Skip to content

An AI-powered search engine with a generative UI

feat: integrate pruneMessages for optimized context window management

Yoshiki Miurae95eed00cce1
diff --git a/bun.lock b/bun.lock
index 40551230..35f6496e 100644
--- a/bun.lock
+++ b/bun.lock
@@ -4,11 +4,11 @@
"": {
"name": "morphic",
"dependencies": {
- "@ai-sdk/anthropic": "^2.0.15",
- "@ai-sdk/gateway": "^1.0.20",
- "@ai-sdk/google": "^2.0.13",
- "@ai-sdk/openai": "^2.0.28",
- "@ai-sdk/react": "^2.0.39",
+ "@ai-sdk/anthropic": "^2.0.44",
+ "@ai-sdk/gateway": "^2.0.9",
+ "@ai-sdk/google": "^2.0.31",
+ "@ai-sdk/openai": "^2.0.65",
+ "@ai-sdk/react": "^2.0.93",
"@aws-sdk/client-s3": "^3.850.0",
"@opentelemetry/api-logs": "0.57.2",
"@opentelemetry/instrumentation": "0.57.2",
@@ -39,7 +39,7 @@
"@upstash/redis": "^1.35.5",
"@vercel/analytics": "^1.5.0",
"@vercel/otel": "^1.13.0",
- "ai": "^5.0.39",
+ "ai": "^5.0.93",
"class-variance-authority": "^0.7.1",
"clsx": "^2.1.0",
"cmdk": "1.0.0",
@@ -97,19 +97,19 @@
"packages": {
"@adobe/css-tools": ["@adobe/[email protected]", "", {}, "sha512-VQKMkwriZbaOgVCby1UDY/LDk5fIjhQicCvVPFqfe+69fWaPWydbWJ3wRt59/YzIwda1I81loas3oCoHxnqvdA=="],
- "@ai-sdk/anthropic": ["@ai-sdk/[email protected]", "", { "dependencies": { "@ai-sdk/provider": "2.0.0", "@ai-sdk/provider-utils": "3.0.8" }, "peerDependencies": { "zod": "^3.25.76 || ^4" } }, "sha512-MxNGoYvKyF7IqMU0k9gogyiJi0/ogwg6i2Baw862BMjM2KJuBcCPqh6/lrpwiDg6pqphGUc+LfjPd6PRFARnng=="],
+ "@ai-sdk/anthropic": ["@ai-sdk/[email protected]", "", { "dependencies": { "@ai-sdk/provider": "2.0.0", "@ai-sdk/provider-utils": "3.0.17" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-o8TfNXRzO/KZkBrcx+CL9LQsPhx7PHyqzUGjza3TJaF9WxfH1S5UQLAmEw8F7lQoHNLU0IX03WT8o8R/4JbUxQ=="],
- "@ai-sdk/gateway": ["@ai-sdk/[email protected]", "", { "dependencies": { "@ai-sdk/provider": "2.0.0", "@ai-sdk/provider-utils": "3.0.8" }, "peerDependencies": { "zod": "^3.25.76 || ^4" } }, "sha512-2K0kGnHyLfT1v2+3xXbLqohfWBJ/vmIh1FTWnrvZfvuUuBdOi2DMgnSQzkLvFVLyM8mhOcx+ZwU6IOOsuyOv/w=="],
+ "@ai-sdk/gateway": ["@ai-sdk/[email protected]", "", { "dependencies": { "@ai-sdk/provider": "2.0.0", "@ai-sdk/provider-utils": "3.0.17", "@vercel/oidc": "3.0.3" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-E6x4h5CPPPJ0za1r5HsLtHbeI+Tp3H+YFtcH8G3dSSPFE6w+PZINzB4NxLZmg1QqSeA5HTP3ZEzzsohp0o2GEw=="],
- "@ai-sdk/google": ["@ai-sdk/[email protected]", "", { "dependencies": { "@ai-sdk/provider": "2.0.0", "@ai-sdk/provider-utils": "3.0.8" }, "peerDependencies": { "zod": "^3.25.76 || ^4" } }, "sha512-5WauM+IrqbllWT4uXZVrfTnPCSKTtkHGNsD2CYD0JgGfeIOpa285UYCYUi0Z4RtcovwnZitvQABq465FfeLwzA=="],
+ "@ai-sdk/google": ["@ai-sdk/[email protected]", "", { "dependencies": { "@ai-sdk/provider": "2.0.0", "@ai-sdk/provider-utils": "3.0.17" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-wOlUkrXHuL73sXPZd251+30BQ378zn2Zo8pW+Hq+8d9FmSJpZXOSK8cmYle1SE5ZWe/TS+eSj3eBCZ6AW+q2EA=="],
