Skip to content

An AI-powered search engine with a generative UI

refactor: replace createUIMessageStream + writer.merge with toUIMessageStreamResponse (#805)

Yoshiki Miura37b0c1ca01c4
diff --git a/lib/streaming/create-chat-stream-response.ts b/lib/streaming/create-chat-stream-response.ts
index 2259e6b8..867341ba 100644
--- a/lib/streaming/create-chat-stream-response.ts
+++ b/lib/streaming/create-chat-stream-response.ts
@@ -1,12 +1,8 @@
import {
consumeStream,
convertToModelMessages,
- createUIMessageStream,
- createUIMessageStreamResponse,
pruneMessages,
- smoothStream,
- UIMessage,
- UIMessageStreamWriter
+ smoothStream
} from 'ai'
import { randomUUID } from 'crypto'
import { Langfuse } from 'langfuse'
@@ -106,141 +102,134 @@ export async function createChatStreamResponse(
trigger,
initialChat,
abortSignal,
- parentTraceId, // Add parent trace ID to context
+ parentTraceId,
isNewChat
}
// Declare titlePromise in outer scope for onFinish access
let titlePromise: Promise<string> | undefined
- // Create the stream
- const stream = createUIMessageStream<UIMessage>({
- execute: async ({ writer }: { writer: UIMessageStreamWriter }) => {
- try {
- // Prepare messages for the model
- const prepareStart = performance.now()
- perfLog(
- `prepareMessages - Invoked: trigger=${trigger}, isNewChat=${isNewChat}`
- )
- const messagesToModel = await prepareMessages(context, message)
- perfTime('prepareMessages completed (stream)', prepareStart)
-
- // Get the researcher agent with parent trace ID and search mode.
- const researchAgent = researcher({
- model: context.modelId,
- modelConfig: model,
- parentTraceId,
- searchMode
- })
-
- // For OpenAI models, strip reasoning parts from UIMessages before conversion
- // OpenAI's Responses API requires reasoning items and their following items to be kept together
- // See: https://github.com/vercel/ai/issues/11036
- const isOpenAI = context.modelId.startsWith('openai:')
- const messagesWithoutSpec = stripSpecFromMessages(messagesToModel)
- const messagesToConvert = isOpenAI
- ? stripReasoningParts(messagesWithoutSpec)
- : messagesWithoutSpec
-
- // Convert to model messages and apply context window management
- let modelMessages = await convertToModelMessages(messagesToConvert)
-
- // Prune messages to reduce token usage while keeping recent context
- modelMessages = pruneMessages({
- messages: modelMessages,
- reasoning: 'before-last-message',
- toolCalls: 'before-last-2-messages',
- emptyMessages: 'remove'
- })
-
- if (shouldTruncateMessages(modelMessages, model)) {
- const maxTokens = getMaxAllowedTokens(model)
- const originalCount = modelMessages.length
- modelMessages = truncateMessages(modelMessages, maxTokens, model.id)
-
- if (process.env.NODE_ENV === 'development') {
- console.log(
- `Context window limit reached. Truncating from ${originalCount} to ${modelMessages.length} messages`
- )
- }
- }
+ try {
+ // Prepare messages for the model
+ const prepareStart = performance.now()
+ perfLog(
+ `prepareMessages - Invoked: trigger=${trigger}, isNewChat=${isNewChat}`
+ )
+ const messagesToModel = await prepareMessages(context, message)
+ perfTime('prepareMessages completed (stream)', prepareStart)
+
+ // Get the researcher agent with parent trace ID and search mode.
