = false;\n\n    if (Array.isArray(content) && blockIndex != null) {\n      content = content.map((block) =>\n        typeof block === 'object' && !('index' in block)\n          ? { ...block, index: blockIndex }\n          : block\n      );\n      contentModified = true;\n    } else if (\n      typeof content === 'string' &&\n      content !== '' &&\n      blockIndex != null &&\n      seenBlockIndices.size > 1\n    ) {\n      content = [{ type: 'text', text: content, index: blockIndex }];\n      contentModified = true;\n    }\n\n    if (!contentModified && !hasMetadataIndex) {\n      return chunk;\n    }\n\n    const cleanedMetadata = hasMetadataIndex\n      ? (this.removeContentBlockIndex(metadata) as Record<string, unknown>)\n      : metadata;\n\n    return new ChatGenerationChunk({\n      text: chunk.text,\n      message: new AIMessageChunk({\n        ...message,\n        content,\n        response_metadata: cleanedMetadata,\n      }),\n      generationInfo: chunk.generationInfo,\n    });\n  }\n\n  /**\n   * Extract `contentBlockIndex` from the top level of response_metadata.\n   * Our custom handlers always place it at the top level.\n   */\n  private extractContentBlockIndex(\n    metadata: Record<string, unknown>\n  ): number | undefined {\n    if (\n      'contentBlockIndex' in metadata &&\n      typeof metadata.contentBlockIndex === 'number'\n    ) {\n      return metadata.contentBlockIndex;\n    }\n    return undefined;\n  }\n\n  private removeContentBlockIndex(obj: unknown): unknown {\n    if (obj === null || obj === undefined) {\n      return obj;\n    }\n\n    if (Array.isArray(obj)) {\n      return obj.map((item) => this.removeContentBlockIndex(item));\n    }\n\n    if (typeof obj === 'object') {\n      const cleaned: Record<string, unknown> = {};\n      for (const [key, value] of Object.entries(obj)) {\n        if (key !== 'contentBlockIndex') {\n          cleaned[key] = this.removeContentBlockIndex(value);\n        }\n      }\n      return cleaned;\n    }\n\n    return obj;\n  }\n}\n\nexport type { ChatBedrockConverseInput };\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA8DA,MAAM,0BAA0B;AAChC,MAAM,8BAA8B;AACpC,MAAM,wBAAwB;AAC9B,MAAM,oBAAoB,IAAI,IAAI;CAAC;CAAK;CAAK;CAAK;CAAK;CAAK;CAAK;AAAG,CAAC;AAUrE,SAAS,wBAAwB,MAAc,SAAyB;CACtE,IAAI,WAAW,KAAK,QAClB,OAAO,KAAK;CAGd,KAAK,IAAI,WAAW,SAAS,WAAW,KAAK,QAAQ,YACnD,IAAI,kBAAkB,IAAI,KAAK,SAAS,GACtC,OAAO,WAAW;CAItB,OAAO,KAAK;AACd;AAEA,SAAS,iBAAiB,MAAwB;CAChD,MAAM,SAAmB,CAAC;CAC1B,IAAI,eAAe;CAEnB,OAAO,eAAe,KAAK,QAAQ;EAEjC,MAAM,YAAY,wBADI,KAAK,MAAM,YAEnB,GACZ,qBACF;EACA,OAAO,KAAK,KAAK,MAAM,cAAc,eAAe,SAAS,CAAC;EAC9D,gBAAgB;CAClB;CAEA,OAAO;AACT;;;;;AAMA,SAAS,mBAAmB,MAAe,eAAgC;CACzE,IAAI,OAAO,SAAS,UAClB,OAAO;CAGT,IAAI,OAAO,kBAAkB,UAC3B,OAAO;CAGT,OAAO;AACT;AAEA,SAAS,uBAAuB,EAC9B,aACA,sBACA,OAKS;CACT,IAAI,eAAe,KAAK,wBAAwB,MAC9C,OAAO;CAET,OAAO,KAAK,IAAI,GAAG,eAAe,MAAM,qBAAqB;AAC/D;AAEA,eAAe,mBACb,OACA,QACe;CACf,IAAI,SAAS,KAAK,gBAAgB,MAAM,GACtC;CAEF,MAAM,IAAI,SAAe,YAAY;EACnC,MAAM,aAA0D,CAAC;EACjE,MAAM,gBAAsB;GAC1B,IAAI,WAAW,SACb,aAAa,WAAW,OAAO;GAEjC,QAAQ,oBAAoB,SAAS,OAAO;GAC5C,QAAQ;EACV;EACA,WAAW,UAAU,iBAAiB;GACpC,QAAQ,oBAAoB,SAAS,OAAO;GAC5C,QAAQ;EACV,GAAG,KAAK;EACR,QAAQ,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;EACzD,IAAI,gBAAgB,MAAM,GACxB,QAAQ;CAEZ,CAAC;AACH;AAEA,SAAS,gBAAgB,QAA+B;CACtD,OAAO,QAAQ,YAAY;AAC7B;AAoEA,IAAa,4BAAb,cAA+CA,eAAAA,oBAAoB;CACjE;;;;CAKA;;;;CAKA;;;;CAKA;;;;CAKA;;;;;;;CAQA;CAEA,YAAY,QAAyC;EACnD,MAAM,MAAM;EACZ,KAAK,cAAc,QAAQ;EAC3B,KAAK,iBAAiB,QAAQ;EAC9B,KAAK,mBAAmB,KAAK,IAAI,GAAG,QAAQ,oBAAoB,CAAC;EACjE,KAAK,8BAA8B,QAAQ;EAC3C,KAAK,cAAc,QAAQ;EAI3B,KAAK,eAAe,QAAQ,SAAS,KAAK;CAC5C;CAEA,OAAO,UAAkB;EACvB,OAAO;CACT;;;;;CAMA,aAA+B;EAC7B,OAAO,KAAK,+BAA+B,KAAK;CAClD;;;;CAKA,iBACE,SAGA;EACA,MAAM,aAAa,MAAM,iBAAiB,OAAO;EACjD,MAAM,aACJ,KAAK,gBAAgB,QAAQC,cAAAA,yBAAyB,KAAK,YAAY,IACnEC,kBAAAA,4BACA,WAAW,YACX,MACAC,cAAAA,6BAA6B,KAAK,gBAAgB,KAAK,YAAY,CACrE,IACE,WAAW;;EAGjB,MAAM,kBAAkB,SAAS,eAAe,KAAK;EAErD,OAAO;GACL,GAAG;GACH;GACA,aAAa,kBAAkB,EAAE,MAAM,gBAAgB,IAAI,KAAA;EAC7D;CACF;;;;;CAMA,MAAe,sBACb,UACA,SACA,YACqB;EACrB,MAAM,gBAAgB,KAAK;EAC3B,IACE,KAAK,+BAA+B,QACpC,KAAK,gCAAgC,IAErC,KAAK,QAAQ,KAAK;EAGpB,IAAI;GACF,OAAO,MAAM,MAAM,sBAAsB,UAAU,SAAS,UAAU;EACxE,UAAU;GACR,KAAK,QAAQ;EACf;CACF;;;;;;;;CASA,OAAgB,sBACd,UACA,SACA,YACqC;EACrC,MAAM,EAAE,kBAAkB,mBACxBC,uBAAAA,0BAA0B,QAAQ;EACpC,MAAM,SAAS,KAAK,iBAAiB,OAAO;EAE5C,IAAI,EAAE,gBAAgB;EACtB,IAAK,QAAoC,gBAAgB,KAAA,GACvD,cAAe,QAAoC;EAGrD,MAAM,UAAU,KAAK,WAAW;EAEhC,oBAAA,kCAAkC;GAChC,cAAc,QAAQ;GACtB,QAAQ;GACR,UAAU;GACV;GACA;EACF,CAAC;EAED,MAAM,UAAU,IAAIC,gCAAAA,sBAAsB;GACxC;GACA,UAAU;GACV,QAAQ;GACR,GAAI;EACN,CAAC;EAED,MAAM,wBAAwB,IAAI,gBAAgB;EAClD,MAAM,oBAA0B,sBAAsB,MAAM;EAC5D,QAAQ,QAAQ,iBAAiB,SAAS,aAAa,EAAE,MAAM,KAAK,CAAC;EACrE,IAAI,gBAAgB,QAAQ,MAAM,GAChC,YAAY;EAGd,IAAI;GAKF,MAAM,UACJ,MALqB,KAAK,OAAO,KAAK,SAAS,EAC/C,aAAa,sBAAsB,OACrC,CAAC,EAAA,CAGU;GACX,IAAI,CAAC,QACH;GAGF,MAAM,mCAAmB,IAAI,IAAY;GACzC,MAAM,sCAAsB,IAAI,IAAY;GAC5C,MAAM,eAAwC,CAAC;GAC/C,MAAM,gBAAoD,EAAE,MAAM,MAAM;GACxE,IAAI,mBAAmB;GACvB,IAAI,qBAAqB;GACzB,IAAI,iBAAiB;GACrB,IAAI;GACJ,IAAI;GACJ,IAAI,2BAA2B;GAC/B,IAAI;;;;;;;GAQJ,MAAM,oBACJ,QAAQ,mBAAmB,QAAQ,KAAK,mBAAmB;GAE7D,MAAM,+BAAqC;IACzC,iBAAiB;IACjB,iBAAiB,KAAA;GACnB;GAEA,MAAM,+BAAqC;IACzC,iBAAiB;IACjB,iBAAiB,KAAA;GACnB;GAEA,MAAM,wBACJ,mBAAmB,aAAa;GAElC,MAAM,4BACJ,aAAa,SAAS;GAExB,MAAM,0BACJ,oBAAoB,KAAK,2BACzB,sBAAsB;GAExB,MAAM,mBAAmB,YAA2B;IAClD,IACE,gBAAgB,KAChB,cAAc,QACd,cAAc,SAAS,MAEvB;IAEF,MAAM,IAAI,SAAe,YAAY;KACnC,iBAAiB;IACnB,CAAC;GACH;GAEA,MAAM,oBAAoB,YAA2B;IACnD,OACE,kBAAkB,KAClB,CAAC,kBACD,CAAC,gBAAgB,QAAQ,MAAM,GAE/B,MAAM,IAAI,SAAe,YAAY;KACnC,MAAM,SAAS,QAAQ;KACvB,MAAM,gBAAsB;MAC1B,QAAQ,oBAAoB,SAAS,OAAO;MAC5C,QAAQ;KACV;KACA,MAAM,gBAAsB;MAC1B,QAAQ,oBAAoB,SAAS,OAAO;MAC5C,QAAQ;KACV;KACA,iBAAiB;KACjB,QAAQ,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;KACzD,IAAI,gBAAgB,MAAM,GACxB,QAAQ;IAEZ,CAAC;GAEL;GAEA,MAAM,gBAAmD;IACvD,IAAI,CAAC,gBAAgB,GACnB;IAEF,MAAM,cAAc,aAAa;IACjC;IACA,IACE,mBAAmB,OACnB,mBAAmB,KAAK,aAAa,QACrC;KACA,aAAa,OAAO,GAAG,gBAAgB;KACvC,mBAAmB;IACrB;IACA,OAAO;GACT;GAEA,MAAM,UAAU,OACd,gBACkB;IAClB,MAAM,kBAAkB;IACxB,IAAI,kBAAkB,gBAAgB,QAAQ,MAAM,GAAG;KACrD,YAAY;KACZ,MAAM,IAAI,MAAM,uCAAuC;IACzD;IACA,aAAa,KAAK,WAAW;IAC7B,IAAI,YAAY,QACd,sBAAsB,YAAY;IAEpC,uBAAuB;GACzB;GAEA,MAAM,eAAe,OAAO,EAC1B,OACA,eACA,gBAAgB,IAChB,SAAS,OACT,aAAa,QAOM;IACnB,MAAM,QAAQ;KACZ;KACA;KACA;KACA;KACA,YAAY,SAAS,aAAa;IACpC,CAAC;GACH;GAEA,MAAM,eAAe,OACnB,sBACkB;IAClB,MAAM,QAAQ,kBAAkB;IAChC,IAAI,SAAS,MACX,MAAM,IAAI,MAAM,kCAAkC;IAGpD,MAAM,MAAM,kBAAkB;IAC9B,IAAI,OAAO,MACT,iBAAiB,IAAI,GAAG;IAG1B,MAAM,OAAO,MAAM;IACnB,MAAM,mBAAmB,MAAM;IAC/B,MAAM,gBAAgB,kBAAkB;IACxC,MAAM,cAAc,mBAAmB,MAAM,aAAa;IAC1D,MAAM,SAAS,KAAK,mBAAmB,KAAK,gBAAgB;IAC5D,MAAM,cAAc,SAChB,iBAAiB,WAAW,IAC5B,CAAC,WAAW;IAEhB,KAAK,MAAM,SAAS,aAAa;KAC/B,IAAI,aAAa;KACjB,IAAI,OAAO,SAAS,UAClB,aAAa;MACX,GAAG;MACH,OAAO,EAAE,MAAM,MAAM;KACvB;UACK,IACL,OAAO,kBAAkB,YACzB,oBAAoB,MAEpB,aAAa;MACX,GAAG;MACH,OAAO,EACL,kBAAkB;OAChB,GAAG;OACH,MAAM;MACR,EACF;KACF;KAGF,MAAM,aAAaC,wBAAAA,sCAAsC,UAAU;KACnE,MAAM,gBAAgB,KAAK,YAAY,YAAY,gBAAgB;KACnE,IAAI,kBAAkB;;;;;;KAMpB,qBAAA,yBAAyB,WAAW,SAAS,cAAc,OAAO;KAEpE,MAAM,aAAa;MACjB,OAAO;MACP,eAAe;MACf,eAAe,WAAW;MAC1B;MACA,YAAY,MAAM;KACpB,CAAC;IACH;GACF;GAEA,MAAM,YAAY,YAA2B;IAC3C,IAAI;KACF,WAAW,MAAM,SAAS,QAAQ;MAChC,IAAI,gBAAgB,QAAQ,MAAM,GAAG;OACnC,YAAY;OACZ,MAAM,IAAI,MAAM,uCAAuC;MACzD;MAEA,IAAI,MAAM,qBAAqB,MAAM;OACnC,MAAM,aAAaC,wBAAAA,sCACjB,MAAM,iBACR;OACA,IAAI,cAAc,MAAM;QACtB,MAAM,MAAM,MAAM,kBAAkB;QACpC,IAAI,OAAO,MAAM;SACf,iBAAiB,IAAI,GAAG;SACxB,IAAI,MAAM,kBAAkB,OAAO,WAAW,MAC5C,oBAAoB,IAAI,GAAG;QAE/B;QACA,MAAM,gBAAgB,KAAK,YACzB,YACA,gBACF;QACA,IAAI,kBAAkB,YACpB,qBAAA,yBACE,WAAW,SACX,cAAc,OAChB;QAEF,MAAM,aAAa;SACjB,OAAO;SACP,eAAe;SACf,eAAe,WAAW;QAC5B,CAAC;OACH;MACF,OAAO,IAAI,MAAM,qBAAqB,MACpC,MAAM,aAAa,MAAM,iBAAiB;WACrC,IAAI,MAAM,YAAY,MAC3B,MAAM,aAAa,EACjB,OAAOC,wBAAAA,6BAA6B,MAAM,UAAU,EAClD,YACF,CAAC,EACH,CAAC;WACI,IAAI,MAAM,oBAAoB,MAAM;OACzC,MAAM,UAAU,MAAM,iBAAiB;OACvC,IAAI,WAAW,MAAM;QACnB,iBAAiB,IAAI,OAAO;QAC5B,IAAI,qBAAqB,oBAAoB,IAAI,OAAO,GAAG;SACzD,MAAM,YAAYC,wBAAAA,+BAA+B,OAAO;SACxD,MAAM,aAAa;UACjB,OAAO;UACP,eAAe;UACf,eAAe,UAAU;SAC3B,CAAC;QACH;OACF;MACF,OACE,MAAM,aAAa,EACjB,OAAO,IAAIC,wBAAAA,oBAAoB;OAC7B,MAAM;OACN,SAAS,IAAIC,yBAAAA,eAAe;QAC1B,SAAS;QACT,mBAAmB,EAAE,GAAG,MAAM;OAChC,CAAC;MACH,CAAC,EACH,CAAC;KAEL;IACF,SAAS,OAAO;KACd,cAAc,QAAQ;IACxB,UAAU;KACR,cAAc,OAAO;KACrB,uBAAuB;IACzB;GACF,EAAA,CAAG;GAEH,IAAI;IACF,IAAI,gBAAgB;IACpB,OAAO,eAAe;KACpB,IAAI,gBAAgB,QAAQ,MAAM,GAAG;MACnC,YAAY;MACZ,MAAM,IAAI,MAAM,uCAAuC;KACzD;KAEA,MAAM,iBAAiB;KACvB,MAAM,cAAc,QAAQ;KAE5B,IAAI,CAAC,aAAa;MAChB,IAAI,cAAc,SAAS,MACzB,MAAM,cAAc;MAEtB,IAAI,cAAc,MAChB,gBAAgB;MAElB;KACF;KAEA,IAAI,YAAY,QAAQ;MACtB,qBAAqB,KAAK,IACxB,GACA,qBAAqB,YAAY,UACnC;MACA,uBAAuB;MACvB,MAAM,mBACJ,uBAAuB;OACrB,aAAa,2BACT,KAAK,mBACL;OACJ;OACA,KAAK,KAAK,IAAI;MAChB,CAAC,GACD,QAAQ,MACV;MACA,IAAI,gBAAgB,QAAQ,MAAM,GAAG;OACnC,YAAY;OACZ,MAAM,IAAI,MAAM,uCAAuC;MACzD;MACA,2BAA2B;MAC3B,uBAAuB,KAAK,IAAI;KAClC,OACE,uBAAuB;KAGzB,MAAM,YAAY;KAElB,IAAI,YAAY,iBAAiB,MAC/B,MAAM,YAAY,kBAChB,YAAY,eACZ,KAAA,GACA,KAAA,GACA,KAAA,GACA,KAAA,GACA,EAAE,OAAO,YAAY,cAAc,CACrC;IAEJ;GACF,UAAU;IACR,iBAAiB;IACjB,IAAI,CAAC,cAAc,MAAM;KACvB,YAAY;KACZ,uBAAuB;IACzB;IACA,MAAM;GACR;EACF,UAAU;GACR,QAAQ,QAAQ,oBAAoB,SAAS,WAAW;GACxD,IAAI,CAAC,sBAAsB,OAAO,SAChC,sBAAsB,MAAM;EAEhC;CACF;;;;;;;;;CAUA,YACE,OACA,kBACqB;EACrB,MAAM,UAAU,MAAM;EACtB,IAAI,EAAE,mBAAmBA,yBAAAA,iBACvB,OAAO;EAGT,MAAM,WAAW,QAAQ;EACzB,MAAM,aAAa,KAAK,yBAAyB,QAAQ;EACzD,MAAM,mBAAmB,cAAc;EAEvC,IAAI,UAAqC,QAAQ;EACjD,IAAI,kBAAkB;EAEtB,IAAI,MAAM,QAAQ,OAAO,KAAK,cAAc,MAAM;GAChD,UAAU,QAAQ,KAAK,UACrB,OAAO,UAAU,YAAY,EAAE,WAAW,SACtC;IAAE,GAAG;IAAO,OAAO;GAAW,IAC9B,KACN;GACA,kBAAkB;EACpB,OAAO,IACL,OAAO,YAAY,YACnB,YAAY,MACZ,cAAc,QACd,iBAAiB,OAAO,GACxB;GACA,UAAU,CAAC;IAAE,MAAM;IAAQ,MAAM;IAAS,OAAO;GAAW,CAAC;GAC7D,kBAAkB;EACpB;EAEA,IAAI,CAAC,mBAAmB,CAAC,kBACvB,OAAO;EAGT,MAAM,kBAAkB,mBACnB,KAAK,wBAAwB,QAAQ,IACtC;EAEJ,OAAO,IAAID,wBAAAA,oBAAoB;GAC7B,MAAM,MAAM;GACZ,SAAS,IAAIC,yBAAAA,eAAe;IAC1B,GAAG;IACH;IACA,mBAAmB;GACrB,CAAC;GACD,gBAAgB,MAAM;EACxB,CAAC;CACH;;;;;CAMA,yBACE,UACoB;EACpB,IACE,uBAAuB,YACvB,OAAO,SAAS,sBAAsB,UAEtC,OAAO,SAAS;CAGpB;CAEA,wBAAgC,KAAuB;EACrD,IAAI,QAAQ,QAAQ,QAAQ,KAAA,GAC1B,OAAO;EAGT,IAAI,MAAM,QAAQ,GAAG,GACnB,OAAO,IAAI,KAAK,SAAS,KAAK,wBAAwB,IAAI,CAAC;EAG7D,IAAI,OAAO,QAAQ,UAAU;GAC3B,MAAM,UAAmC,CAAC;GAC1C,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,GAAG,GAC3C,IAAI,QAAQ,qBACV,QAAQ,OAAO,KAAK,wBAAwB,KAAK;GAGrD,OAAO;EACT;EAEA,OAAO;CACT;AACF"}