- "@ai-sdk/openai": ["@ai-sdk/[email protected]", "", { "dependencies": { "@ai-sdk/provider": "2.0.0", "@ai-sdk/provider-utils": "3.0.8" }, "peerDependencies": { "zod": "^3.25.76 || ^4" } }, "sha512-Z2mG7PjUKbpT8fMexE6yrorxXVzGHSl3jKF293w2i6s9Dc6X81Gf6Z0OGNnkrftLtW4PXr7RZ/9xoyusBZW4uA=="],
+ "@ai-sdk/openai": ["@ai-sdk/[email protected]", "", { "dependencies": { "@ai-sdk/provider": "2.0.0", "@ai-sdk/provider-utils": "3.0.17" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-Wqo4iNCsxUYJFmEAzAHO0XDnnS76rqvPRX2AUhEAOX3cgL9UEfouFmiNQS2jM9AZhMdWj5GIG41hgr4YOr20tA=="],
"@ai-sdk/provider": ["@ai-sdk/[email protected]", "", { "dependencies": { "json-schema": "^0.4.0" } }, "sha512-6o7Y2SeO9vFKB8lArHXehNuusnpddKPk7xqL7T2/b+OvXMRIXUO1rR4wcv1hAFUAT9avGZshty3Wlua/XA7TvA=="],
- "@ai-sdk/provider-utils": ["@ai-sdk/[email protected]", "", { "dependencies": { "@ai-sdk/provider": "2.0.0", "@standard-schema/spec": "^1.0.0", "eventsource-parser": "^3.0.5" }, "peerDependencies": { "zod": "^3.25.76 || ^4" } }, "sha512-cDj1iigu7MW2tgAQeBzOiLhjHOUM9vENsgh4oAVitek0d//WdgfPCsKO3euP7m7LyO/j9a1vr/So+BGNdpFXYw=="],
+ "@ai-sdk/provider-utils": ["@ai-sdk/[email protected]", "", { "dependencies": { "@ai-sdk/provider": "2.0.0", "@standard-schema/spec": "^1.0.0", "eventsource-parser": "^3.0.6" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-TR3Gs4I3Tym4Ll+EPdzRdvo/rc8Js6c4nVhFLuvGLX/Y4V9ZcQMa/HTiYsHEgmYrf1zVi6Q145UEZUfleOwOjw=="],
- "@ai-sdk/react": ["@ai-sdk/[email protected]", "", { "dependencies": { "@ai-sdk/provider-utils": "3.0.8", "ai": "5.0.39", "swr": "^2.2.5", "throttleit": "2.1.0" }, "peerDependencies": { "react": "^18 || ^19 || ^19.0.0-rc", "zod": "^3.25.76 || ^4" }, "optionalPeers": ["zod"] }, "sha512-cBQZQG3ZtEebBPCcDVJGS4BU6e4ADgnO2PRhqvkPcPeFM68LK+x5E0VJcpLWUuoAW1Lv/yqy62LN3Ql1mFbzYQ=="],
+ "@ai-sdk/react": ["@ai-sdk/[email protected]", "", { "dependencies": { "@ai-sdk/provider-utils": "3.0.17", "ai": "5.0.93", "swr": "^2.2.5", "throttleit": "2.1.0" }, "peerDependencies": { "react": "^18 || ^19 || ^19.0.0-rc", "zod": "^3.25.76 || ^4.1.8" }, "optionalPeers": ["zod"] }, "sha512-2TzhpQr10HuWxpqyHpSAUMRUqD1G2O73J2sAaJChomVDbjr7BwpM0mdR3aRamCXNtuLiJmTFQhbNzw8fXMBdYw=="],
"@alloc/quick-lru": ["@alloc/[email protected]", "", {}, "sha512-UrcABB+4bUrFABwbluTIBErXwvbsU/V7TZWfmbgJfbkwiBuziS9gxdODUyuiecfdGQ85jglMW6juS3+z5TsKLw=="],
@@ -971,6 +971,8 @@
"@vercel/analytics": ["@vercel/[email protected]", "", { "peerDependencies": { "@remix-run/react": "^2", "@sveltejs/kit": "^1 || ^2", "next": ">= 13", "react": "^18 || ^19 || ^19.0.0-rc", "svelte": ">= 4", "vue": "^3", "vue-router": "^4" }, "optionalPeers": ["@remix-run/react", "@sveltejs/kit", "next", "react", "svelte", "vue", "vue-router"] }, "sha512-MYsBzfPki4gthY5HnYN7jgInhAZ7Ac1cYDoRWFomwGHWEX7odTEzbtg9kf/QSo7XEsEAqlQugA6gJ2WS2DEa3g=="],
