feat: Add reasoning token support with expandable thinking blocks

- Add reasoning field to AssistantMessage type
- Create ReasoningBlock component with truncate/expand functionality
- Add reasoning-capable models (Qwen QwQ, DeepSeek R1, Qwen 3 32B)
- Implement reasoning event handling in stream processing
- Enable reasoning format for Groq and Anthropic models
- Style reasoning blocks with grey border and reduced font size
- Display reasoning above assistant messages when available

The reasoning blocks show truncated thinking tokens by default and
expand on click to show full reasoning process from models like
Qwen3-32b that support step-by-step reasoning output.
This commit is contained in:
Glen Maddern 2025-06-24 16:41:56 +10:00
commit 62000dcdac
6 changed files with 384 additions and 300 deletions

View file

@ -5,6 +5,7 @@ import rehypeHighlight from "rehype-highlight";
import { type Message } from "../types"; import { type Message } from "../types";
import ToolCallMessage from "./ToolCallMessage"; import ToolCallMessage from "./ToolCallMessage";
import ToolResultMessage from "./ToolResultMessage"; import ToolResultMessage from "./ToolResultMessage";
import ReasoningBlock from "./ReasoningBlock";
interface ChatMessageProps { interface ChatMessageProps {
message: Message; message: Message;
@ -37,6 +38,11 @@ const ChatMessage: React.FC<ChatMessageProps> = ({
: " text-zinc-900 w-full" : " text-zinc-900 w-full"
}`} }`}
> >
{/* Show reasoning block for assistant messages with reasoning */}
{message.role === "assistant" && "reasoning" in message && message.reasoning && (
<ReasoningBlock reasoning={message.reasoning} />
)}
<div className="prose prose-zinc"> <div className="prose prose-zinc">
<ReactMarkdown <ReactMarkdown
remarkPlugins={[remarkGfm]} remarkPlugins={[remarkGfm]}

View file

@ -168,7 +168,7 @@ const ConversationThread: React.FC<ConversationThreadProps> = ({
} }
}, [conversations]) }, [conversations])
console.log({ currentConversation }) // console.log({ currentConversation })
return ( return (
<div className={`flex flex-col min-h-screen w-full ${currentConversation.messages.length === 0 ? 'justify-center' : ''}`}> <div className={`flex flex-col min-h-screen w-full ${currentConversation.messages.length === 0 ? 'justify-center' : ''}`}>
@ -233,7 +233,7 @@ const ConversationThread: React.FC<ConversationThreadProps> = ({
<span className="text-lg">🧠</span> <span className="text-lg">🧠</span>
<span className="text-sm text-zinc-500"> <span className="text-sm text-zinc-500">
{hasApiKey(selectedModel.provider.id) ? ( {hasApiKey(selectedModel.provider.id) ? (
selectedModel.name.toLowerCase() selectedModel.name.toLowerCase() + ' (' + selectedModel.provider.name.toLowerCase() + ')'
) : ( ) : (
"none" "none"
)} )}

View file

@ -0,0 +1,54 @@
import React, { useState } from "react";
interface ReasoningBlockProps {
reasoning: string;
}
const ReasoningBlock: React.FC<ReasoningBlockProps> = ({ reasoning }) => {
const [isExpanded, setIsExpanded] = useState(false);
if (!reasoning || reasoning.trim().length === 0) {
return null;
}
const toggleExpanded = () => {
setIsExpanded(!isExpanded);
};
// Truncate the reasoning text for the collapsed state
const maxLength = 120;
const shouldTruncate = reasoning.length > maxLength;
const displayText = isExpanded || !shouldTruncate
? reasoning
: reasoning.substring(0, maxLength) + "...";
return (
<div className="mb-3">
<div
className={`text-sm text-zinc-600 border border-zinc-200 rounded-lg p-3 bg-zinc-50 cursor-pointer hover:bg-zinc-100 transition-colors ${shouldTruncate && !isExpanded ? "line-clamp-1" : ""
}`}
onClick={toggleExpanded}
>
<div className="flex items-start gap-2">
<span className="text-zinc-400 text-xs mt-0.5 flex-shrink-0">💭</span>
<div className={`${shouldTruncate && !isExpanded ? "truncate" : "whitespace-pre-wrap"}`}>
{displayText}
</div>
{shouldTruncate && (
<button
className="text-zinc-400 hover:text-zinc-600 text-xs ml-auto flex-shrink-0"
onClick={(e) => {
e.stopPropagation();
toggleExpanded();
}}
>
{isExpanded ? "↑" : "↓"}
</button>
)}
</div>
</div>
</div>
);
};
export default ReasoningBlock;