+ const researchAgent = researcher({
+ model: context.modelId,
+ modelConfig: model,
+ parentTraceId,
+ searchMode
+ })
- // Start title generation in parallel if it's a new chat
- if (!initialChat && message) {
- const userContent = getTextFromParts(message.parts)
- titlePromise = generateChatTitle({
- userMessageContent: userContent,
- modelId: context.modelId,
- abortSignal,
- parentTraceId
- }).catch(error => {
- console.error('Error generating title:', error)
- return DEFAULT_CHAT_TITLE
- })
- }
+ // For OpenAI models, strip reasoning parts from UIMessages before conversion
+ // OpenAI's Responses API requires reasoning items and their following items to be kept together
+ // See: https://github.com/vercel/ai/issues/11036
+ const isOpenAI = context.modelId.startsWith('openai:')
+ const messagesWithoutSpec = stripSpecFromMessages(messagesToModel)
+ const messagesToConvert = isOpenAI
+ ? stripReasoningParts(messagesWithoutSpec)
+ : messagesWithoutSpec
+
+ // Convert to model messages and apply context window management
+ let modelMessages = await convertToModelMessages(messagesToConvert)
+
+ // Prune messages to reduce token usage while keeping recent context
+ modelMessages = pruneMessages({
+ messages: modelMessages,
+ reasoning: 'before-last-message',
+ toolCalls: 'before-last-2-messages',
+ emptyMessages: 'remove'
+ })
- const llmStart = performance.now()
- perfLog(
- `researchAgent.stream - Start: model=${context.modelId}, searchMode=${searchMode}`
- )
- const result = await researchAgent.stream({
- messages: modelMessages,
- abortSignal,
- experimental_transform: smoothStream({ chunking: 'word' })
- })
- result.consumeStream()
- // Stream with the research agent, including metadata
- writer.merge(
- result.toUIMessageStream({
- messageMetadata: ({ part }) => {
- // Send metadata when streaming starts
- if (part.type === 'start') {
- return {
- traceId: parentTraceId,
- searchMode,
- modelId: context.modelId
- }
- }
- }
- })
- )
+ if (shouldTruncateMessages(modelMessages, model)) {
+ const maxTokens = getMaxAllowedTokens(model)
+ const originalCount = modelMessages.length
+ modelMessages = truncateMessages(modelMessages, maxTokens, model.id)
- await result.response
- perfTime('researchAgent.stream completed', llmStart)
- } catch (error) {
- console.error('Stream execution error:', error)
- throw error // This error will be handled by the onError callback
- } finally {
- // Flush Langfuse traces if enabled
- if (langfuse) {
- await langfuse.flushAsync()
- }
+ if (process.env.NODE_ENV === 'development') {
+ console.log(
+ `Context window limit reached. Truncating from ${originalCount} to ${modelMessages.length} messages`
+ )
}
- },
- onError: (error: any) => {
- // console.error('Stream error:', error)
- return error instanceof Error ? error.message : String(error)
- },
- onFinish: async ({ responseMessage, isAborted }) => {
- if (isAborted || !responseMessage) return
-
- // Persist stream results to database
- await persistStreamResults(
- responseMessage,
- chatId,
- userId,
- titlePromise,
- parentTraceId,
- searchMode,
- context.modelId,
- context.pendingInitialSave,
- context.pendingInitialUserMessage
- )
}
- })
- return createUIMessageStreamResponse({
- stream,
- consumeSseStream: consumeStream
- })
+ // Start title generation in parallel if it's a new chat
+ if (!initialChat && message) {
+ const userContent = getTextFromParts(message.parts)
+ titlePromise = generateChatTitle({
+ userMessageContent: userContent,
+ modelId: context.modelId,
+ abortSignal,
+ parentTraceId
+ }).catch(error => {
+ console.error('Error generating title:', error)
+ return DEFAULT_CHAT_TITLE
+ })
+ }
+
+ const llmStart = performance.now()
+ perfLog(
+ `researchAgent.stream - Start: model=${context.modelId}, searchMode=${searchMode}`
+ )
+ const result = await researchAgent.stream({
+ messages: modelMessages,
+ abortSignal,
+ experimental_transform: smoothStream({ chunking: 'word' })
+ })
+ result.consumeStream()
+
+ return result.toUIMessageStreamResponse({
+ messageMetadata: ({ part }) => {
+ if (part.type === 'start') {
+ return {
+ traceId: parentTraceId,
+ searchMode,
+ modelId: context.modelId
+ }
+ }
+ },
+ onFinish: async ({ responseMessage, isAborted }) => {
+ try {
+ perfTime('researchAgent.stream completed', llmStart)
+ if (isAborted || !responseMessage) return
+
+ // Persist stream results to database
+ await persistStreamResults(
+ responseMessage,
+ chatId,
+ userId,
+ titlePromise,
+ parentTraceId,
+ searchMode,
+ context.modelId,
