mirror of
https://github.com/mudler/LocalAI.git
synced 2026-05-20 01:38:45 -05:00
feat(api): support 'reasoning' api field (#7959)
This PR adds support to support the 'reasoning' API field of the OpenAI spec. LocalAI now will extract automatically thinking tags in both SSE and non-SSE mode. The changes are adapted as well to the Chat UI now that will use the reasoning field to extract the thinking process and display it in the chat. This fixes https://github.com/mudler/LocalAI/issues/7944 Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
This commit is contained in:
Ettore Di Giacinto
committed by
GitHub
GitHub
parent
5ca8f0aea0
commit
c88074a19e
@@ -3,6 +3,7 @@ package openai
|
||||
import (
|
||||
"encoding/json"
|
||||
"fmt"
|
||||
"strings"
|
||||
"time"
|
||||
|
||||
"github.com/google/uuid"
|
||||
@@ -34,11 +35,54 @@ func ChatEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, evaluator
|
||||
Created: created,
|
||||
Model: req.Model, // we have to return what the user sent here, due to OpenAI spec.
|
||||
Choices: []schema.Choice{{Delta: &schema.Message{Role: "assistant"}, Index: 0, FinishReason: nil}},
|
||||
Object: "chat.completion.chunk",
|
||||
}
|
||||
responses <- initialMessage
|
||||
|
||||
// Track accumulated content for reasoning extraction
|
||||
accumulatedContent := ""
|
||||
lastEmittedReasoning := ""
|
||||
lastEmittedCleanedContent := ""
|
||||
|
||||
_, _, err := ComputeChoices(req, s, config, cl, startupOptions, loader, func(s string, c *[]schema.Choice) {}, func(s string, tokenUsage backend.TokenUsage) bool {
|
||||
accumulatedContent += s
|
||||
// Extract reasoning from accumulated content
|
||||
currentReasoning, cleanedContent := functions.ExtractReasoning(accumulatedContent)
|
||||
|
||||
// Calculate new reasoning delta (what we haven't emitted yet)
|
||||
var reasoningDelta *string
|
||||
if currentReasoning != lastEmittedReasoning {
|
||||
// Extract only the new part
|
||||
if len(currentReasoning) > len(lastEmittedReasoning) && strings.HasPrefix(currentReasoning, lastEmittedReasoning) {
|
||||
newReasoning := currentReasoning[len(lastEmittedReasoning):]
|
||||
reasoningDelta = &newReasoning
|
||||
lastEmittedReasoning = currentReasoning
|
||||
} else if currentReasoning != "" {
|
||||
// If reasoning changed in a non-append way, emit the full current reasoning
|
||||
reasoningDelta = ¤tReasoning
|
||||
lastEmittedReasoning = currentReasoning
|
||||
}
|
||||
}
|
||||
|
||||
// Calculate content delta from cleaned content
|
||||
var deltaContent string
|
||||
if len(cleanedContent) > len(lastEmittedCleanedContent) && strings.HasPrefix(cleanedContent, lastEmittedCleanedContent) {
|
||||
deltaContent = cleanedContent[len(lastEmittedCleanedContent):]
|
||||
lastEmittedCleanedContent = cleanedContent
|
||||
} else if cleanedContent != lastEmittedCleanedContent {
|
||||
// If cleaned content changed but not in a simple append, extract delta from cleaned content
|
||||
// This handles cases where thinking tags are removed mid-stream
|
||||
if lastEmittedCleanedContent == "" {
|
||||
deltaContent = cleanedContent
|
||||
lastEmittedCleanedContent = cleanedContent
|
||||
} else {
|
||||
// Content changed in non-append way, use the new cleaned content
|
||||
deltaContent = cleanedContent
|
||||
lastEmittedCleanedContent = cleanedContent
|
||||
}
|
||||
}
|
||||
// Only emit content if there's actual content (not just thinking tags)
|
||||
// If deltaContent is empty, we still emit the response but with empty content
|
||||
|
||||
usage := schema.OpenAIUsage{
|
||||
PromptTokens: tokenUsage.Prompt,
|
||||
CompletionTokens: tokenUsage.Completion,
|
||||
@@ -49,11 +93,20 @@ func ChatEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, evaluator
|
||||
usage.TimingPromptProcessing = tokenUsage.TimingPromptProcessing
|
||||
}
|
||||
|
||||
delta := &schema.Message{}
|
||||
// Only include content if there's actual content (not just thinking tags)
|
||||
if deltaContent != "" {
|
||||
delta.Content = &deltaContent
|
||||
}
|
||||
if reasoningDelta != nil && *reasoningDelta != "" {
|
||||
delta.Reasoning = reasoningDelta
|
||||
}
|
||||
|
||||
resp := schema.OpenAIResponse{
|
||||
ID: id,
|
||||
Created: created,
|
||||
Model: req.Model, // we have to return what the user sent here, due to OpenAI spec.
|
||||
Choices: []schema.Choice{{Delta: &schema.Message{Content: &s}, Index: 0, FinishReason: nil}},
|
||||
Choices: []schema.Choice{{Delta: delta, Index: 0, FinishReason: nil}},
|
||||
Object: "chat.completion.chunk",
|
||||
Usage: usage,
|
||||
}
|
||||
@@ -176,6 +229,10 @@ func ChatEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, evaluator
|
||||
if err != nil {
|
||||
return err
|
||||
}
|
||||
// Extract reasoning before processing tool calls
|
||||
reasoning, cleanedResult := functions.ExtractReasoning(result)
|
||||
result = cleanedResult
|
||||
|
||||
textContentToReturn = functions.ParseTextContent(result, config.FunctionsConfig)
|
||||
result = functions.CleanupLLMResult(result, config.FunctionsConfig)
|
||||
functionResults := functions.ParseFunctionCall(result, config.FunctionsConfig)
|
||||
@@ -208,11 +265,20 @@ func ChatEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, evaluator
|
||||
usage.TimingPromptProcessing = tokenUsage.TimingPromptProcessing
|
||||
}
|
||||
|
||||
var deltaReasoning *string
|
||||
if reasoning != "" {
|
||||
deltaReasoning = &reasoning
|
||||
}
|
||||
delta := &schema.Message{Content: &result}
|
||||
if deltaReasoning != nil {
|
||||
delta.Reasoning = deltaReasoning
|
||||
}
|
||||
|
||||
resp := schema.OpenAIResponse{
|
||||
ID: id,
|
||||
Created: created,
|
||||
Model: req.Model, // we have to return what the user sent here, due to OpenAI spec.