View file

@ -80,6 +80,19 @@ export const useStreamResponse = ({
return aiTools return aiTools
} }
// Check if a model supports reasoning
const supportsReasoning = (model: Model) => {
const reasoningModels = [
'qwen/qwen3-32b',
'qwen-qwq-32b',
'deepseek-r1-distill-llama-70b',
'claude-4-opus-20250514',
'claude-4-sonnet-20250514',
'claude-3-7-sonnet-20250219'
]
return reasoningModels.includes(model.modelId)
}
const getModelInstance = (model: Model, apiKey: string) => { const getModelInstance = (model: Model, apiKey: string) => {
switch (model.provider.id) { switch (model.provider.id) {
case 'groq': { case 'groq': {
@ -102,6 +115,7 @@ export const useStreamResponse = ({
const streamResponse = async (messages: Message[]) => { const streamResponse = async (messages: Message[]) => {
let aiResponse = '' let aiResponse = ''
let reasoning = ''
let assistantMessageIndex = -1 // Track the index of our assistant message let assistantMessageIndex = -1 // Track the index of our assistant message
let assistantMessageCreated = false // Track if we've created the assistant message yet let assistantMessageCreated = false // Track if we've created the assistant message yet
@ -138,7 +152,8 @@ export const useStreamResponse = ({
content: hasContent(m) ? m.content.substring(0, 100) + '...' : 'no content' content: hasContent(m) ? m.content.substring(0, 100) + '...' : 'no content'
}))) })))
const result = await streamText({ // Prepare stream options
const streamOptions: any = {
model: modelInstance, model: modelInstance,
messages: messages.filter(msg => messages: messages.filter(msg =>
['user', 'assistant', 'system'].includes(msg.role) && hasContent(msg) ['user', 'assistant', 'system'].includes(msg.role) && hasContent(msg)
@ -152,7 +167,30 @@ export const useStreamResponse = ({
tools: Object.keys(aiTools).length > 0 ? aiTools : undefined, tools: Object.keys(aiTools).length > 0 ? aiTools : undefined,
maxSteps: 5, // Allow up to 5 steps for tool calling maxSteps: 5, // Allow up to 5 steps for tool calling
abortSignal: controller.signal, abortSignal: controller.signal,
}) }
// Add reasoning parameters for supported models
if (supportsReasoning(selectedModel)) {
debugLog(`[useStreamResponse] Enabling reasoning for model: ${selectedModel.modelId}`)
if (selectedModel.provider.id === 'groq') {
// For Groq models, use reasoning_format parameter
streamOptions.experimental_providerOptions = {
groq: {
reasoning_format: 'parsed'
}
}
} else if (selectedModel.provider.id === 'anthropic') {
// For Anthropic models, enable thinking
streamOptions.experimental_providerOptions = {
anthropic: {
thinking: { type: 'enabled', budgetTokens: 12000 }
}
}
}
}
const result = await streamText(streamOptions)
debugLog(`[useStreamResponse] streamText result obtained, starting to process fullStream`) debugLog(`[useStreamResponse] streamText result obtained, starting to process fullStream`)
@ -163,7 +201,10 @@ export const useStreamResponse = ({
try { try {
debugLog(`[useStreamResponse] Full stream event:`, event.type, event) debugLog(`[useStreamResponse] Full stream event:`, event.type, event)
if (event.type === 'tool-call') { if (event.type === 'reasoning') {
debugLog(`[useStreamResponse] Reasoning event:`, event)
reasoning += (event as any).textDelta || ''
} else if (event.type === 'tool-call') {
const eventKey = `tool-call-${event.toolCallId}` const eventKey = `tool-call-${event.toolCallId}`
if (processedEvents.has(eventKey)) { if (processedEvents.has(eventKey)) {
debugLog(`[useStreamResponse] Skipping duplicate tool call:`, eventKey) debugLog(`[useStreamResponse] Skipping duplicate tool call:`, eventKey)
@ -196,40 +237,6 @@ export const useStreamResponse = ({
} else { } else {
console.warn('[useStreamResponse] Tool call event missing toolName:', event) console.warn('[useStreamResponse] Tool call event missing toolName:', event)
} }
} else if (event.type === 'tool-result') {
const eventKey = `tool-result-${event.toolCallId}`
if (processedEvents.has(eventKey)) {
debugLog(`[useStreamResponse] Skipping duplicate tool result:`, eventKey)
continue
}
processedEvents.add(eventKey)
debugLog(`[useStreamResponse] Tool result event:`, event)
if (event.toolName) {
setConversations((prev) => {
const updated = [...prev]
const conv = updated.find((c) => c.id === conversationId)
if (conv) {
// Check if this tool result already exists
const existingToolResult = conv.messages.find(
(msg) => msg.role === 'tool-result' && 'callId' in msg && msg.callId === event.toolCallId
)
if (!existingToolResult) {
conv.messages.push({
role: 'tool-result',
toolName: event.toolName,
toolArgs: event.args || {},
toolResult: event.result,
callId: event.toolCallId || 'unknown',
})
}
}
return updated
})
scrollToBottom()
} else {
console.warn('[useStreamResponse] Tool result event missing toolName:', event)
}
} else if (event.type === 'text-delta') { } else if (event.type === 'text-delta') {
debugLog(`[useStreamResponse] Text delta:`, event.textDelta) debugLog(`[useStreamResponse] Text delta:`, event.textDelta)
@ -240,7 +247,7 @@ export const useStreamResponse = ({
const updated = [...prev] const updated = [...prev]
const conv = updated.find((c) => c.id === conversationId) const conv = updated.find((c) => c.id === conversationId)
if (conv) { if (conv) {
conv.messages.push({ role: 'assistant', content: '' }) conv.messages.push({ role: 'assistant', content: '', reasoning: reasoning || undefined })
assistantMessageIndex = conv.messages.length - 1 assistantMessageIndex = conv.messages.length - 1
debugLog(`[useStreamResponse] Created assistant message at index ${assistantMessageIndex}`) debugLog(`[useStreamResponse] Created assistant message at index ${assistantMessageIndex}`)
} }
@ -277,6 +284,10 @@ export const useStreamResponse = ({
const assistantMessage = conv.messages[assistantMessageIndex] const assistantMessage = conv.messages[assistantMessageIndex]
if (assistantMessage && hasContent(assistantMessage) && assistantMessage.role === 'assistant') { if (assistantMessage && hasContent(assistantMessage) && assistantMessage.role === 'assistant') {
assistantMessage.content = aiResponse assistantMessage.content = aiResponse
// Update reasoning if available
if (reasoning) {
assistantMessage.reasoning = reasoning
}
debugLog(`[useStreamResponse] Updated assistant message at index ${assistantMessageIndex}`) debugLog(`[useStreamResponse] Updated assistant message at index ${assistantMessageIndex}`)
} else { } else {
debugLog(`[useStreamResponse] Could not find assistant message at index ${assistantMessageIndex}`, assistantMessage) debugLog(`[useStreamResponse] Could not find assistant message at index ${assistantMessageIndex}`, assistantMessage)