+ "@vercel/oidc": ["@vercel/[email protected]", "", {}, "sha512-yNEQvPcVrK9sIe637+I0jD6leluPxzwJKx/Haw6F4H77CdDsszUn5V3o96LPziXkSNE2B83+Z3mjqGKBK/R6Gg=="],
+
"@vercel/otel": ["@vercel/[email protected]", "", { "peerDependencies": { "@opentelemetry/api": ">=1.7.0 <2.0.0", "@opentelemetry/api-logs": ">=0.46.0 <0.200.0", "@opentelemetry/instrumentation": ">=0.46.0 <0.200.0", "@opentelemetry/resources": ">=1.19.0 <2.0.0", "@opentelemetry/sdk-logs": ">=0.46.0 <0.200.0", "@opentelemetry/sdk-metrics": ">=1.19.0 <2.0.0", "@opentelemetry/sdk-trace-base": ">=1.19.0 <2.0.0" } }, "sha512-esRkt470Y2jRK1B1g7S1vkt4Csu44gp83Zpu8rIyPoqy2BKgk4z7ik1uSMswzi45UogLHFl6yR5TauDurBQi4Q=="],
"@vitejs/plugin-react": ["@vitejs/[email protected]", "", { "dependencies": { "@babel/core": "^7.28.0", "@babel/plugin-transform-react-jsx-self": "^7.27.1", "@babel/plugin-transform-react-jsx-source": "^7.27.1", "@rolldown/pluginutils": "1.0.0-beta.27", "@types/babel__core": "^7.20.5", "react-refresh": "^0.17.0" }, "peerDependencies": { "vite": "^4.2.0 || ^5.0.0 || ^6.0.0 || ^7.0.0" } }, "sha512-gUu9hwfWvvEDBBmgtAowQCojwZmJ5mcLn3aufeCsitijs3+f2NsrPtlAWIR6OPiqljl96GVCUbLe0HyqIpVaoA=="],
@@ -997,7 +999,7 @@
"agent-base": ["[email protected]", "", {}, "sha512-MnA+YT8fwfJPgBx3m60MNqakm30XOkyIoH1y6huTQvC0PwZG7ki8NacLBcrPbNoo8vEZy7Jpuk7+jMO+CUovTQ=="],
- "ai": ["[email protected]", "", { "dependencies": { "@ai-sdk/gateway": "1.0.20", "@ai-sdk/provider": "2.0.0", "@ai-sdk/provider-utils": "3.0.8", "@opentelemetry/api": "1.9.0" }, "peerDependencies": { "zod": "^3.25.76 || ^4" } }, "sha512-1AOjTHY8MUY4T/X/I+otGTbvKmMQCCGWffuVDyQ21l/2Vv/QoLZcw+ZZHVvp+wvQcPOsfXjURGSFZtin7rnghA=="],
+ "ai": ["[email protected]", "", { "dependencies": { "@ai-sdk/gateway": "2.0.9", "@ai-sdk/provider": "2.0.0", "@ai-sdk/provider-utils": "3.0.17", "@opentelemetry/api": "1.9.0" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-9eGcu+1PJgPg4pRNV4L7tLjRR3wdJC9CXQoNMvtqvYNOLZHFCzjHtVIOr2SIkoJJeu2+sOy3hyiSuTmy2MA40g=="],
"ajv": ["[email protected]", "", { "dependencies": { "fast-deep-equal": "^3.1.1", "fast-json-stable-stringify": "^2.0.0", "json-schema-traverse": "^0.4.1", "uri-js": "^4.2.2" } }, "sha512-j3fVLgvTo527anyYyJOGTYJbG+vnnQYvE0m5mmkc1TK+nxAppkCLMIL0aZ4dblVCNoGShhm+kzE4ZUykBoMg4g=="],
@@ -2411,6 +2413,8 @@
"node-fetch/whatwg-url": ["[email protected]", "", { "dependencies": { "tr46": "~0.0.3", "webidl-conversions": "^3.0.0" } }, "sha512-saE57nupxk6v3HY35+jzBwYa0rKSy0XR8JSxZPwgLr7ys0IBzhGviA1/TUGJLmSVqs8pb9AnvICXEuOHLprYTw=="],
+ "ollama-ai-provider-v2/@ai-sdk/provider-utils": ["@ai-sdk/[email protected]", "", { "dependencies": { "@ai-sdk/provider": "2.0.0", "@standard-schema/spec": "^1.0.0", "eventsource-parser": "^3.0.5" }, "peerDependencies": { "zod": "^3.25.76 || ^4" } }, "sha512-cDj1iigu7MW2tgAQeBzOiLhjHOUM9vENsgh4oAVitek0d//WdgfPCsKO3euP7m7LyO/j9a1vr/So+BGNdpFXYw=="],
+
"parse-entities/@types/unist": ["@types/[email protected]", "", {}, "sha512-CmBKiL6NNo/OqgmMn95Fk9Whlp2mtvIv+KNpQKN2F4SjvrEesubTRWGYSg+BnWZOnlCaSTU1sMpsBOzgbYhnsA=="],