|
||||
Choices: []schema.Choice{{Delta: &schema.Message{Content: &result}, Index: 0, FinishReason: nil}},
|
||||
Choices: []schema.Choice{{Delta: delta, Index: 0, FinishReason: nil}},
|
||||
Object: "chat.completion.chunk",
|
||||
Usage: usage,
|
||||
}
|
||||
@@ -553,10 +619,18 @@ func ChatEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, evaluator
|
||||
default:
|
||||
|
||||
tokenCallback := func(s string, c *[]schema.Choice) {
|
||||
// Extract reasoning from the response
|
||||
reasoning, cleanedS := functions.ExtractReasoning(s)
|
||||
s = cleanedS
|
||||
|
||||
if !shouldUseFn {
|
||||
// no function is called, just reply and use stop as finish reason
|
||||
stopReason := FinishReasonStop
|
||||
*c = append(*c, schema.Choice{FinishReason: &stopReason, Index: 0, Message: &schema.Message{Role: "assistant", Content: &s}})
|
||||
message := &schema.Message{Role: "assistant", Content: &s}
|
||||
if reasoning != "" {
|
||||
message.Reasoning = &reasoning
|
||||
}
|
||||
*c = append(*c, schema.Choice{FinishReason: &stopReason, Index: 0, Message: message})
|
||||
return
|
||||
}
|
||||
|
||||
@@ -575,9 +649,13 @@ func ChatEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, evaluator
|
||||
}
|
||||
|
||||
stopReason := FinishReasonStop
|
||||
message := &schema.Message{Role: "assistant", Content: &result}
|
||||
if reasoning != "" {
|
||||
message.Reasoning = &reasoning
|
||||
}
|
||||
*c = append(*c, schema.Choice{
|
||||
FinishReason: &stopReason,
|
||||
Message: &schema.Message{Role: "assistant", Content: &result}})
|
||||
Message: message})
|
||||
default:
|
||||
toolCallsReason := FinishReasonToolCalls
|
||||
toolChoice := schema.Choice{
|
||||
@@ -586,6 +664,9 @@ func ChatEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, evaluator
|
||||
Role: "assistant",
|
||||
},
|
||||
}
|
||||
if reasoning != "" {
|
||||
toolChoice.Message.Reasoning = &reasoning
|
||||
}
|
||||
|
||||
for _, ss := range results {
|
||||
name, args := ss.Name, ss.Arguments
|
||||
@@ -606,16 +687,20 @@ func ChatEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, evaluator
|
||||
} else {
|
||||
// otherwise we return more choices directly (deprecated)
|
||||
functionCallReason := FinishReasonFunctionCall
|
||||
message := &schema.Message{
|
||||
Role: "assistant",
|
||||
Content: &textContentToReturn,
|
||||
FunctionCall: map[string]interface{}{
|
||||
"name": name,
|
||||
"arguments": args,
|
||||
},
|
||||
}
|
||||
if reasoning != "" {
|
||||
message.Reasoning = &reasoning
|
||||
}
|
||||
*c = append(*c, schema.Choice{
|
||||
FinishReason: &functionCallReason,
|
||||
Message: &schema.Message{
|
||||
Role: "assistant",
|
||||
Content: &textContentToReturn,
|
||||
FunctionCall: map[string]interface{}{
|
||||
"name": name,
|
||||
"arguments": args,
|
||||
},
|
||||
},
|
||||
Message: message,
|
||||
})
|
||||
}
|
||||
}
|
||||
|
||||
+203
-41
@@ -1368,6 +1368,7 @@ async function promptGPT(systemPrompt, input) {
|
||||
let lastAssistantMessageIndex = -1;
|
||||
let lastThinkingMessageIndex = -1;
|
||||
let lastThinkingScrollTime = 0;
|
||||
let hasReasoningFromAPI = false; // Track if we're receiving reasoning from API (skip tag-based detection)
|
||||
const THINKING_SCROLL_THROTTLE = 200; // Throttle scrolling to every 200ms
|
||||
|
||||
try {
|
||||
@@ -1401,19 +1402,24 @@ async function promptGPT(systemPrompt, input) {
|
||||
// Handle different event types
|
||||
switch (eventData.type) {
|
||||
case "reasoning":
|
||||
hasReasoningFromAPI = true; // Mark that we're receiving reasoning from API
|
||||
if (eventData.content) {
|
||||
// Insert reasoning before assistant message if it exists
|
||||
const currentChat = chatStore.getChat(chatId);
|
||||
if (!currentChat) break; // Chat was deleted
|
||||
const isMCPMode = currentChat.mcpMode || false;
|
||||
const shouldExpand = !isMCPMode; // Expanded in non-MCP mode, collapsed in MCP mode
|
||||
// Insert thinking before assistant message if it exists (always use "thinking" role)
|
||||
if (lastAssistantMessageIndex >= 0 && targetHistory[lastAssistantMessageIndex]?.role === "assistant") {
|
||||
targetHistory.splice(lastAssistantMessageIndex, 0, {
|
||||
role: "reasoning",
|
||||
role: "thinking",
|
||||
content: eventData.content,
|
||||
html: DOMPurify.sanitize(marked.parse(eventData.content)),
|
||||
image: [],
|
||||
audio: [],
|
||||
expanded: false // Reasoning is always collapsed
|
||||
expanded: shouldExpand
|
||||
});
|
||||
lastAssistantMessageIndex++; // Adjust index since we inserted
|
||||
// Scroll smoothly after adding reasoning
|
||||
// Scroll smoothly after adding thinking
|
||||
setTimeout(() => {
|
||||
const chatContainer = document.getElementById('chat');
|
||||
if (chatContainer) {
|
||||
@@ -1425,7 +1431,7 @@ async function promptGPT(systemPrompt, input) {
|
||||
}, 100);
|
||||
} else {
|
||||
// No assistant message yet, just add normally
|
||||
chatStore.add("reasoning", eventData.content, null, null, chatId);
|
||||
chatStore.add("thinking", eventData.content, null, null, chatId);
|
||||
}
|
||||
}
|
||||
break;
|
||||
@@ -1491,14 +1497,17 @@ async function promptGPT(systemPrompt, input) {
|
||||
// Only update display if this is the active chat (interval will handle it)
|
||||
// Don't call updateTokensPerSecond here to avoid unnecessary updates
|
||||
|
||||
// Check for thinking tags in the chunk (incremental detection)
|
||||
if (contentChunk.includes("<thinking>") || contentChunk.includes("<think>")) {
|
||||
isThinking = true;
|
||||
thinkingContent = "";
|
||||
lastThinkingMessageIndex = -1;
|
||||
}
|
||||
|
||||
if (contentChunk.includes("</thinking>") || contentChunk.includes("</think>")) {
|
||||
// Only check for thinking tags if we're NOT receiving reasoning from API
|
||||
// This prevents duplicate thinking/reasoning messages
|
||||
if (!hasReasoningFromAPI) {
|
||||
// Check for thinking tags in the chunk (incremental detection)
|
||||
if (contentChunk.includes("<thinking>") || contentChunk.includes("<think>")) {
|
||||
isThinking = true;
|
||||
thinkingContent = "";
|
||||
lastThinkingMessageIndex = -1;
|
||||
}
|
||||
|
||||
if (contentChunk.includes("</thinking>") || contentChunk.includes("</think>")) {
|
||||
isThinking = false;
|
||||
// When closing tag is detected, process the accumulated thinking content
|
||||
if (thinkingContent.trim()) {
|
||||
@@ -1552,10 +1561,11 @@ async function promptGPT(systemPrompt, input) {
|
||||
}
|
||||
thinkingContent = "";
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// Handle content based on thinking state
|
||||
if (isThinking) {
|
||||
// Handle content based on thinking state (only if not receiving reasoning from API)
|
||||
if (!hasReasoningFromAPI && isThinking) {
|
||||
thinkingContent += contentChunk;
|
||||
const currentChat = chatStore.getChat(chatId);
|
||||
if (!currentChat) break; // Chat was deleted
|
||||
@@ -1637,7 +1647,10 @@ async function promptGPT(systemPrompt, input) {
|
||||
|
||||
// Process any thinking tags that might be in the accumulated content
|
||||
// This handles cases where tags are split across chunks
|
||||
const { regularContent: processedRegular, thinkingContent: processedThinking } = processThinkingTags(regularContent);
|
||||
// Only process if we're NOT receiving reasoning from API (to avoid duplicates)
|
||||
const { regularContent: processedRegular, thinkingContent: processedThinking } = hasReasoningFromAPI
|
||||
? { regularContent: regularContent, thinkingContent: "" }
|
||||
: processThinkingTags(regularContent);
|
||||
|
||||
// Update or create assistant message with processed regular content
|
||||
const currentChat = chatStore.getChat(chatId);
|
||||
@@ -1645,10 +1658,10 @@ async function promptGPT(systemPrompt, input) {
|
||||
const request = activeRequests.get(chatId);
|
||||
const requestModel = request?.model || null;
|
||||
if (lastAssistantMessageIndex === -1) {
|
||||
if (processedRegular && processedRegular.trim()) {
|
||||