View file

@ -12,6 +12,7 @@ export interface SystemMessage extends BaseMessage {
export interface AssistantMessage extends BaseMessage { export interface AssistantMessage extends BaseMessage {
role: "assistant"; role: "assistant";
reasoning?: string;
} }
export interface ToolCallMessage { export interface ToolCallMessage {

View file

@ -42,10 +42,22 @@ export const availableModels: Model[] = [
}, },
{ {
id: 'qwen-3-32b', id: 'qwen-3-32b',
name: 'Qwen 3 32B', name: 'Qwen 3 32B (Reasoning)',
provider: providers.groq, provider: providers.groq,
modelId: 'qwen/qwen3-32b' modelId: 'qwen/qwen3-32b'
}, },
{
id: 'qwen-qwq-32b',
name: 'Qwen QwQ 32B (Reasoning)',
provider: providers.groq,
modelId: 'qwen-qwq-32b'
},
{
id: 'deepseek-r1-distill-llama-70b',
name: 'DeepSeek R1 Distil Llama 70B (Reasoning)',
provider: providers.groq,
modelId: 'deepseek-r1-distill-llama-70b'
},
{ {
id: 'claude-3-5-sonnet-20241022', id: 'claude-3-5-sonnet-20241022',
name: 'Claude 4 Sonnet', name: 'Claude 4 Sonnet',