+ context.pendingInitialSave,
+ context.pendingInitialUserMessage
+ )
+ } finally {
+ if (langfuse) {
+ await langfuse.flushAsync()
+ }
+ }
+ },
+ onError: (error: unknown) => {
+ return error instanceof Error ? error.message : String(error)
+ },
+ consumeSseStream: consumeStream
+ })
+ } catch (error) {
+ if (langfuse) {
+ await langfuse.flushAsync()
+ }
+ console.error('Stream execution error:', error)
+ const errorMessage = error instanceof Error ? error.message : String(error)
+ return new Response(errorMessage, {
+ status: 500,
+ statusText: 'Internal Server Error'
+ })
+ }
}
diff --git a/lib/streaming/create-ephemeral-chat-stream-response.ts b/lib/streaming/create-ephemeral-chat-stream-response.ts
index fc525fe1..80edf300 100644
--- a/lib/streaming/create-ephemeral-chat-stream-response.ts
+++ b/lib/streaming/create-ephemeral-chat-stream-response.ts
@@ -1,12 +1,9 @@
+import type { UIMessage } from 'ai'
import {
consumeStream,
convertToModelMessages,
- createUIMessageStream,
- createUIMessageStreamResponse,
pruneMessages,
- smoothStream,
- UIMessage,
- UIMessageStreamWriter
+ smoothStream
} from 'ai'
import { randomUUID } from 'crypto'
import { Langfuse } from 'langfuse'
@@ -64,70 +61,69 @@ export async function createEphemeralChatStreamResponse(
})
}
- const stream = createUIMessageStream<UIMessage>({
- execute: async ({ writer }: { writer: UIMessageStreamWriter }) => {
- try {
- const isOpenAI = `${model.providerId}:${model.id}`.startsWith('openai:')
- const messagesWithoutSpec = stripSpecFromMessages(messages)
- const messagesToConvert = isOpenAI
- ? stripReasoningParts(messagesWithoutSpec)
- : messagesWithoutSpec
-
- let modelMessages = await convertToModelMessages(messagesToConvert)
-
- modelMessages = pruneMessages({
- messages: modelMessages,
- reasoning: 'before-last-message',
- toolCalls: 'before-last-2-messages',
- emptyMessages: 'remove'
- })
-
- if (shouldTruncateMessages(modelMessages, model)) {
- const maxTokens = getMaxAllowedTokens(model)
- modelMessages = truncateMessages(modelMessages, maxTokens, model.id)
- }
+ try {
+ const isOpenAI = `${model.providerId}:${model.id}`.startsWith('openai:')
+ const messagesWithoutSpec = stripSpecFromMessages(messages)
+ const messagesToConvert = isOpenAI
+ ? stripReasoningParts(messagesWithoutSpec)
+ : messagesWithoutSpec
+
+ let modelMessages = await convertToModelMessages(messagesToConvert)
+
+ modelMessages = pruneMessages({
+ messages: modelMessages,
+ reasoning: 'before-last-message',
+ toolCalls: 'before-last-2-messages',
+ emptyMessages: 'remove'
+ })
+
+ if (shouldTruncateMessages(modelMessages, model)) {
+ const maxTokens = getMaxAllowedTokens(model)
+ modelMessages = truncateMessages(modelMessages, maxTokens, model.id)
+ }
+
+ const researchAgent = researcher({
+ model: `${model.providerId}:${model.id}`,
+ modelConfig: model,
+ parentTraceId,
+ searchMode
+ })
- const researchAgent = researcher({
- model: `${model.providerId}:${model.id}`,
- modelConfig: model,
- parentTraceId,
- searchMode
- })
-
- const result = await researchAgent.stream({
- messages: modelMessages,
- abortSignal,
- experimental_transform: smoothStream({ chunking: 'word' })
- })
- result.consumeStream()
- writer.merge(
- result.toUIMessageStream({
- messageMetadata: ({ part }) => {
- if (part.type === 'start') {
- return {
- traceId: parentTraceId,
- searchMode,
- modelId: `${model.providerId}:${model.id}`
- }
- }
- }
- })
- )
-
- await result.response
- } finally {
+ const result = await researchAgent.stream({
+ messages: modelMessages,
+ abortSignal,
+ experimental_transform: smoothStream({ chunking: 'word' })
+ })
+ result.consumeStream()
+
+ return result.toUIMessageStreamResponse({
+ messageMetadata: ({ part }) => {
+ if (part.type === 'start') {
+ return {
+ traceId: parentTraceId,
+ searchMode,
+ modelId: `${model.providerId}:${model.id}`
+ }
+ }
+ },
+ onFinish: async () => {
if (langfuse) {
await langfuse.flushAsync()
}
- }
- },
- onError: (error: any) => {
- return error instanceof Error ? error.message : String(error)
+ },
+ onError: (error: unknown) => {
+ return error instanceof Error ? error.message : String(error)
+ },
+ consumeSseStream: consumeStream
+ })
+ } catch (error) {
+ if (langfuse) {
+ await langfuse.flushAsync()
}
- })
-
- return createUIMessageStreamResponse({
- stream,
- consumeSseStream: consumeStream
- })
+ const message = error instanceof Error ? error.message : String(error)
+ return new Response(message, {
+ status: 500,
+ statusText: 'Internal Server Error'
+ })
+ }
}