chatStore.add("assistant", processedRegular, null, null, chatId, requestModel);
|
||||
lastAssistantMessageIndex = targetHistory.length - 1;
|
||||
}
|
||||
// Create assistant message if we have any content (even if empty string after processing)
|
||||
// This ensures the message is created and can be updated with more content later
|
||||
chatStore.add("assistant", processedRegular || "", null, null, chatId, requestModel);
|
||||
lastAssistantMessageIndex = targetHistory.length - 1;
|
||||
} else {
|
||||
const lastMessage = targetHistory[lastAssistantMessageIndex];
|
||||
if (lastMessage && lastMessage.role === "assistant") {
|
||||
@@ -1686,7 +1699,10 @@ async function promptGPT(systemPrompt, input) {
|
||||
if (assistantContentBuffer.length > 0) {
|
||||
const regularContent = assistantContentBuffer.join("");
|
||||
// Process any remaining thinking tags that might be in the buffer
|
||||
const { regularContent: processedRegular, thinkingContent: processedThinking } = processThinkingTags(regularContent);
|
||||
// Only process if we're NOT receiving reasoning from API (to avoid duplicates)
|
||||
const { regularContent: processedRegular, thinkingContent: processedThinking } = hasReasoningFromAPI
|
||||
? { regularContent: regularContent, thinkingContent: "" }
|
||||
: processThinkingTags(regularContent);
|
||||
|
||||
const currentChat = chatStore.getChat(chatId);
|
||||
if (!currentChat) {
|
||||
@@ -1719,23 +1735,26 @@ async function promptGPT(systemPrompt, input) {
|
||||
}
|
||||
|
||||
// Then update or create assistant message
|
||||
// Always create/update assistant message if we have any content
|
||||
if (lastAssistantMessageIndex !== -1) {
|
||||
const lastMessage = targetHistory[lastAssistantMessageIndex];
|
||||
if (lastMessage && lastMessage.role === "assistant") {
|
||||
lastMessage.content = (lastMessage.content || "") + (processedRegular || "");
|
||||
lastMessage.html = DOMPurify.sanitize(marked.parse(lastMessage.content));
|
||||
}
|
||||
} else if (processedRegular && processedRegular.trim()) {
|
||||
} else {
|
||||
// Create assistant message (even if empty, so it can be updated with more content)
|
||||
const request = activeRequests.get(chatId);
|
||||
const requestModel = request?.model || null;
|
||||
chatStore.add("assistant", processedRegular, null, null, chatId, requestModel);
|
||||
chatStore.add("assistant", processedRegular || "", null, null, chatId, requestModel);
|
||||
lastAssistantMessageIndex = targetHistory.length - 1;
|
||||
}
|
||||
}
|
||||
|
||||
// Final thinking content flush if any data remains (from incremental detection)
|
||||
// Only process if we're NOT receiving reasoning from API (to avoid duplicates)
|
||||
const finalChat = chatStore.getChat(chatId);
|
||||
if (finalChat && thinkingContent.trim() && lastThinkingMessageIndex === -1) {
|
||||
if (finalChat && !hasReasoningFromAPI && thinkingContent.trim() && lastThinkingMessageIndex === -1) {
|
||||
const finalHistory = finalChat.history;
|
||||
// Extract thinking content if tags are present
|
||||
const thinkingMatch = thinkingContent.match(/<(?:thinking|redacted_reasoning)>(.*?)<\/(?:thinking|redacted_reasoning)>/s);
|
||||
@@ -1891,9 +1910,13 @@ async function promptGPT(systemPrompt, input) {
|
||||
let buffer = "";
|
||||
let contentBuffer = [];
|
||||
let thinkingContent = "";
|
||||
let reasoningContent = ""; // Track reasoning from API reasoning field
|
||||
let isThinking = false;
|
||||
let lastThinkingMessageIndex = -1;
|
||||
let lastReasoningMessageIndex = -1; // Track reasoning message separately
|
||||
let lastAssistantMessageIndex = -1; // Track assistant message for reasoning placement
|
||||
let lastThinkingScrollTime = 0;
|
||||
let hasReasoningFromAPI = false; // Track if we're receiving reasoning from API (skip tag-based detection)
|
||||
const THINKING_SCROLL_THROTTLE = 200; // Throttle scrolling to every 200ms
|
||||
|
||||
try {
|
||||
@@ -1929,30 +1952,100 @@ async function promptGPT(systemPrompt, input) {
|
||||
chatStore.updateTokenUsage(jsonData.usage, chatId);
|
||||
}
|
||||
|
||||
const token = jsonData.choices[0].delta.content;
|
||||
const token = jsonData.choices?.[0]?.delta?.content;
|
||||
const reasoningDelta = jsonData.choices?.[0]?.delta?.reasoning;
|
||||
|
||||
if (token) {
|
||||
// Check for thinking tags
|
||||
if (token.includes("<thinking>") || token.includes("<think>")) {
|
||||
isThinking = true;
|
||||
thinkingContent = "";
|
||||
lastThinkingMessageIndex = -1;
|
||||
// Handle reasoning from API reasoning field - always use "thinking" role
|
||||
if (reasoningDelta && reasoningDelta.trim() !== "") {
|
||||
hasReasoningFromAPI = true; // Mark that we're receiving reasoning from API
|
||||
reasoningContent += reasoningDelta;
|
||||
const currentChat = chatStore.getChat(chatId);
|
||||
if (!currentChat) {
|
||||
// Chat was deleted, skip this line
|
||||
return;
|
||||
}
|
||||
if (token.includes("</thinking>") || token.includes("</think>")) {
|
||||
isThinking = false;
|
||||
if (thinkingContent.trim()) {
|
||||
// Only add the final thinking message if we don't already have one
|
||||
if (lastThinkingMessageIndex === -1) {
|
||||
chatStore.add("thinking", thinkingContent, null, null, chatId);
|
||||
const isMCPMode = currentChat.mcpMode || false;
|
||||
const shouldExpand = !isMCPMode; // Expanded in non-MCP mode, collapsed in MCP mode
|
||||
|
||||
// Only create/update thinking message if we have actual content
|
||||
if (reasoningContent.trim() !== "") {
|
||||
// Update or create thinking message (always use "thinking" role, not "reasoning")
|
||||
if (lastReasoningMessageIndex === -1) {
|
||||
// Find the last assistant message index to insert thinking before it
|
||||
const targetHistory = currentChat.history;
|
||||
const assistantIndex = targetHistory.length - 1;
|
||||
if (assistantIndex >= 0 && targetHistory[assistantIndex]?.role === "assistant") {
|
||||
// Insert thinking before assistant message
|
||||
targetHistory.splice(assistantIndex, 0, {
|
||||
role: "thinking",
|
||||
content: reasoningContent,
|
||||
html: DOMPurify.sanitize(marked.parse(reasoningContent)),
|
||||
image: [],
|
||||
audio: [],
|
||||
expanded: shouldExpand
|
||||
});
|
||||
lastReasoningMessageIndex = assistantIndex;
|
||||
lastAssistantMessageIndex = assistantIndex + 1; // Adjust for inserted thinking
|
||||
} else {
|
||||
// No assistant message yet, just add normally
|
||||
chatStore.add("thinking", reasoningContent, null, null, chatId);
|
||||
lastReasoningMessageIndex = currentChat.history.length - 1;
|
||||
}
|
||||
} else {
|
||||
// Update existing thinking message
|
||||
const targetHistory = currentChat.history;
|
||||
if (lastReasoningMessageIndex >= 0 && lastReasoningMessageIndex < targetHistory.length) {
|
||||
const thinkingMessage = targetHistory[lastReasoningMessageIndex];
|
||||
if (thinkingMessage && thinkingMessage.role === "thinking") {
|
||||
thinkingMessage.content = reasoningContent;
|
||||
thinkingMessage.html = DOMPurify.sanitize(marked.parse(reasoningContent));
|
||||
}
|
||||
}
|
||||
}
|
||||
return;
|
||||
}
|
||||
|
||||
// Scroll when reasoning is updated (throttled)
|
||||
const now = Date.now();
|
||||
if (now - lastThinkingScrollTime > THINKING_SCROLL_THROTTLE) {
|
||||
lastThinkingScrollTime = now;
|
||||
setTimeout(() => {
|
||||
const chatContainer = document.getElementById('chat');
|
||||
if (chatContainer) {
|
||||
chatContainer.scrollTo({
|
||||
top: chatContainer.scrollHeight,
|
||||
behavior: 'smooth'
|
||||
});
|
||||
}
|
||||
scrollThinkingBoxToBottom();
|
||||
}, 100);
|
||||
}
|
||||
}
|
||||
|
||||
// Handle content based on thinking state
|
||||
if (isThinking) {
|
||||
thinkingContent += token;
|
||||
if (token && token.trim() !== "") {
|
||||
// Only check for thinking tags if we're NOT receiving reasoning from API
|
||||
// This prevents duplicate thinking/reasoning messages
|
||||
if (!hasReasoningFromAPI) {
|
||||
// Check for thinking tags (legacy support - models that output tags directly)
|
