Skip to content

An AI-powered search engine with a generative UI

fix: strip reasoning parts for OpenAI models to fix multi-turn errors

Yoshiki Miurad6fa6f7160d2
diff --git a/lib/streaming/create-chat-stream-response.ts b/lib/streaming/create-chat-stream-response.ts
index 29dd4360..a3e08997 100644
--- a/lib/streaming/create-chat-stream-response.ts
+++ b/lib/streaming/create-chat-stream-response.ts
@@ -27,6 +27,7 @@ import { perfLog, perfTime } from '../utils/perf-logging'
import { persistStreamResults } from './helpers/persist-stream-results'
import { prepareMessages } from './helpers/prepare-messages'
import { streamRelatedQuestions } from './helpers/stream-related-questions'
+import { stripReasoningParts } from './helpers/strip-reasoning-parts'
import type { StreamContext } from './helpers/types'
import { BaseStreamConfig } from './types'
@@ -135,18 +136,23 @@ export async function createChatStreamResponse(
modelType
})
+ // 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 messagesToConvert = isOpenAI
+ ? stripReasoningParts(messagesToModel)
+ : messagesToModel
+
// Convert to model messages and apply context window management
- let modelMessages = await convertToModelMessages(messagesToModel)
+ let modelMessages = await convertToModelMessages(messagesToConvert)
- // Prune messages to remove unnecessary tool calls and empty messages
- // This reduces token usage while keeping recent context
- // Note: reasoning is kept ('none') for OpenAI reasoning model compatibility
- // (removing reasoning without its paired tool-call causes API errors)
+ // Prune messages to reduce token usage while keeping recent context
modelMessages = pruneMessages({
messages: modelMessages,
- reasoning: 'none', // Keep all reasoning for OpenAI compatibility
- toolCalls: 'before-last-2-messages', // Keep recent tool calls (last 2 messages)
- emptyMessages: 'remove' // Remove messages that become empty after pruning
+ reasoning: 'before-last-message',
+ toolCalls: 'before-last-2-messages',
+ emptyMessages: 'remove'
})
if (shouldTruncateMessages(modelMessages, model)) {
diff --git a/lib/streaming/helpers/strip-reasoning-parts.ts b/lib/streaming/helpers/strip-reasoning-parts.ts
new file mode 100644
index 00000000..4856cb5a
--- /dev/null
+++ b/lib/streaming/helpers/strip-reasoning-parts.ts
@@ -0,0 +1,30 @@
+import { UIMessage } from 'ai'
+
+/**
+ * Strips reasoning parts from UIMessages for OpenAI models.
+ *
+ * OpenAI's Responses API requires reasoning items and their following items
+ * (tool-calls or text) to be kept together. The AI SDK's convertToModelMessages
+ * doesn't properly handle these requirements, causing errors like:
+ * "Item 'rs_...' of type 'reasoning' was provided without its required following item"
+ *
+ * By stripping reasoning parts before conversion, we avoid this compatibility issue.
+ *
+ * @see https://github.com/vercel/ai/issues/11036
+ */
+export function stripReasoningParts(messages: UIMessage[]): UIMessage[] {
+ return messages.map(msg => {
+ if (msg.role !== 'assistant' || !msg.parts) {
+ return msg
+ }
+
+ const filteredParts = msg.parts.filter(part => part.type !== 'reasoning')
+
+ // If all parts were reasoning, keep the original message
+ if (filteredParts.length === 0) {
+ return msg
+ }
+
+ return { ...msg, parts: filteredParts }
+ })
+}