+{"version":3,"file":"index.cjs","names":["ChatBedrockConverse","resolveStreamDelay","supportsBedrockToolCache","insertBedrockToolCachePoint","resolveBedrockPromptCacheTtl","convertToConverseMessages","ConverseStreamCommand","isSignalAborted","handleConverseStreamContentBlockDelta","smoothStream","handleConverseStreamContentBlockStart","handleConverseStreamMetadata","createConverseToolUseStopChunk","ChatGenerationChunk","AIMessageChunk"],"sources":["../../../../src/llm/bedrock/index.ts"],"sourcesContent":["/**\n * Optimized ChatBedrockConverse wrapper that fixes content block merging for\n * streaming responses and adds support for latest @langchain/aws features:\n *\n * - Application Inference Profiles (PR #9129)\n * - Service Tiers (Priority/Standard/Flex) (PR #9785) - requires AWS SDK 3.966.0+\n *\n * Bedrock's `@langchain/aws` library does not include an `index` property on content\n * blocks (unlike Anthropic/OpenAI), which causes LangChain's `_mergeLists` to append\n * each streaming chunk as a separate array entry instead of merging by index.\n *\n * This wrapper takes full ownership of the stream by directly interfacing with the\n * AWS SDK client (`this.client`) and using custom handlers from `./utils/` that\n * include `contentBlockIndex` in response_metadata for every delta type. It then\n * promotes `contentBlockIndex` to an `index` property on each content block\n * (mirroring Anthropic's pattern) and strips it from metadata to avoid\n * `_mergeDicts` conflicts.\n *\n * When multiple content block types are present (e.g. reasoning + text), text deltas\n * are promoted from strings to array form with `index` so they merge correctly once\n * the accumulated content is already an array.\n */\n\nimport { ChatBedrockConverse } from '@langchain/aws';\nimport { AIMessageChunk } from '@langchain/core/messages';\nimport { ChatGenerationChunk, ChatResult } from '@langchain/core/outputs';\nimport {\n  ConverseStreamCommand,\n  type ConverseStreamOutput,\n  type GuardrailConfiguration,\n  type GuardrailStreamConfiguration,\n} from '@aws-sdk/client-bedrock-runtime';\nimport type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';\nimport type { BaseMessage, ResponseMetadata } from '@langchain/core/messages';\nimport type { ChatBedrockConverseInput } from '@langchain/aws';\nimport type { SmoothItem } from '@/llm/stream/smoother';\nimport type { ContentBlockDeltaEvent } from './types';\nimport {\n  convertToConverseMessages,\n  createConverseToolUseStopChunk,\n  handleConverseStreamContentBlockStart,\n  handleConverseStreamContentBlockDelta,\n  handleConverseStreamMetadata,\n} from './utils';\nimport {\n  resolveBedrockPromptCacheTtl,\n  supportsBedrockToolCache,\n  type PromptCacheTtl,\n} from '@/messages/cache';\nimport { smoothStream, resolveStreamDelay, isSignalAborted } from '@/llm/stream/smoother';\nimport { linkStreamLimitCanonical } from '@/llm/streamLimits';\nimport { applyCachePointsToConversePayload } from './cachePoints';\nimport { insertBedrockToolCachePoint } from './toolCache';\n\n/**\n * Service tier type for Bedrock invocations.\n * Requires AWS SDK >= 3.966.0 to actually work.\n * @see https://docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html\n */\nexport type ServiceTierType = 'priority' | 'default' | 'flex' | 'reserved';\n\nexport type CustomGuardrailConfiguration = GuardrailConfiguration &\n  Pick<GuardrailStreamConfiguration, 'streamProcessingMode'>;\n\ntype BedrockEmittedChunk = {\n  chunk: ChatGenerationChunk;\n  callbackChunk?: ChatGenerationChunk;\n  callbackToken: string;\n};\n\n/**\n * Resolves the text a delta contributes to the smoothing cadence, preferring a\n * text delta over a reasoning delta and ignoring non-string payloads.\n */\nfunction resolveVisibleText(text?: string, reasoningText?: string): string {\n  if (typeof text === 'string') {\n    return text;\n  }\n\n  if (typeof reasoningText === 'string') {\n    return reasoningText;\n  }\n\n  return '';\n}\n\n/**\n * Extended input interface with additional features:\n * - applicationInferenceProfile: Use an inference profile ARN instead of model ID\n * - serviceTier: Specify service tier (Priority, Standard, Flex, Reserved)\n */\nexport interface CustomChatBedrockConverseInput\n  extends ChatBedrockConverseInput {\n  /**\n   * Enables Bedrock prompt cache checkpoints for message and tool prefixes.\n   */\n  promptCache?: boolean;\n\n  /**\n   * Prompt-cache checkpoint TTL. Defaults to `'1h'` (extended cache) when\n   * `promptCache` is enabled; set `'5m'` for the legacy 5-minute behavior.\n   * Bedrock models that don't support the 1-hour TTL downgrade to 5m\n   * server-side (verified on Sonnet/Opus 4.6), so the default is safe to leave\n   * on; use `'5m'` for any model that rejects it.\n   */\n  promptCacheTtl?: PromptCacheTtl;\n\n  /**\n   * Minimum delay in milliseconds between visible streamed content deltas.\n   */\n  _lc_stream_delay?: number;\n\n  /**\n   * Guardrail configuration for Converse and ConverseStream invocations.\n   * `streamProcessingMode` is only used by ConverseStream.