||||
if (token.includes("<thinking>") || token.includes("<think>")) {
|
||||
isThinking = true;
|
||||
thinkingContent = "";
|
||||
lastThinkingMessageIndex = -1;
|
||||
return;
|
||||
}
|
||||
if (token.includes("</thinking>") || token.includes("</think>")) {
|
||||
isThinking = false;
|
||||
if (thinkingContent.trim()) {
|
||||
// Only add the final thinking message if we don't already have one
|
||||
if (lastThinkingMessageIndex === -1) {
|
||||
chatStore.add("thinking", thinkingContent, null, null, chatId);
|
||||
}
|
||||
}
|
||||
return;
|
||||
}
|
||||
|
||||
// Handle content based on thinking state
|
||||
if (isThinking) {
|
||||
thinkingContent += token;
|
||||
// Count tokens for rate calculation (per chat)
|
||||
const request = activeRequests.get(chatId);
|
||||
if (request) {
|
||||
@@ -1995,7 +2088,42 @@ async function promptGPT(systemPrompt, input) {
|
||||
}, 100);
|
||||
}
|
||||
} else {
|
||||
// Not in thinking state, add to content buffer
|
||||
contentBuffer.push(token);
|
||||
// Track assistant message index for reasoning placement
|
||||
if (lastAssistantMessageIndex === -1) {
|
||||
const currentChat = chatStore.getChat(chatId);
|
||||
if (currentChat) {
|
||||
const targetHistory = currentChat.history;
|
||||
// Find or create assistant message index
|
||||
for (let i = targetHistory.length - 1; i >= 0; i--) {
|
||||
if (targetHistory[i].role === "assistant") {
|
||||
lastAssistantMessageIndex = i;
|
||||
break;
|
||||
}
|
||||
}
|
||||
// If no assistant message yet, it will be created when we flush contentBuffer
|
||||
}
|
||||
}
|
||||
}
|
||||
} else {
|
||||
// Receiving reasoning from API, just add token to content buffer
|
||||
contentBuffer.push(token);
|
||||
// Track assistant message index for reasoning placement
|
||||
if (lastAssistantMessageIndex === -1) {
|
||||
const currentChat = chatStore.getChat(chatId);
|
||||
if (currentChat) {
|
||||
const targetHistory = currentChat.history;
|
||||
// Find or create assistant message index
|
||||
for (let i = targetHistory.length - 1; i >= 0; i--) {
|
||||
if (targetHistory[i].role === "assistant") {
|
||||
lastAssistantMessageIndex = i;
|
||||
break;
|
||||
}
|
||||
}
|
||||
// If no assistant message yet, it will be created when we flush contentBuffer
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
} catch (error) {
|
||||
@@ -2007,6 +2135,17 @@ async function promptGPT(systemPrompt, input) {
|
||||
// Efficiently update the chat in batch
|
||||
if (contentBuffer.length > 0) {
|
||||
addToChat(contentBuffer.join(""));
|
||||
// Update assistant message index after adding content
|
||||
const currentChat = chatStore.getChat(chatId);
|
||||
if (currentChat) {
|
||||
const targetHistory = currentChat.history;
|
||||
for (let i = targetHistory.length - 1; i >= 0; i--) {
|
||||
if (targetHistory[i].role === "assistant") {
|
||||
lastAssistantMessageIndex = i;
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
contentBuffer = [];
|
||||
// Scroll when assistant content is updated (this will also show thinking messages above)
|
||||
setTimeout(() => {
|
||||
@@ -2025,7 +2164,30 @@ async function promptGPT(systemPrompt, input) {
|
||||
if (contentBuffer.length > 0) {
|
||||
addToChat(contentBuffer.join(""));
|
||||
}
|
||||
|
||||
// Final reasoning flush if any data remains - always use "thinking" role
|
||||
const finalChat = chatStore.getChat(chatId);
|
||||
if (finalChat && reasoningContent.trim() && lastReasoningMessageIndex === -1) {
|
||||
const isMCPMode = finalChat.mcpMode || false;
|
||||
const shouldExpand = !isMCPMode;
|
||||
const targetHistory = finalChat.history;
|
||||
// Find assistant message to insert before
|
||||
const assistantIndex = targetHistory.length - 1;
|
||||
if (assistantIndex >= 0 && targetHistory[assistantIndex]?.role === "assistant") {
|
||||
targetHistory.splice(assistantIndex, 0, {
|
||||
role: "thinking",
|
||||
content: reasoningContent,
|
||||
html: DOMPurify.sanitize(marked.parse(reasoningContent)),
|
||||
image: [],
|
||||
audio: [],
|
||||
expanded: shouldExpand
|
||||
});
|
||||
} else {
|
||||
chatStore.add("thinking", reasoningContent, null, null, chatId);
|
||||
}
|
||||
}
|
||||
|
||||
// Final thinking content flush (legacy tag-based thinking)
|
||||
if (finalChat && thinkingContent.trim() && lastThinkingMessageIndex === -1) {
|
||||
chatStore.add("thinking", thinkingContent, null, null, chatId);
|
||||
}
|
||||
|
||||
+260
-49
@@ -41,7 +41,7 @@ SOFTWARE.
|
||||
__chatContextSize = {{ .ContextSize }};
|
||||
{{ end }}
|
||||
|
||||
// Store gallery configs for header icon display
|
||||
// Store gallery configs for header icon display and model info modal
|
||||
window.__galleryConfigs = {};
|
||||
{{ $allGalleryConfigs:=.GalleryConfig }}
|
||||
{{ range $modelName, $galleryConfig := $allGalleryConfigs }}
|
||||
@@ -49,6 +49,16 @@ SOFTWARE.
|
||||
{{ if $galleryConfig.Icon }}
|
||||
window.__galleryConfigs["{{$modelName}}"].Icon = "{{$galleryConfig.Icon}}";
|
||||
{{ end }}
|
||||
{{ if $galleryConfig.Description }}
|
||||
window.__galleryConfigs["{{$modelName}}"].Description = {{ printf "%q" $galleryConfig.Description }};
|
||||
{{ end }}
|
||||
{{ if $galleryConfig.URLs }}
|
||||
window.__galleryConfigs["{{$modelName}}"].URLs = [
|
||||
{{ range $idx, $url := $galleryConfig.URLs }}
|
||||
{{ if $idx }},{{ end }}{{ printf "%q" $url }}
|
||||
{{ end }}
|
||||
];
|
||||
{{ end }}
|
||||
{{ end }}
|
||||
|
||||
// Function to initialize store
|
||||
@@ -326,10 +336,10 @@ SOFTWARE.
|
||||
c += DOMPurify.sanitize(marked.parse(line));
|
||||
});
|
||||
}
|
||||
// Set expanded state: thinking is expanded by default in non-MCP mode, collapsed in MCP mode
|
||||
// Reasoning, tool_call, and tool_result are always collapsed by default
|
||||
// Set expanded state: thinking and reasoning are expanded by default in non-MCP mode, collapsed in MCP mode
|
||||
// tool_call and tool_result are always collapsed by default
|
||||
const isMCPMode = chat.mcpMode || false;
|
||||
const shouldExpand = (role === "thinking" && !isMCPMode) || false;
|
||||
const shouldExpand = ((role === "thinking" || role === "reasoning") && !isMCPMode) || false;
|
||||
chat.history.push({ role, content, html: c, image, audio, expanded: shouldExpand, model: messageModel });
|
||||
|
||||
// Auto-name chat from first user message
|
||||
@@ -497,6 +507,11 @@ SOFTWARE.
|
||||
activeChat.model = modelName;
|
||||
activeChat.updatedAt = Date.now();
|
||||
|
||||
// Update model info modal with new model
|
||||
if (window.updateModelInfoModal) {
|
||||
window.updateModelInfoModal(modelName);
|
||||
}
|
||||
|
||||
// Get context size from data attribute
|
||||
let contextSize = null;
|
||||
if (selectedOption.dataset.contextSize) {
|
||||
@@ -536,18 +551,23 @@ SOFTWARE.
|
||||
}
|
||||
|
||||
// Update model selector to reflect the change (ensure it stays in sync)
|
||||
// Note: We don't dispatch a change event here to avoid infinite loop
|
||||
// The selector is already updated via user interaction or programmatic change
|
||||
const modelSelector = document.getElementById('modelSelector');
|
||||
if (modelSelector) {
|
||||
// Find and select the option matching the model
|
||||
const optionValue = 'chat/' + modelName;
|
||||
for (let i = 0; i < modelSelector.options.length; i++) {
|
||||
if (modelSelector.options[i].value === optionValue) {
|
||||
modelSelector.selectedIndex = i;
|
||||
// Only update if it's different to avoid unnecessary updates
|
||||
if (modelSelector.selectedIndex !== i) {
|
||||
modelSelector.selectedIndex = i;
|
||||
}
|
||||
break;
|
||||
}
|
||||
}
|
||||
// Trigger Alpine reactivity by dispatching change event
|
||||
modelSelector.dispatchEvent(new Event('change', { bubbles: true }));
|
||||
// Don't dispatch change event here - it would cause infinite recursion
|
||||
// The selector is already in sync with the model
|
||||
}
|
||||
|
||||
// Trigger MCP availability check in Alpine component
|
||||
@@ -603,27 +623,52 @@ SOFTWARE.