"path-scurry/lru-cache": ["[email protected]", "", {}, "sha512-JNAzZcXrCt42VGLuYz0zfAzDfAvJWW6AfYlDBQyDV5DClI2m5sAmK+OIO7s59XfsRsWHp02jAJrRadPRGTt6SQ=="],
diff --git a/lib/streaming/__tests__/prune-messages-integration.test.ts b/lib/streaming/__tests__/prune-messages-integration.test.ts
new file mode 100644
index 00000000..bc9feb4b
--- /dev/null
+++ b/lib/streaming/__tests__/prune-messages-integration.test.ts
@@ -0,0 +1,135 @@
+import { describe, expect, it } from 'vitest'
+
+import { pruneMessages } from 'ai'
+import type { ModelMessage } from 'ai'
+
+describe('pruneMessages integration', () => {
+ it('should prune messages according to configuration', () => {
+ const messages: ModelMessage[] = [
+ {
+ role: 'user',
+ content: 'First question'
+ },
+ {
+ role: 'assistant',
+ content: 'Let me think about that...'
+ },
+ {
+ role: 'user',
+ content: 'Second question'
+ },
+ {
+ role: 'assistant',
+ content: 'Here is my response'
+ }
+ ]
+
+ const pruned = pruneMessages({
+ messages,
+ reasoning: 'before-last-message',
+ toolCalls: 'before-last-2-messages',
+ emptyMessages: 'remove'
+ })
+
+ // Should return pruned messages
+ expect(Array.isArray(pruned)).toBe(true)
+ expect(pruned.length).toBeGreaterThan(0)
+ expect(pruned.length).toBeLessThanOrEqual(messages.length)
+ })
+
+ it('should handle messages with text content', () => {
+ const messages: ModelMessage[] = [
+ {
+ role: 'user',
+ content: 'Question'
+ },
+ {
+ role: 'assistant',
+ content: 'Answer'
+ },
+ {
+ role: 'user',
+ content: 'Another question'
+ }
+ ]
+
+ const pruned = pruneMessages({
+ messages,
+ reasoning: 'all',
+ toolCalls: 'all',
+ emptyMessages: 'remove'
+ })
+
+ // Should keep all messages with text content
+ expect(pruned.length).toBe(messages.length)
+ expect(pruned.every((msg: ModelMessage) => msg.content)).toBe(true)
+ })
+
+ it('should preserve message structure', () => {
+ const messages: ModelMessage[] = [
+ {
+ role: 'user',
+ content: 'Question 1'
+ },
+ {
+ role: 'assistant',
+ content: 'Answer to question 1'
+ },
+ {
+ role: 'user',
+ content: 'Question 2'
+ },
+ {
+ role: 'assistant',
+ content: 'Answer to question 2'
+ }
+ ]
+
+ const pruned = pruneMessages({
+ messages,
+ reasoning: 'before-last-message',
+ toolCalls: 'none',
+ emptyMessages: 'remove'
+ })
+
+ // Last assistant message should be preserved
+ const lastAssistant = pruned
+ .slice()
+ .reverse()
+ .find((msg: ModelMessage) => msg.role === 'assistant')
+ expect(lastAssistant).toBeDefined()
+ expect(lastAssistant!.content).toBeTruthy()
+ })
+
+ it('should handle simple conversation messages', () => {
+ const messages: ModelMessage[] = [
+ {
+ role: 'user',
+ content: 'Hello'
+ },
+ {
+ role: 'assistant',
+ content: 'Hi there!'
+ },
+ {
+ role: 'user',
+ content: 'How are you?'
+ },
+ {
+ role: 'assistant',
+ content: 'I am doing well, thank you!'