\n   */\n  guardrailConfig?: CustomGuardrailConfiguration;\n\n  /**\n   * Application Inference Profile ARN to use for the model.\n   * For example, \"arn:aws:bedrock:eu-west-1:123456789102:application-inference-profile/fm16bt65tzgx\"\n   * When provided, this ARN will be used for the actual inference calls instead of the model ID.\n   * Must still provide `model` as normal modelId to benefit from all the metadata.\n   * @see https://docs.aws.amazon.com/bedrock/latest/userguide/inference-profiles-create.html\n   */\n  applicationInferenceProfile?: string;\n\n  /**\n   * Service tier for model invocation.\n   * Specifies the processing tier type used for serving the request.\n   * Supported values are 'priority', 'default', 'flex', and 'reserved'.\n   *\n   * - 'priority': Prioritized processing for lower latency\n   * - 'default': Standard processing tier\n   * - 'flex': Flexible processing tier with lower cost\n   * - 'reserved': Reserved capacity for consistent performance\n   *\n   * If not provided, AWS uses the default tier.\n   * Note: Requires AWS SDK >= 3.966.0 to work.\n   * @see https://docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html\n   */\n  serviceTier?: ServiceTierType;\n}\n\n/**\n * Extended call options with serviceTier override support.\n */\nexport interface CustomChatBedrockConverseCallOptions {\n  serviceTier?: ServiceTierType;\n  guardrailConfig?: CustomGuardrailConfiguration;\n}\n\nexport class CustomChatBedrockConverse extends ChatBedrockConverse {\n  _lc_stream_delay: number;\n\n  /**\n   * Whether to insert Bedrock prompt cache checkpoints when available.\n   */\n  promptCache?: boolean;\n\n  /**\n   * Prompt-cache checkpoint TTL (`'5m'` legacy or `'1h'` extended cache).\n   */\n  promptCacheTtl?: PromptCacheTtl;\n\n  /**\n   * Application Inference Profile ARN to use instead of model ID.\n   */\n  applicationInferenceProfile?: string;\n\n  /**\n   * Service tier for model invocation.\n   */\n  serviceTier?: ServiceTierType;\n\n  /**\n   * The configured model id, captured at construction so it survives the\n   * temporary `this.model` swap to an application-inference-profile ARN during\n   * generation. Used to gate the Bedrock tool cache point to Claude models\n   * (see {@link supportsBedrockToolCache}).\n   */\n  private readonly cacheModelId: string;\n\n  constructor(fields?: CustomChatBedrockConverseInput) {\n    super(fields);\n    this.promptCache = fields?.promptCache;\n    this.promptCacheTtl = fields?.promptCacheTtl;\n    this._lc_stream_delay = resolveStreamDelay(fields?._lc_stream_delay);\n    this.applicationInferenceProfile = fields?.applicationInferenceProfile;\n    this.serviceTier = fields?.serviceTier;\n    // `super(fields)` initializes `this.model` to LangChain's default Claude\n    // model when `fields.model` is omitted, so fall back to it rather than ''\n    // (which would treat the default Claude model as tool-cache-unsupported).\n    this.cacheModelId = fields?.model ?? this.model;\n  }\n\n  static lc_name(): string {\n    return 'LibreChatBedrockConverse';\n  }\n\n  /**\n   * Get the model ID to use for API calls.\n   * Returns applicationInferenceProfile if set, otherwise returns this.model.\n   */\n  protected getModelId(): string {\n    return this.applicationInferenceProfile ?? this.model;\n  }\n\n  /**\n   * Override invocationParams to add serviceTier support.\n   */\n  override invocationParams(\n    options?: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions\n  ): ReturnType<ChatBedrockConverse['invocationParams']> & {\n    serviceTier?: { type: ServiceTierType };\n  } {\n    const baseParams = super.invocationParams(options);\n    const toolConfig =\n      this.promptCache === true && supportsBedrockToolCache(this.cacheModelId)\n        ? insertBedrockToolCachePoint(\n          baseParams.toolConfig,\n          true,\n          resolveBedrockPromptCacheTtl(this.promptCacheTtl, this.cacheModelId)\n        )\n        : baseParams.toolConfig;\n\n    /** Service tier from options or fall back to class-level setting */\n    const serviceTierType = options?.serviceTier ?? this.serviceTier;\n\n    return {\n      ...baseParams,\n      toolConfig,\n      serviceTier: serviceTierType ? { type: serviceTierType } : undefined,\n    };\n  }\n\n  /**\n   * Override _generateNonStreaming to use applicationInferenceProfile as modelId.\n   * Uses the same model-swapping pattern as streaming for consistency.\n   */\n  override async _generateNonStreaming(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions,\n    runManager?: CallbackManagerForLLMRun\n  ): Promise<ChatResult> {\n    const originalModel = this.model;\n    if (\n      this.applicationInferenceProfile != null &&\n      this.applicationInferenceProfile !== ''\n    ) {\n      this.model = this.applicationInferenceProfile;\n    }\n\n    try {\n      return await super._generateNonStreaming(messages, options, runManager);\n    } finally {\n      this.model = originalModel;\n    }\n  }\n\n  /**\n   * Own the stream end-to-end so we have direct access to every\n   * `contentBlockDelta.contentBlockIndex` from the AWS SDK.