|
||||
<div class="flex items-center justify-between gap-2">
|
||||
<label class="text-xs font-medium text-[var(--color-text-secondary)] uppercase tracking-wide flex-shrink-0">Model</label>
|
||||
<div class="flex items-center gap-1 flex-shrink-0">
|
||||
{{ if $model }}
|
||||
{{ $galleryConfig:= index $allGalleryConfigs $model}}
|
||||
{{ if $galleryConfig }}
|
||||
<button
|
||||
data-twe-ripple-init
|
||||
data-twe-ripple-color="light"
|
||||
class="text-[var(--color-text-secondary)] hover:text-[var(--color-primary)] transition-colors text-xs p-1 rounded hover:bg-[var(--color-bg-primary)]"
|
||||
data-modal-target="model-info-modal"
|
||||
data-modal-toggle="model-info-modal"
|
||||
title="Model Information">
|
||||
<i class="fas fa-info-circle"></i>
|
||||
</button>
|
||||
{{ end }}
|
||||
{{ end }}
|
||||
{{ if $model }}
|
||||
<a href="/models/edit/{{$model}}"
|
||||
class="text-[var(--color-text-secondary)] hover:text-[var(--color-warning)] transition-colors text-xs p-1 rounded hover:bg-[var(--color-bg-primary)]"
|
||||
title="Edit Model Configuration">
|
||||
<i class="fas fa-edit"></i>
|
||||
</a>
|
||||
{{ end }}
|
||||
<!-- Info button - reactive to active chat model -->
|
||||
<template x-if="$store.chat.activeChat() && $store.chat.activeChat().model && window.__galleryConfigs && window.__galleryConfigs[$store.chat.activeChat().model]">
|
||||
<button
|
||||
data-twe-ripple-init
|
||||
data-twe-ripple-color="light"
|
||||
class="text-[var(--color-text-secondary)] hover:text-[var(--color-primary)] transition-colors text-xs p-1 rounded hover:bg-[var(--color-bg-primary)]"
|
||||
data-modal-target="model-info-modal"
|
||||
data-modal-toggle="model-info-modal"
|
||||
:data-model-name="$store.chat.activeChat().model"
|
||||
@click="if (window.updateModelInfoModal) { window.updateModelInfoModal($store.chat.activeChat().model, true); }"
|
||||
title="Model Information">
|
||||
<i class="fas fa-info-circle"></i>
|
||||
</button>
|
||||
</template>
|
||||
<!-- Fallback info button for initial model from server -->
|
||||
<template x-if="(!$store.chat.activeChat() || !$store.chat.activeChat().model) && window.__galleryConfigs && window.__galleryConfigs['{{$model}}']">
|
||||
<button
|
||||
data-twe-ripple-init
|
||||
data-twe-ripple-color="light"
|
||||
class="text-[var(--color-text-secondary)] hover:text-[var(--color-primary)] transition-colors text-xs p-1 rounded hover:bg-[var(--color-bg-primary)]"
|
||||
data-modal-target="model-info-modal"
|
||||
data-modal-toggle="model-info-modal"
|
||||
data-model-name="{{$model}}"
|
||||
@click="if (window.updateModelInfoModal) { window.updateModelInfoModal('{{$model}}', true); }"
|
||||
title="Model Information">
|
||||
<i class="fas fa-info-circle"></i>
|
||||
</button>
|
||||
</template>
|
||||
<!-- Edit button - reactive to active chat model -->
|
||||
<template x-if="$store.chat.activeChat() && $store.chat.activeChat().model">
|
||||
<a :href="'/models/edit/' + $store.chat.activeChat().model"
|
||||
class="text-[var(--color-text-secondary)] hover:text-[var(--color-warning)] transition-colors text-xs p-1 rounded hover:bg-[var(--color-bg-primary)]"
|
||||
title="Edit Model Configuration">
|
||||
<i class="fas fa-edit"></i>
|
||||
</a>
|
||||
</template>
|
||||
<!-- Fallback edit button for initial model from server -->
|
||||
<template x-if="!$store.chat.activeChat() || !$store.chat.activeChat().model">
|
||||
{{ if $model }}
|
||||
<a href="/models/edit/{{$model}}"
|
||||
class="text-[var(--color-text-secondary)] hover:text-[var(--color-warning)] transition-colors text-xs p-1 rounded hover:bg-[var(--color-bg-primary)]"
|
||||
title="Edit Model Configuration">
|
||||
<i class="fas fa-edit"></i>
|
||||
</a>
|
||||
{{ end }}
|
||||
</template>
|
||||
</div>
|
||||
</div>
|
||||
<select
|
||||
@@ -1488,17 +1533,14 @@ SOFTWARE.
|
||||
</div>
|
||||
</div>
|
||||
|
||||
<!-- Modal moved outside of sidebar to appear in center of page -->
|
||||
{{ if $model }}
|
||||
{{ $galleryConfig:= index $allGalleryConfigs $model}}
|
||||
{{ if $galleryConfig }}
|
||||
<div id="model-info-modal" tabindex="-1" aria-hidden="true" class="hidden overflow-y-auto overflow-x-hidden fixed top-0 right-0 left-0 z-50 flex justify-center items-center w-full md:inset-0 h-[calc(100%-1rem)] max-h-full">
|
||||
<!-- Modal moved outside of sidebar to appear in center of page - Always available, content updated dynamically -->
|
||||
<div id="model-info-modal" tabindex="-1" aria-hidden="true" class="hidden overflow-y-auto overflow-x-hidden fixed top-0 right-0 left-0 z-50 flex justify-center items-center w-full h-full md:inset-0 max-h-full" style="padding: 1rem;">
|
||||
<div class="relative p-4 w-full max-w-2xl max-h-full">
|
||||
<div class="relative p-4 w-full max-w-2xl max-h-full bg-white rounded-lg shadow dark:bg-gray-700">
|
||||
<!-- Header -->
|
||||
<div class="flex items-center justify-between p-4 md:p-5 border-b rounded-t dark:border-gray-600">
|
||||
<h3 class="text-xl font-semibold text-gray-900 dark:text-white">{{ $model }}</h3>
|
||||
<button class="text-gray-400 bg-transparent hover:bg-gray-200 hover:text-gray-900 rounded-lg text-sm w-8 h-8 ms-auto inline-flex justify-center items-center dark:hover:bg-gray-600 dark:hover:text-white" data-modal-hide="model-info-modal">
|
||||
<h3 id="model-info-modal-title" class="text-xl font-semibold text-gray-900 dark:text-white">{{ if $model }}{{ $model }}{{ end }}</h3>
|
||||
<button class="text-gray-400 bg-transparent hover:bg-gray-200 hover:text-gray-900 rounded-lg text-sm w-8 h-8 ms-auto inline-flex justify-center items-center dark:hover:bg-gray-600 dark:hover:text-white" data-modal-hide="model-info-modal" @click="if (window.closeModelInfoModal) { window.closeModelInfoModal(); }">
|
||||
<svg class="w-3 h-3" aria-hidden="true" xmlns="http://www.w3.org/2000/svg" fill="none" viewBox="0 0 14 14">
|
||||
<path stroke="currentColor" stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="m1 1 6 6m0 0 6 6M7 7l6-6M7 7l-6 6"/>
|
||||
</svg>
|
||||
@@ -1509,29 +1551,24 @@ SOFTWARE.
|
||||
<!-- Body -->
|
||||
<div class="p-4 md:p-5 space-y-4">
|
||||
<div class="flex justify-center items-center">
|
||||
{{ if $galleryConfig.Icon }}<img class="lazy rounded-t-lg max-h-48 max-w-96 object-cover mt-3 entered loaded" src="{{$galleryConfig.Icon}}" loading="lazy"/>{{end}}
|
||||
<img id="model-info-modal-icon" class="lazy rounded-t-lg max-h-48 max-w-96 object-cover mt-3 entered loaded" style="display: none;" loading="lazy"/>
|
||||
</div>
|
||||
<div id="model-info-description" class="text-base leading-relaxed text-gray-500 dark:text-gray-400 break-words max-w-full">{{ $galleryConfig.Description }}</div>
|
||||
<div id="model-info-description" class="text-base leading-relaxed text-gray-500 dark:text-gray-400 break-words max-w-full"></div>
|
||||
<hr>
|
||||
<p class="text-sm font-semibold text-gray-900 dark:text-white">Links</p>
|
||||
<ul>
|
||||
{{range $galleryConfig.URLs}}
|
||||
<li><a href="{{ . }}" target="_blank">{{ . }}</a></li>
|
||||
{{end}}
|
||||
<ul id="model-info-links">
|
||||
</ul>
|
||||
</div>
|
||||
|
||||
<!-- Footer -->
|
||||
<div class="flex items-center p-4 md:p-5 border-t border-gray-200 rounded-b dark:border-gray-600">
|
||||
<button data-modal-hide="model-info-modal" class="py-2.5 px-5 ms-3 text-sm font-medium text-gray-900 focus:outline-none bg-white rounded-lg border border-gray-200 hover:bg-gray-100 hover:text-blue-700 focus:z-10 focus:ring-4 focus:ring-gray-100 dark:focus:ring-gray-700 dark:bg-gray-800 dark:text-gray-400 dark:border-gray-600 dark:hover:text-white dark:hover:bg-gray-700">
|
||||
<button data-modal-hide="model-info-modal" class="py-2.5 px-5 ms-3 text-sm font-medium text-gray-900 focus:outline-none bg-white rounded-lg border border-gray-200 hover:bg-gray-100 hover:text-blue-700 focus:z-10 focus:ring-4 focus:ring-gray-100 dark:focus:ring-gray-700 dark:bg-gray-800 dark:text-gray-400 dark:border-gray-600 dark:hover:text-white dark:hover:bg-gray-700" @click="if (window.closeModelInfoModal) { window.closeModelInfoModal(); }">
|
||||
Close
|
||||
</button>
|
||||
</div>
|
||||
</div>
|
||||
</div>
|
||||
</div>
|
||||
{{ end }}
|
||||
{{ end }}
|
||||
|
||||
<!-- Alpine store initialization and utilities -->
|
||||
<script>
|
||||
@@ -1742,10 +1779,20 @@ SOFTWARE.