+ }
+ ]
+
+ const pruned = pruneMessages({
+ messages,
+ reasoning: 'before-last-message',
+ toolCalls: 'before-last-2-messages',
+ emptyMessages: 'remove'
+ })
+
+ // Should preserve conversation flow
+ expect(pruned.length).toBeGreaterThan(0)
+ expect(pruned.every((msg) => msg.content)).toBe(true)
+ })
+})
diff --git a/lib/streaming/create-chat-stream-response.ts b/lib/streaming/create-chat-stream-response.ts
index b91b1b33..6233d355 100644
--- a/lib/streaming/create-chat-stream-response.ts
+++ b/lib/streaming/create-chat-stream-response.ts
@@ -3,6 +3,7 @@ import {
convertToModelMessages,
createUIMessageStream,
createUIMessageStreamResponse,
+ pruneMessages,
UIMessage,
UIMessageStreamWriter
} from 'ai'
@@ -22,7 +23,6 @@ import {
import { getTextFromParts } from '../utils/message-utils'
import { perfLog, perfTime } from '../utils/perf-logging'
-import { filterReasoningParts } from './helpers/filter-reasoning-parts'
import { persistStreamResults } from './helpers/persist-stream-results'
import { prepareMessages } from './helpers/prepare-messages'
import { streamRelatedQuestions } from './helpers/stream-related-questions'
@@ -135,12 +135,17 @@ export async function createChatStreamResponse(
modelType
})
- // Filter out reasoning parts from messages before converting to model messages
- // OpenAI API requires reasoning messages to be followed by assistant messages
- const filteredMessages = filterReasoningParts(messagesToModel)
-
// Convert to model messages and apply context window management
- let modelMessages = convertToModelMessages(filteredMessages)
+ let modelMessages = convertToModelMessages(messagesToModel)
+
+ // Prune messages to remove unnecessary reasoning, tool calls, and empty messages
+ // This reduces token usage while keeping recent context
+ modelMessages = pruneMessages({
+ messages: modelMessages,
+ reasoning: 'before-last-message', // Keep reasoning in last message for LLM context
+ toolCalls: 'before-last-2-messages', // Keep recent tool calls (last 2 messages)
+ emptyMessages: 'remove' // Remove messages that become empty after pruning
+ })
if (shouldTruncateMessages(modelMessages, model)) {
const maxTokens = getMaxAllowedTokens(model)
diff --git a/lib/streaming/helpers/filter-reasoning-parts.ts b/lib/streaming/helpers/filter-reasoning-parts.ts
deleted file mode 100644
index a1312695..00000000
--- a/lib/streaming/helpers/filter-reasoning-parts.ts
+++ /dev/null
@@ -1,29 +0,0 @@
-import { UIMessage } from 'ai'
-
-/**
- * Filter out reasoning parts from messages before sending to the model API.
- *
- * OpenAI API requires that reasoning messages are followed by assistant messages,
- * and empty reasoning parts can cause API errors. This helper removes all reasoning
- * parts from the message history to prevent these issues.
- *
- * @param messages - Array of UI messages from the chat history
- * @returns Filtered messages without reasoning parts
- */
-export function filterReasoningParts(messages: UIMessage[]): UIMessage[] {
- return messages.map(msg => {
- if (msg.parts && Array.isArray(msg.parts)) {
- const filteredParts = msg.parts.filter(part => {
- // Remove all reasoning parts (including empty ones)
- // These are displayed in the UI but should not be sent to the API
- return part.type !== 'reasoning'
- })
-
- return {
- ...msg,
- parts: filteredParts
- }
- }
- return msg
- })
-}
diff --git a/package.json b/package.json
index 51d89d9e..45fad491 100644
--- a/package.json
+++ b/package.json
@@ -17,11 +17,11 @@
"chat": "bun run scripts/chat-cli.ts"
},
"dependencies": {
- "@ai-sdk/anthropic": "^2.0.15",
- "@ai-sdk/gateway": "^1.0.20",
- "@ai-sdk/google": "^2.0.13",
- "@ai-sdk/openai": "^2.0.28",
- "@ai-sdk/react": "^2.0.39",
+ "@ai-sdk/anthropic": "^2.0.44",
+ "@ai-sdk/gateway": "^2.0.9",
+ "@ai-sdk/google": "^2.0.31",
+ "@ai-sdk/openai": "^2.0.65",
+ "@ai-sdk/react": "^2.0.93",
"@aws-sdk/client-s3": "^3.850.0",
"@opentelemetry/api-logs": "0.57.2",
"@opentelemetry/instrumentation": "0.57.2",
@@ -52,7 +52,7 @@
"@upstash/redis": "^1.35.5",
"@vercel/analytics": "^1.5.0",
"@vercel/otel": "^1.13.0",
- "ai": "^5.0.39",
+ "ai": "^5.0.93",
"class-variance-authority": "^0.7.1",
"clsx": "^2.1.0",
"cmdk": "1.0.0",