\n   *\n   * This replaces the parent's implementation which strips contentBlockIndex\n   * from text and reasoning deltas, making it impossible to merge correctly.\n   */\n  override async *_streamResponseChunks(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions,\n    runManager?: CallbackManagerForLLMRun\n  ): AsyncGenerator<ChatGenerationChunk> {\n    const { converseMessages, converseSystem } =\n      convertToConverseMessages(messages);\n    const params = this.invocationParams(options);\n\n    let { streamUsage } = this;\n    if ((options as Record<string, unknown>).streamUsage !== undefined) {\n      streamUsage = (options as Record<string, unknown>).streamUsage as boolean;\n    }\n\n    const modelId = this.getModelId();\n\n    applyCachePointsToConversePayload({\n      cacheControl: options.cache_control,\n      system: converseSystem,\n      messages: converseMessages,\n      params,\n      modelId,\n    });\n\n    const command = new ConverseStreamCommand({\n      modelId,\n      messages: converseMessages,\n      system: converseSystem,\n      ...(params as Record<string, unknown>),\n    });\n\n    const streamAbortController = new AbortController();\n    const abortStream = (): void => streamAbortController.abort();\n    options.signal?.addEventListener('abort', abortStream, { once: true });\n    if (isSignalAborted(options.signal)) {\n      abortStream();\n    }\n\n    try {\n      const response = await this.client.send(command, {\n        abortSignal: streamAbortController.signal,\n      });\n\n      const stream: AsyncIterable<ConverseStreamOutput> | undefined =\n        response.stream;\n      if (!stream) {\n        return;\n      }\n\n      const seenBlockIndices = new Set<number>();\n      const toolUseBlockIndices = new Set<number>();\n\n      /**\n       * Guardrails can reject an already-streamed toolUse block at\n       * `messageStop` (`guardrail_intervened`), after `contentBlockStop` has\n       * passed. Only emit eager-execution seals when no guardrails are\n       * configured, so a later intervention can't race an eagerly started tool.\n       */\n      const sealToolUseOnStop =\n        options.guardrailConfig == null && this.guardrailConfig == null;\n\n      /**\n       * Builds the emission for one piece of a delta, reproducing the exact\n       * per-piece pipeline (sliced delta → chunk → enrichment → stream-limit\n       * link). `indicesSnapshot` is the arrival-time copy of the seen block\n       * indices, so lazily emitted pieces observe the same enrichment\n       * decisions the delta saw when it arrived.\n       */\n      const buildDeltaEmission = (\n        contentBlockDelta: ContentBlockDeltaEvent,\n        token: string,\n        indicesSnapshot: Set<number>\n      ): BedrockEmittedChunk => {\n        const delta = contentBlockDelta.delta;\n        const text = delta?.text;\n        const reasoningContent = delta?.reasoningContent;\n        const reasoningText = reasoningContent?.text;\n\n        let splitDelta = contentBlockDelta;\n        if (typeof text === 'string') {\n          splitDelta = {\n            ...contentBlockDelta,\n            delta: { text: token },\n          };\n        } else if (\n          typeof reasoningText === 'string' &&\n          reasoningContent != null\n        ) {\n          splitDelta = {\n            ...contentBlockDelta,\n            delta: {\n              reasoningContent: {\n                ...reasoningContent,\n                text: token,\n              },\n            },\n          };\n        }\n\n        const deltaChunk = handleConverseStreamContentBlockDelta(splitDelta);\n        const enrichedChunk = this.enrichChunk(deltaChunk, indicesSnapshot);\n        if (enrichedChunk !== deltaChunk) {\n          /** The callback copy is the same emission as the enriched yield;\n           * without the link, stream-limit accounting charges both message\n           * objects and Bedrock tool arguments falsely trip near half the\n           * cap. Linked on the messages, which are what the accounting\n           * observes. */\n          linkStreamLimitCanonical(deltaChunk.message, enrichedChunk.message);\n        }\n        return {\n          chunk: enrichedChunk,\n          callbackChunk: deltaChunk,\n          callbackToken: deltaChunk.text,\n        };\n      };\n\n      const enrichChunk = this.enrichChunk.bind(this);\n      const source = (async function* (): AsyncGenerator<\n        SmoothItem<BedrockEmittedChunk>\n        > {\n        for await (const event of stream) {\n          if (event.contentBlockStart != null) {\n            const startChunk = handleConverseStreamContentBlockStart(\n              event.contentBlockStart\n            );\n            if (startChunk != null) {\n              const idx = event.contentBlockStart.contentBlockIndex;\n              if (idx != null) {\n                seenBlockIndices.add(idx);\n                if (event.contentBlockStart.start?.toolUse != null) {\n                  toolUseBlockIndices.add(idx);\n                }\n              }\n              const enrichedStart = enrichChunk(startChunk, seenBlockIndices);\n              if (enrichedStart !