|
||||
});
|
||||
|
||||
// Also listen for click events on modal toggle buttons
|
||||
document.querySelectorAll('[data-modal-toggle="model-info-modal"]').forEach(button => {
|
||||
button.addEventListener('click', () => {
|
||||
// Use event delegation to handle dynamically created buttons
|
||||
document.addEventListener('click', (e) => {
|
||||
const button = e.target.closest('[data-modal-toggle="model-info-modal"]');
|
||||
if (button) {
|
||||
// Update modal with current model before showing
|
||||
if (window.Alpine && window.Alpine.store("chat")) {
|
||||
const activeChat = window.Alpine.store("chat").activeChat();
|
||||
const modelName = activeChat ? activeChat.model : (button.dataset.modelName || (document.getElementById("chat-model") ? document.getElementById("chat-model").value : null));
|
||||
if (modelName && window.updateModelInfoModal) {
|
||||
window.updateModelInfoModal(modelName, true);
|
||||
}
|
||||
}
|
||||
setTimeout(processMarkdown, 300);
|
||||
});
|
||||
}
|
||||
});
|
||||
|
||||
// Process on initial load if libraries are ready
|
||||
@@ -1786,12 +1833,176 @@ SOFTWARE.
|
||||
syncModelSelectorOnLoad();
|
||||
}
|
||||
|
||||
// Function to update model info modal with current model
|
||||
// Set openModal to true to actually open the modal, false to just update content
|
||||
window.updateModelInfoModal = function(modelName, openModal = false) {
|
||||
if (!modelName) {
|
||||
return;
|
||||
}
|
||||
if (!window.__galleryConfigs) {
|
||||
return;
|
||||
}
|
||||
|
||||
const galleryConfig = window.__galleryConfigs[modelName];
|
||||
// Check if galleryConfig exists and has at least one property
|
||||
if (!galleryConfig || Object.keys(galleryConfig).length === 0) {
|
||||
// Still update the modal title even if no config, so user can see which model they clicked
|
||||
const titleEl = document.getElementById('model-info-modal-title');
|
||||
if (titleEl) {
|
||||
titleEl.textContent = modelName;
|
||||
}
|
||||
// Show message that no info is available
|
||||
const descEl = document.getElementById('model-info-description');
|
||||
if (descEl) {
|
||||
descEl.textContent = 'No additional information available for this model.';
|
||||
}
|
||||
const linksEl = document.getElementById('model-info-links');
|
||||
if (linksEl) {
|
||||
linksEl.innerHTML = '';
|
||||
}
|
||||
const iconEl = document.getElementById('model-info-modal-icon');
|
||||
if (iconEl) {
|
||||
iconEl.style.display = 'none';
|
||||
}
|
||||
// Only open the modal if explicitly requested
|
||||
if (openModal) {
|
||||
const modalElement = document.getElementById('model-info-modal');
|
||||
if (modalElement) {
|
||||
modalElement.classList.remove('hidden');
|
||||
modalElement.setAttribute('aria-hidden', 'false');
|
||||
// Add backdrop
|
||||
let backdrop = document.querySelector('.modal-backdrop');
|
||||
if (!backdrop) {
|
||||
backdrop = document.createElement('div');
|
||||
backdrop.className = 'modal-backdrop fixed inset-0 bg-gray-900 bg-opacity-50 dark:bg-opacity-80 z-40';
|
||||
document.body.appendChild(backdrop);
|
||||
backdrop.addEventListener('click', () => {
|
||||
closeModelInfoModal();
|
||||
});
|
||||
}
|
||||
}
|
||||
}
|
||||
return;
|
||||
}
|
||||
|
||||
// Update modal title
|
||||
const titleEl = document.getElementById('model-info-modal-title');
|
||||
if (titleEl) {
|
||||
titleEl.textContent = modelName;
|
||||
}
|
||||
|
||||
// Update icon
|
||||
const iconEl = document.getElementById('model-info-modal-icon');
|
||||
if (iconEl) {
|
||||
if (galleryConfig.Icon) {
|
||||
iconEl.src = galleryConfig.Icon;
|
||||
iconEl.style.display = 'block';
|
||||
} else {
|
||||
iconEl.style.display = 'none';
|
||||
}
|
||||
}
|
||||
|
||||
// Update description
|
||||
const descEl = document.getElementById('model-info-description');
|
||||
if (descEl) {
|
||||
descEl.textContent = galleryConfig.Description || 'No description available.';
|
||||
}
|
||||
|
||||
// Update links
|
||||
const linksEl = document.getElementById('model-info-links');
|
||||
if (linksEl && galleryConfig.URLs && Array.isArray(galleryConfig.URLs) && galleryConfig.URLs.length > 0) {
|
||||
linksEl.innerHTML = '';
|
||||
galleryConfig.URLs.forEach(url => {
|
||||
const li = document.createElement('li');
|
||||
const a = document.createElement('a');
|
||||
a.href = url;
|
||||
a.target = '_blank';
|
||||
a.textContent = url;
|
||||
li.appendChild(a);
|
||||
linksEl.appendChild(li);
|
||||
});
|
||||
} else if (linksEl) {
|
||||
linksEl.innerHTML = '<li>No links available</li>';
|
||||
}
|
||||
|
||||
// Only open the modal if explicitly requested
|
||||
if (openModal) {
|
||||
const modalElement = document.getElementById('model-info-modal');
|
||||
if (modalElement) {
|
||||
// Ensure positioning classes are present (they might have been removed)
|
||||
if (!modalElement.classList.contains('flex')) {
|
||||
modalElement.classList.add('flex');
|
||||
}
|
||||
if (!modalElement.classList.contains('justify-center')) {
|
||||
modalElement.classList.add('justify-center');
|
||||
}
|
||||
if (!modalElement.classList.contains('items-center')) {
|
||||
modalElement.classList.add('items-center');
|
||||
}
|
||||
// Ensure fixed positioning
|
||||
if (!modalElement.classList.contains('fixed')) {
|
||||
modalElement.classList.add('fixed');
|
||||
}
|
||||
// Ensure full width and height
|
||||
if (!modalElement.classList.contains('w-full')) {
|
||||
modalElement.classList.add('w-full');
|
||||
}
|
||||
if (!modalElement.classList.contains('h-full')) {
|
||||
modalElement.classList.add('h-full');
|
||||
}
|
||||
// Ensure padding is set
|
||||
if (!modalElement.style.padding) {
|
||||
modalElement.style.padding = '1rem';
|
||||
}
|
||||
// Remove hidden class if present
|
||||
modalElement.classList.remove('hidden');
|
||||
// Set aria-hidden to false
|
||||
modalElement.setAttribute('aria-hidden', 'false');
|
||||
// Add backdrop if needed
|
||||
let backdrop = document.querySelector('.modal-backdrop');
|
||||
if (!backdrop) {
|
||||
backdrop = document.createElement('div');
|
||||
backdrop.className = 'modal-backdrop fixed inset-0 bg-gray-900 bg-opacity-50 dark:bg-opacity-80 z-40';
|
||||
document.body.appendChild(backdrop);
|
||||
backdrop.addEventListener('click', () => {
|
||||
window.closeModelInfoModal();
|
||||
});
|
||||
}
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
// Function to close the model info modal
|
||||
window.closeModelInfoModal = function() {
|
||||
const modalElement = document.getElementById('model-info-modal');
|
||||
if (modalElement) {
|
||||
modalElement.classList.add('hidden');
|
||||
modalElement.setAttribute('aria-hidden', 'true');
|
||||
}
|
||||
const backdrop = document.querySelector('.modal-backdrop');
|
||||
if (backdrop) {
|
||||
backdrop.remove();
|
||||
}
|
||||
};
|
||||
|
||||
// Also sync after Alpine initializes (in case it runs after DOMContentLoaded)
|
||||
function initializeModelInfo() {
|
||||
syncModelSelectorOnLoad();
|
||||
// Initialize model info modal content with current model (but don't open it)
|
||||
if (window.updateModelInfoModal && window.Alpine && window.Alpine.store("chat")) {
|
||||
const activeChat = window.Alpine.store("chat").activeChat();
|
||||
const modelName = activeChat ? activeChat.model : (document.getElementById("chat-model") ? document.getElementById("chat-model").value : null);
|
||||
if (modelName) {
|
||||
window.updateModelInfoModal(modelName, false); // false = don't open, just update content
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if (window.Alpine) {
|
||||
Alpine.nextTick(syncModelSelectorOnLoad);
|
||||
Alpine.nextTick(initializeModelInfo);
|
||||
} else {
|
||||
document.addEventListener('alpine:init', () => {
|
||||
Alpine.nextTick(syncModelSelectorOnLoad);
|
||||
Alpine.nextTick(initializeModelInfo);
|
||||
});
|
||||
}
|
||||
</script>
|
||||
|
||||
@@ -27,6 +27,9 @@ type Message struct {
|
||||
FunctionCall interface{} `json:"function_call,omitempty" yaml:"function_call,omitempty"`
|
||||
|
||||
ToolCalls []ToolCall `json:"tool_calls,omitempty" yaml:"tool_call,omitempty"`
|
||||
|
||||
// Reasoning content extracted from <thinking>...</thinking> tags
|
||||
Reasoning *string `json:"reasoning,omitempty" yaml:"reasoning,omitempty"`
|
||||
}
|
||||
|
||||
type ToolCall struct {
|
||||
@@ -78,8 +81,8 @@ func (messages Messages) ToProto() []*proto.Message {
|
||||
}
|
||||
}
|
||||
|
||||
// Note: tool_call_id and reasoning_content are not in schema.Message yet
|
||||
// They may need to be added to schema.Message if needed in the future
|
||||
// Note: tool_call_id is not in schema.Message yet
|
||||
// Reasoning field is now available in schema.Message but not yet in proto.Message
|
||||
}
|
||||
return protoMessages
|
||||
}
|
||||
|
||||
@@ -0,0 +1,114 @@
|
||||
package functions
|
||||
|
||||
import (
|
||||
"strings"
|
||||
)
|
||||
|
||||
// ExtractReasoning extracts reasoning content from thinking tags and returns
|
||||
// both the extracted reasoning and the cleaned content (with tags removed).