== startChunk) {\n                linkStreamLimitCanonical(\n                  startChunk.message,\n                  enrichedStart.message\n                );\n              }\n              yield {\n                text: '',\n                smooth: false,\n                emit: (): BedrockEmittedChunk => ({\n                  chunk: enrichedStart,\n                  callbackChunk: startChunk,\n                  callbackToken: startChunk.text,\n                }),\n              };\n            }\n          } else if (event.contentBlockDelta != null) {\n            const contentBlockDelta = event.contentBlockDelta;\n            const delta = contentBlockDelta.delta;\n            if (delta == null) {\n              throw new Error('No delta found in content block.');\n            }\n\n            const idx = contentBlockDelta.contentBlockIndex;\n            if (idx != null) {\n              seenBlockIndices.add(idx);\n            }\n\n            const visibleText = resolveVisibleText(\n              delta.text,\n              delta.reasoningContent?.text\n            );\n            const indicesSnapshot = new Set(seenBlockIndices);\n\n            if (visibleText === '') {\n              yield {\n                text: '',\n                smooth: false,\n                emit: (): BedrockEmittedChunk =>\n                  buildDeltaEmission(\n                    contentBlockDelta,\n                    visibleText,\n                    indicesSnapshot\n                  ),\n              };\n              continue;\n            }\n\n            yield {\n              text: visibleText,\n              smooth: true,\n              emit: (piece): BedrockEmittedChunk =>\n                buildDeltaEmission(\n                  contentBlockDelta,\n                  piece.text,\n                  indicesSnapshot\n                ),\n            };\n          } else if (event.metadata != null) {\n            const metadataChunk = handleConverseStreamMetadata(event.metadata, {\n              streamUsage,\n            });\n            yield {\n              text: '',\n              smooth: false,\n              emit: (): BedrockEmittedChunk => ({ chunk: metadataChunk, callbackToken: '' }),\n            };\n          } else if (event.contentBlockStop != null) {\n            const stopIdx = event.contentBlockStop.contentBlockIndex;\n            if (stopIdx != null) {\n              seenBlockIndices.add(stopIdx);\n              if (sealToolUseOnStop && toolUseBlockIndices.has(stopIdx)) {\n                const sealChunk = createConverseToolUseStopChunk(stopIdx);\n                yield {\n                  text: '',\n                  smooth: false,\n                  emit: (): BedrockEmittedChunk => ({\n                    chunk: sealChunk,\n                    callbackChunk: sealChunk,\n                    callbackToken: sealChunk.text,\n                  }),\n                };\n              }\n            }\n          } else {\n            const eventChunk = new ChatGenerationChunk({\n              text: '',\n              message: new AIMessageChunk({\n                content: '',\n                response_metadata: { ...event } as ResponseMetadata,\n              }),\n            });\n            yield {\n              text: '',\n              smooth: false,\n              emit: (): BedrockEmittedChunk => ({ chunk: eventChunk, callbackToken: '' }),\n            };\n          }\n        }\n      })();\n\n      const smoothed = smoothStream({\n        source,\n        delayMs: this._lc_stream_delay,\n        signal: options.signal,\n        abortUpstream: abortStream,\n      });\n\n      for await (const emitted of smoothed) {\n        yield emitted.chunk;\n\n        if (emitted.callbackChunk != null) {\n          await runManager?.handleLLMNewToken(\n            emitted.callbackToken,\n            undefined,\n            undefined,\n            undefined,\n            undefined,\n            { chunk: emitted.callbackChunk }\n          );\n        }\n      }\n    } finally {\n      options.signal?.removeEventListener('abort', abortStream);\n      if (!streamAbortController.signal.aborted) {\n        streamAbortController.abort();\n      }\n    }\n  }\n\n  /**\n   * Inject `index` on content blocks for proper merge behaviour, then strip\n   * `contentBlockIndex` from response_metadata to prevent `_mergeDicts` conflicts.\n   *\n   * Text string content is promoted to array form only when the stream contains\n   * multiple content block indices (e.g. reasoning at index 0, text at index 1),\n   * ensuring text merges correctly with the already-array accumulated content.\n   */\n  private enrichChunk(\n    chunk: ChatGenerationChunk,\n    seenBlockIndices: Set<number>\n  ): ChatGenerationChunk {\n    const message = chunk.message;\n    if (!(message instanceof AIMessageChunk)) {\n      return chunk;\n    }\n\n    const metadata = message.response_metadata as Record<string, unknown>;\n    const blockIndex = this.extractContentBlockIndex(metadata);\n    const hasMetadataIndex = blockIndex != null;\n\n    let content: AIMessageChunk['content'] = message.content;\n    let contentModified = false;\n\n    if (Array.isArray(content) && blockIndex != null) {\n      content = content.map((block) =>\n        typeof block === 'object' && !