|
||||
// It handles <thinking>...</thinking> and <think>...</think> tags.
|
||||
// Multiple reasoning blocks are concatenated with newlines.
|
||||
func ExtractReasoning(content string) (reasoning string, cleanedContent string) {
|
||||
if content == "" {
|
||||
return "", content
|
||||
}
|
||||
|
||||
var reasoningParts []string
|
||||
var cleanedParts []string
|
||||
remaining := content
|
||||
|
||||
// Define tag pairs to look for
|
||||
tagPairs := []struct {
|
||||
start string
|
||||
end string
|
||||
}{
|
||||
{"<thinking>", "</thinking>"},
|
||||
{"<think>", "</think>"},
|
||||
}
|
||||
|
||||
// Track the last position we've processed
|
||||
lastPos := 0
|
||||
|
||||
for {
|
||||
// Find the earliest tag start
|
||||
earliestStart := -1
|
||||
earliestEnd := -1
|
||||
isUnclosed := false
|
||||
var matchedTag struct {
|
||||
start string
|
||||
end string
|
||||
}
|
||||
|
||||
for _, tagPair := range tagPairs {
|
||||
startIdx := strings.Index(remaining[lastPos:], tagPair.start)
|
||||
if startIdx == -1 {
|
||||
continue
|
||||
}
|
||||
startIdx += lastPos
|
||||
|
||||
// Find the corresponding end tag
|
||||
endIdx := strings.Index(remaining[startIdx+len(tagPair.start):], tagPair.end)
|
||||
if endIdx == -1 {
|
||||
// Unclosed tag - extract what we have
|
||||
if earliestStart == -1 || startIdx < earliestStart {
|
||||
earliestStart = startIdx
|
||||
earliestEnd = len(remaining)
|
||||
isUnclosed = true
|
||||
matchedTag = tagPair
|
||||
}
|
||||
continue
|
||||
}
|
||||
endIdx += startIdx + len(tagPair.start)
|
||||
|
||||
// Found a complete tag pair
|
||||
if earliestStart == -1 || startIdx < earliestStart {
|
||||
earliestStart = startIdx
|
||||
earliestEnd = endIdx + len(tagPair.end)
|
||||
isUnclosed = false
|
||||
matchedTag = tagPair
|
||||
}
|
||||
}
|
||||
|
||||
if earliestStart == -1 {
|
||||
// No more tags found, add remaining content
|
||||
if lastPos < len(remaining) {
|
||||
cleanedParts = append(cleanedParts, remaining[lastPos:])
|
||||
}
|
||||
break
|
||||
}
|
||||
|
||||
// Add content before the tag
|
||||
if earliestStart > lastPos {
|
||||
cleanedParts = append(cleanedParts, remaining[lastPos:earliestStart])
|
||||
}
|
||||
|
||||
// Extract reasoning content
|
||||
reasoningStart := earliestStart + len(matchedTag.start)
|
||||
// For unclosed tags, earliestEnd is already at the end of the string
|
||||
// For closed tags, earliestEnd points to after the closing tag, so we subtract the end tag length
|
||||
var reasoningEnd int
|
||||
if isUnclosed {
|
||||
// Unclosed tag - extract everything to the end
|
||||
reasoningEnd = len(remaining)
|
||||
} else {
|
||||
// Closed tag - exclude the end tag
|
||||
reasoningEnd = earliestEnd - len(matchedTag.end)
|
||||
}
|
||||
if reasoningEnd > reasoningStart {
|
||||
reasoningContent := strings.TrimSpace(remaining[reasoningStart:reasoningEnd])
|
||||
if reasoningContent != "" {
|
||||
reasoningParts = append(reasoningParts, reasoningContent)
|
||||
}
|
||||
}
|
||||
|
||||
// Move past this tag
|
||||
lastPos = earliestEnd
|
||||
}
|
||||
|
||||
// Combine reasoning parts
|
||||
reasoning = strings.Join(reasoningParts, "\n\n")
|
||||
// Combine cleaned content parts
|
||||
cleanedContent = strings.Join(cleanedParts, "")
|
||||
|
||||
return reasoning, cleanedContent
|
||||
}
|
||||
@@ -0,0 +1,261 @@
|
||||
package functions_test
|
||||
|
||||
import (
|
||||
"strings"
|
||||
|
||||
. "github.com/mudler/LocalAI/pkg/functions"
|
||||
. "github.com/onsi/ginkgo/v2"
|
||||
. "github.com/onsi/gomega"
|
||||
)
|
||||
|
||||
var _ = Describe("ExtractReasoning", func() {
|
||||
Context("when content has no reasoning tags", func() {
|
||||
It("should return empty reasoning and original content", func() {
|
||||
content := "This is regular content without any tags."