('index' in block)\n          ? { ...block, index: blockIndex }\n          : block\n      );\n      contentModified = true;\n    } else if (\n      typeof content === 'string' &&\n      content !== '' &&\n      blockIndex != null &&\n      seenBlockIndices.size > 1\n    ) {\n      content = [{ type: 'text', text: content, index: blockIndex }];\n      contentModified = true;\n    }\n\n    if (!contentModified && !hasMetadataIndex) {\n      return chunk;\n    }\n\n    const cleanedMetadata = hasMetadataIndex\n      ? (this.removeContentBlockIndex(metadata) as Record<string, unknown>)\n      : metadata;\n\n    return new ChatGenerationChunk({\n      text: chunk.text,\n      message: new AIMessageChunk({\n        ...message,\n        content,\n        response_metadata: cleanedMetadata,\n      }),\n      generationInfo: chunk.generationInfo,\n    });\n  }\n\n  /**\n   * Extract `contentBlockIndex` from the top level of response_metadata.\n   * Our custom handlers always place it at the top level.\n   */\n  private extractContentBlockIndex(\n    metadata: Record<string, unknown>\n  ): number | undefined {\n    if (\n      'contentBlockIndex' in metadata &&\n      typeof metadata.contentBlockIndex === 'number'\n    ) {\n      return metadata.contentBlockIndex;\n    }\n    return undefined;\n  }\n\n  private removeContentBlockIndex(obj: unknown): unknown {\n    if (obj === null || obj === undefined) {\n      return obj;\n    }\n\n    if (Array.isArray(obj)) {\n      return obj.map((item) => this.removeContentBlockIndex(item));\n    }\n\n    if (typeof obj === 'object') {\n      const cleaned: Record<string, unknown> = {};\n      for (const [key, value] of Object.entries(obj)) {\n        if (key !== 'contentBlockIndex') {\n          cleaned[key] = this.removeContentBlockIndex(value);\n        }\n      }\n      return cleaned;\n    }\n\n    return obj;\n  }\n}\n\nexport type { ChatBedrockConverseInput };\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA0EA,SAAS,mBAAmB,MAAe,eAAgC;CACzE,IAAI,OAAO,SAAS,UAClB,OAAO;CAGT,IAAI,OAAO,kBAAkB,UAC3B,OAAO;CAGT,OAAO;AACT;AAoEA,IAAa,4BAAb,cAA+CA,eAAAA,oBAAoB;CACjE;;;;CAKA;;;;CAKA;;;;CAKA;;;;CAKA;;;;;;;CAQA;CAEA,YAAY,QAAyC;EACnD,MAAM,MAAM;EACZ,KAAK,cAAc,QAAQ;EAC3B,KAAK,iBAAiB,QAAQ;EAC9B,KAAK,mBAAmBC,iBAAAA,mBAAmB,QAAQ,gBAAgB;EACnE,KAAK,8BAA8B,QAAQ;EAC3C,KAAK,cAAc,QAAQ;EAI3B,KAAK,eAAe,QAAQ,SAAS,KAAK;CAC5C;CAEA,OAAO,UAAkB;EACvB,OAAO;CACT;;;;;CAMA,aAA+B;EAC7B,OAAO,KAAK,+BAA+B,KAAK;CAClD;;;;CAKA,iBACE,SAGA;EACA,MAAM,aAAa,MAAM,iBAAiB,OAAO;EACjD,MAAM,aACJ,KAAK,gBAAgB,QAAQC,cAAAA,yBAAyB,KAAK,YAAY,IACnEC,kBAAAA,4BACA,WAAW,YACX,MACAC,cAAAA,6BAA6B,KAAK,gBAAgB,KAAK,YAAY,CACrE,IACE,WAAW;;EAGjB,MAAM,kBAAkB,SAAS,eAAe,KAAK;EAErD,OAAO;GACL,GAAG;GACH;GACA,aAAa,kBAAkB,EAAE,MAAM,gBAAgB,IAAI,KAAA;EAC7D;CACF;;;;;CAMA,MAAe,sBACb,UACA,SACA,YACqB;EACrB,MAAM,gBAAgB,KAAK;EAC3B,IACE,KAAK,+BAA+B,QACpC,KAAK,gCAAgC,IAErC,KAAK,QAAQ,KAAK;EAGpB,IAAI;GACF,OAAO,MAAM,MAAM,sBAAsB,UAAU,SAAS,UAAU;EACxE,UAAU;GACR,KAAK,QAAQ;EACf;CACF;;;;;;;;CASA,OAAgB,sBACd,UACA,SACA,YACqC;EACrC,MAAM,EAAE,kBAAkB,mBACxBC,uBAAAA,0BAA0B,QAAQ;EACpC,MAAM,SAAS,KAAK,iBAAiB,OAAO;EAE5C,IAAI,EAAE,gBAAgB;EACtB,IAAK,QAAoC,gBAAgB,KAAA,GACvD,cAAe,QAAoC;EAGrD,MAAM,UAAU,KAAK,WAAW;EAEhC,oBAAA,kCAAkC;GAChC,cAAc,QAAQ;GACtB,QAAQ;GACR,UAAU;GACV;GACA;EACF,CAAC;EAED,MAAM,UAAU,IAAIC,gCAAAA,sBAAsB;GACxC;GACA,UAAU;GACV,QAAQ;GACR,GAAI;EACN,CAAC;EAED,MAAM,wBAAwB,IAAI,gBAAgB;EAClD,MAAM,oBAA0B,sBAAsB,MAAM;EAC5D,QAAQ,QAAQ,iBAAiB,SAAS,aAAa,EAAE,MAAM,KAAK,CAAC;EACrE,IAAIC,iBAAAA,gBAAgB,QAAQ,MAAM,GAChC,YAAY;EAGd,IAAI;GAKF,MAAM,UACJ,MALqB,KAAK,OAAO,KAAK,SAAS,EAC/C,aAAa,sBAAsB,OACrC,CAAC,EAAA,CAGU;GACX,IAAI,CAAC,QACH;GAGF,MAAM,mCAAmB,IAAI,IAAY;GACzC,MAAM,sCAAsB,IAAI,IAAY;;;;;;;GAQ5C,MAAM,oBACJ,QAAQ,mBAAmB,QAAQ,KAAK,mBAAmB;;;;;;;;GAS7D,MAAM,sBACJ,mBACA,OACA,oBACwB;IACxB,MAAM,QAAQ,kBAAkB;IAChC,MAAM,OAAO,OAAO;IACpB,MAAM,mBAAmB,OAAO;IAChC,MAAM,gBAAgB,kBAAkB;IAExC,IAAI,aAAa;IACjB,IAAI,OAAO,SAAS,UAClB,aAAa;KACX,GAAG;KACH,OAAO,EAAE,MAAM,MAAM;IACvB;SACK,IACL,OAAO,kBAAkB,YACzB,oBAAoB,MAEpB,aAAa;KACX,GAAG;KACH,OAAO,EACL,kBAAkB;MAChB,GAAG;MACH,MAAM;KACR,EACF;IACF;IAGF,MAAM,aAAaC,wBAAAA,sCAAsC,UAAU;IACnE,MAAM,gBAAgB,KAAK,YAAY,YAAY,eAAe;IAClE,IAAI,kBAAkB;;;;;;IAMpB,qBAAA,yBAAyB,WAAW,SAAS,cAAc,OAAO;IAEpE,OAAO;KACL,OAAO;KACP,eAAe;KACf,eAAe,WAAW;IAC5B;GACF;GAEA,MAAM,cAAc,KAAK,YAAY,KAAK,IAAI;GAuH9C,MAAM,WAAWC,iBAAAA,aAAa;IAC5B,SAvHc,mBAEZ;KACF,WAAW,MAAM,SAAS,QACxB,IAAI,MAAM,qBAAqB,MAAM;MACnC,MAAM,aAAaC,wBAAAA,sCACjB,MAAM,iBACR;MACA,IAAI,cAAc,MAAM;OACtB,MAAM,MAAM,MAAM,kBAAkB;OACpC,IAAI,OAAO,MAAM;QACf,iBAAiB,IAAI,GAAG;QACxB,IAAI,MAAM,kBAAkB,OAAO,WAAW,MAC5C,oBAAoB,IAAI,GAAG;OAE/B;OACA,MAAM,gBAAgB,YAAY,YAAY,gBAAgB;OAC9D,IAAI,kBAAkB,YACpB,qBAAA,yBACE,WAAW,SACX,cAAc,OAChB;OAEF,MAAM;QACJ,MAAM;QACN,QAAQ;QACR,aAAkC;SAChC,OAAO;SACP,eAAe;SACf,eAAe,WAAW;QAC5B;OACF;MACF;KACF,OAAO,IAAI,MAAM,qBAAqB,MAAM;MAC1C,MAAM,oBAAoB,MAAM;MAChC,MAAM,QAAQ,kBAAkB;MAChC,IAAI,SAAS,MACX,MAAM,IAAI,MAAM,kCAAkC;MAGpD,MAAM,MAAM,kBAAkB;MAC9B,IAAI,OAAO,MACT,iBAAiB,IAAI,GAAG;MAG1B,MAAM,cAAc,mBAClB,MAAM,MACN,MAAM,kBAAkB,IAC1B;MACA,MAAM,kBAAkB,IAAI,IAAI,gBAAgB;MAEhD,IAAI,gBAAgB,IAAI;OACtB,MAAM;QACJ,MAAM;QACN,QAAQ;QACR,YACE,mBACE,mBACA,aACA,eACF;OACJ;OACA;MACF;MAEA,MAAM;OACJ,MAAM;OACN,QAAQ;OACR,OAAO,UACL,mBACE,mBACA,MAAM,MACN,eACF;MACJ;KACF,OAAO,IAAI,MAAM,YAAY,MAAM;MACjC,MAAM,gBAAgBC,wBAAAA,6BAA6B,MAAM,UAAU,EACjE,YACF,CAAC;MACD,MAAM;OACJ,MAAM;OACN,QAAQ;OACR,aAAkC;QAAE,OAAO;QAAe,eAAe;OAAG;MAC9E;KACF,OAAO,IAAI,MAAM,oBAAoB,MAAM;MACzC,MAAM,UAAU,MAAM,iBAAiB;MACvC,IAAI,WAAW,MAAM;OACnB,iBAAiB,IAAI,OAAO;OAC5B,IAAI,qBAAqB,oBAAoB,IAAI,OAAO,GAAG;QACzD,MAAM,YAAYC,wBAAAA,+BAA+B,OAAO;QACxD,MAAM;SACJ,MAAM;SACN,QAAQ;SACR,aAAkC;UAChC,OAAO;UACP,eAAe;UACf,eAAe,UAAU;SAC3B;QACF;OACF;MACF;KACF,OAAO;MACL,MAAM,aAAa,IAAIC,wBAAAA,oBAAoB;OACzC,MAAM;OACN,SAAS,IAAIC,yBAAAA,eAAe;QAC1B,SAAS;QACT,mBAAmB,EAAE,GAAG,MAAM;OAChC,CAAC;MACH,CAAC;MACD,MAAM;OACJ,MAAM;OACN,QAAQ;OACR,aAAkC;QAAE,OAAO;QAAY,eAAe;OAAG;MAC3E;KACF;IAEJ,EAAA,CAGO;IACL,SAAS,KAAK;IACd,QAAQ,QAAQ;IAChB,eAAe;GACjB,CAAC;GAED,WAAW,MAAM,WAAW,UAAU;IACpC,MAAM,QAAQ;IAEd,IAAI,QAAQ,iBAAiB,MAC3B,MAAM,YAAY,kBAChB,QAAQ,eACR,KAAA,GACA,KAAA,GACA,KAAA,GACA,KAAA,GACA,EAAE,OAAO,QAAQ,cAAc,CACjC;GAEJ;EACF,UAAU;GACR,QAAQ,QAAQ,oBAAoB,SAAS,WAAW;GACxD,IAAI,CAAC,sBAAsB,OAAO,SAChC,sBAAsB,MAAM;EAEhC;CACF;;;;;;;;;CAUA,YACE,OACA,kBACqB;EACrB,MAAM,UAAU,MAAM;EACtB,IAAI,EAAE,mBAAmBA,yBAAAA,iBACvB,OAAO;EAGT,MAAM,WAAW,QAAQ;EACzB,MAAM,aAAa,KAAK,yBAAyB,QAAQ;EACzD,MAAM,mBAAmB,cAAc;EAEvC,IAAI,UAAqC,QAAQ;EACjD,IAAI,kBAAkB;EAEtB,IAAI,MAAM,QAAQ,OAAO,KAAK,cAAc,MAAM;GAChD,UAAU,QAAQ,KAAK,UACrB,OAAO,UAAU,YAAY,EAAE,WAAW,SACtC;IAAE,GAAG;IAAO,OAAO;GAAW,IAC9B,KACN;GACA,kBAAkB;EACpB,OAAO,IACL,OAAO,YAAY,YACnB,YAAY,MACZ,cAAc,QACd,iBAAiB,OAAO,GACxB;GACA,UAAU,CAAC;IAAE,MAAM;IAAQ,MAAM;IAAS,OAAO;GAAW,CAAC;GAC7D,kBAAkB;EACpB;EAEA,IAAI,CAAC,mBAAmB,CAAC,kBACvB,OAAO;EAGT,MAAM,kBAAkB,mBACnB,KAAK,wBAAwB,QAAQ,IACtC;EAEJ,OAAO,IAAID,wBAAAA,oBAAoB;GAC7B,MAAM,MAAM;GACZ,SAAS,IAAIC,yBAAAA,eAAe;IAC1B,GAAG;IACH;IACA,mBAAmB;GACrB,CAAC;GACD,gBAAgB,MAAM;EACxB,CAAC;CACH;;;;;CAMA,yBACE,UACoB;EACpB,IACE,uBAAuB,YACvB,OAAO,SAAS,sBAAsB,UAEtC,OAAO,SAAS;CAGpB;CAEA,wBAAgC,KAAuB;EACrD,IAAI,QAAQ,QAAQ,QAAQ,KAAA,GAC1B,OAAO;EAGT,IAAI,MAAM,QAAQ,GAAG,GACnB,OAAO,IAAI,KAAK,SAAS,KAAK,wBAAwB,IAAI,CAAC;EAG7D,IAAI,OAAO,QAAQ,UAAU;GAC3B,MAAM,UAAmC,CAAC;GAC1C,KAAK,MAAM,CAAC,KAAK,UAAU,OAAO,QAAQ,GAAG,GAC3C,IAAI,QAAQ,qBACV,QAAQ,OAAO,KAAK,wBAAwB,KAAK;GAGrD,OAAO;EACT;EAEA,OAAO;CACT;AACF"}

package/dist/cjs/llm/google/index.cjs

@@ -1,3 +1,5 @@
+const require_smoother = require("../stream/smoother.cjs");
+const require_chunkAdapters = require("../stream/chunkAdapters.cjs");
 const require_common = require("./utils/common.cjs");
 let _langchain_core_messages = require("@langchain/core/messages");
 let _langchain_core_outputs = require("@langchain/core/outputs");
@@ -6,6 +8,7 @@ let _langchain_core_utils_env = require("@langchain/core/utils/env");
 let _google_generative_ai = require("@google/generative-ai");
 //#region src/llm/google/index.ts
 var CustomChatGoogleGenerativeAI = class extends _langchain_google_genai.ChatGoogleGenerativeAI {
+	_lc_stream_delay;
 	thinkingConfig;
 	includeServerSideToolInvocations;
 	/**
@@ -16,6 +19,7 @@ var CustomChatGoogleGenerativeAI = class extends _langchain_google_genai.ChatGoo
 	}
 	constructor(fields) {
 		super(fields);
+		this._lc_stream_delay = require_smoother.resolveStreamDelay(fields._lc_stream_delay);
 		this.model = fields.model.replace(/^models\//, "");
 		this.maxOutputTokens = fields.maxOutputTokens ?? this.maxOutputTokens;
 		if (this.maxOutputTokens != null && this.maxOutputTokens < 0) throw new Error("`maxOutputTokens` must be a positive integer");
@@ -136,6 +140,14 @@ var CustomChatGoogleGenerativeAI = class extends _langchain_google_genai.ChatGoo
 		return generationResult;
 	}
 	async *_streamResponseChunks(messages, options, runManager) {
+		yield* require_chunkAdapters.smoothGenerationChunks({
+			chunks: this._streamProviderChunks(messages, options),
+			delayMs: this._lc_stream_delay,
+			signal: options.signal,
+			runManager
+		});
+	}
+	async *_streamProviderChunks(messages, options) {
 		const prompt = require_common.convertBaseMessagesToContent(messages, this._isMultimodalModel, this.useSystemInstruction, this.model);
 		let actualPrompt = prompt;
 		if (prompt?.[0].role === "system") {
@@ -165,19 +177,14 @@ var CustomChatGoogleGenerativeAI = class extends _langchain_google_genai.ChatGoo
 			index += 1;
 			if (!chunk) continue;
 			yield chunk;
-			await runManager?.handleLLMNewToken(chunk.text || "", void 0, void 0, void 0, void 0, { chunk });
-		}
-		if (lastUsageMetadata) {
-			const finalChunk = new _langchain_core_outputs.ChatGenerationChunk({
-				text: "",
-				message: new _langchain_core_messages.AIMessageChunk({
-					content: "",
-					usage_metadata: lastUsageMetadata
-				})
-			});
-			yield finalChunk;
-			await runManager?.handleLLMNewToken(finalChunk.text || "", void 0, void 0, void 0, void 0, { chunk: finalChunk });
 		}
+		if (lastUsageMetadata) yield new _langchain_core_outputs.ChatGenerationChunk({
+			text: "",
+			message: new _langchain_core_messages.AIMessageChunk({
+				content: "",
+				usage_metadata: lastUsageMetadata
+			})
+		});
 	}
 };
 //#endregion