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
Expect(reasoning).To(BeEmpty())
|
||||
Expect(cleaned).To(Equal(content))
|
||||
})
|
||||
|
||||
It("should handle empty string", func() {
|
||||
content := ""
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
Expect(reasoning).To(BeEmpty())
|
||||
Expect(cleaned).To(BeEmpty())
|
||||
})
|
||||
|
||||
It("should handle content with only whitespace", func() {
|
||||
content := " \n\t "
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
Expect(reasoning).To(BeEmpty())
|
||||
Expect(cleaned).To(Equal(content))
|
||||
})
|
||||
})
|
||||
|
||||
Context("when content has <thinking> tags", func() {
|
||||
It("should extract reasoning from single thinking block", func() {
|
||||
content := "Some text <thinking>This is my reasoning</thinking> More text"
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
Expect(reasoning).To(Equal("This is my reasoning"))
|
||||
Expect(cleaned).To(Equal("Some text More text"))
|
||||
})
|
||||
|
||||
It("should extract reasoning and preserve surrounding content", func() {
|
||||
content := "Before <thinking>Reasoning here</thinking> After"
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
Expect(reasoning).To(Equal("Reasoning here"))
|
||||
Expect(cleaned).To(Equal("Before After"))
|
||||
})
|
||||
|
||||
It("should handle thinking block at the start", func() {
|
||||
content := "<thinking>Start reasoning</thinking> Regular content"
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
Expect(reasoning).To(Equal("Start reasoning"))
|
||||
Expect(cleaned).To(Equal(" Regular content"))
|
||||
})
|
||||
|
||||
It("should handle thinking block at the end", func() {
|
||||
content := "Regular content <thinking>End reasoning</thinking>"
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
Expect(reasoning).To(Equal("End reasoning"))
|
||||
Expect(cleaned).To(Equal("Regular content "))
|
||||
})
|
||||
|
||||
It("should handle only thinking block", func() {
|
||||
content := "<thinking>Only reasoning</thinking>"
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
Expect(reasoning).To(Equal("Only reasoning"))
|
||||
Expect(cleaned).To(BeEmpty())
|
||||
})
|
||||
|
||||
It("should trim whitespace from reasoning content", func() {
|
||||
content := "Text <thinking> \n Reasoning with spaces \n </thinking> More"
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
Expect(reasoning).To(Equal("Reasoning with spaces"))
|
||||
Expect(cleaned).To(Equal("Text More"))
|
||||
})
|
||||
})
|
||||
|
||||
Context("when content has <think> tags", func() {
|
||||
It("should extract reasoning from redacted_reasoning block", func() {
|
||||
content := "Text <think>Redacted reasoning</think> More"
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
Expect(reasoning).To(Equal("Redacted reasoning"))
|
||||
Expect(cleaned).To(Equal("Text More"))
|
||||
})
|
||||
|
||||
It("should handle redacted_reasoning with multiline content", func() {
|
||||
content := "Before <think>Line 1\nLine 2\nLine 3</think> After"
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
Expect(reasoning).To(Equal("Line 1\nLine 2\nLine 3"))
|
||||
Expect(cleaned).To(Equal("Before After"))
|
||||
})
|
||||
|
||||
It("should handle redacted_reasoning with complex content", func() {
|
||||
content := "Start <think>Complex reasoning\nwith\nmultiple\nlines</think> End"
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
Expect(reasoning).To(Equal("Complex reasoning\nwith\nmultiple\nlines"))
|
||||
Expect(cleaned).To(Equal("Start End"))
|
||||
})
|
||||
})
|
||||
|
||||
Context("when content has multiple reasoning blocks", func() {
|
||||
It("should concatenate multiple thinking blocks with newlines", func() {
|
||||
content := "Text <thinking>First</thinking> Middle <thinking>Second</thinking> End"
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
Expect(reasoning).To(Equal("First\n\nSecond"))
|
||||
Expect(cleaned).To(Equal("Text Middle End"))
|
||||
})
|
||||
|
||||
It("should handle multiple different tag types", func() {
|
||||
content := "A <thinking>One</thinking> B <think>Two</think> C <think>Three</think> D"
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
Expect(reasoning).To(ContainSubstring("One"))
|
||||
Expect(reasoning).To(ContainSubstring("Two"))
|
||||
Expect(reasoning).To(ContainSubstring("Three"))
|
||||
Expect(cleaned).To(Equal("A B C D"))
|
||||
})
|
||||
|
||||
It("should handle nested tags correctly (extracts first match)", func() {
|
||||
content := "Text <thinking>Outer <think>Inner</think></thinking> More"
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
// Should extract the outer thinking block
|
||||
Expect(reasoning).To(ContainSubstring("Outer"))
|
||||
Expect(reasoning).To(ContainSubstring("Inner"))
|
||||
Expect(cleaned).To(Equal("Text More"))
|
||||
})
|
||||
})
|
||||
|
||||
Context("when content has unclosed reasoning tags", func() {
|
||||
It("should extract unclosed thinking block", func() {
|
||||
content := "Text <thinking>Unclosed reasoning"
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
Expect(reasoning).To(Equal("Unclosed reasoning"))
|
||||
Expect(cleaned).To(Equal("Text "))
|
||||
})
|
||||
|
||||
It("should extract unclosed think block", func() {
|
||||
content := "Before <think>Incomplete"
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
Expect(reasoning).To(Equal("Incomplete"))
|
||||
Expect(cleaned).To(Equal("Before "))
|
||||
})
|
||||
|
||||
It("should extract unclosed redacted_reasoning block", func() {
|
||||
content := "Start <think>Partial reasoning content"
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
Expect(reasoning).To(Equal("Partial reasoning content"))
|
||||
Expect(cleaned).To(Equal("Start "))
|
||||
})
|
||||
|
||||
It("should handle unclosed tag at the end", func() {
|
||||
content := "Regular content <thinking>Unclosed at end"
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
Expect(reasoning).To(Equal("Unclosed at end"))
|
||||
Expect(cleaned).To(Equal("Regular content "))
|
||||
})
|
||||
})
|
||||
|
||||
Context("when content has empty reasoning blocks", func() {
|
||||
It("should ignore empty thinking block", func() {
|
||||
content := "Text <thinking></thinking> More"
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
Expect(reasoning).To(BeEmpty())
|
||||
Expect(cleaned).To(Equal("Text More"))
|
||||
})
|
||||
|
||||
It("should ignore thinking block with only whitespace", func() {
|
||||
content := "Text <thinking> \n\t </thinking> More"
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
Expect(reasoning).To(BeEmpty())
|
||||
Expect(cleaned).To(Equal("Text More"))
|
||||
})
|
||||
})
|
||||
|
||||
Context("when content has reasoning tags with special characters", func() {
|
||||
It("should handle reasoning with newlines", func() {
|
||||
content := "Before <thinking>Line 1\nLine 2\nLine 3</thinking> After"
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
Expect(reasoning).To(Equal("Line 1\nLine 2\nLine 3"))
|
||||
Expect(cleaned).To(Equal("Before After"))
|
||||
})
|
||||
|
||||
It("should handle reasoning with code blocks", func() {
|
||||
content := "Text <thinking>Reasoning with ```code``` blocks</thinking> More"
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
Expect(reasoning).To(Equal("Reasoning with ```code``` blocks"))
|
||||
Expect(cleaned).To(Equal("Text More"))
|
||||
})
|
||||
|
||||
It("should handle reasoning with JSON", func() {
|
||||
content := "Before <think>{\"key\": \"value\"}</think> After"
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
Expect(reasoning).To(Equal("{\"key\": \"value\"}"))
|
||||
Expect(cleaned).To(Equal("Before After"))
|
||||
})
|
||||
|
||||
It("should handle reasoning with HTML-like content", func() {
|
||||
content := "Text <thinking>Reasoning with <tags> inside</thinking> More"
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
Expect(reasoning).To(Equal("Reasoning with <tags> inside"))
|
||||
Expect(cleaned).To(Equal("Text More"))
|
||||
})
|
||||
})
|
||||
|
||||
Context("when content has reasoning mixed with regular content", func() {
|
||||
It("should preserve content order correctly", func() {
|
||||
content := "Start <thinking>Reasoning</thinking> Middle <think>More reasoning</think> End"
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
Expect(reasoning).To(ContainSubstring("Reasoning"))
|
||||
Expect(reasoning).To(ContainSubstring("More reasoning"))
|
||||
Expect(cleaned).To(Equal("Start Middle End"))
|
||||
})
|
||||
|
||||
It("should handle reasoning in the middle of a sentence", func() {
|
||||
content := "This is a <thinking>reasoning</thinking> sentence."
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
Expect(reasoning).To(Equal("reasoning"))
|
||||
Expect(cleaned).To(Equal("This is a sentence."))
|
||||
})
|
||||
})
|
||||
|
||||
Context("edge cases", func() {
|
||||
It("should handle content with only opening tag", func() {
|
||||
content := "<thinking>"
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
Expect(reasoning).To(BeEmpty())
|
||||
Expect(cleaned).To(Equal(""))
|
||||
})
|
||||
|
||||
It("should handle content with only closing tag", func() {
|
||||
content := "</thinking>"
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
Expect(reasoning).To(BeEmpty())
|
||||
Expect(cleaned).To(Equal("</thinking>"))
|
||||
})
|
||||
|
||||
It("should handle mismatched tags", func() {
|
||||
content := "<thinking>Content</think>"
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
// Should extract unclosed thinking block
|
||||
Expect(reasoning).To(ContainSubstring("Content"))
|
||||
Expect(cleaned).To(Equal(""))
|
||||
})
|
||||
|
||||
It("should handle very long reasoning content", func() {
|
||||
longReasoning := strings.Repeat("This is reasoning content. ", 100)
|
||||
content := "Text <thinking>" + longReasoning + "</thinking> More"
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
// TrimSpace is applied, so we need to account for that
|
||||
Expect(reasoning).To(Equal(strings.TrimSpace(longReasoning)))
|
||||
Expect(cleaned).To(Equal("Text More"))
|
||||
})
|
||||
|
||||
It("should handle reasoning with unicode characters", func() {
|
||||
content := "Text <thinking>Reasoning with 中文 and emoji 🧠</thinking> More"
|
||||
reasoning, cleaned := ExtractReasoning(content)
|
||||
Expect(reasoning).To(Equal("Reasoning with 中文 and emoji 🧠"))
|
||||
Expect(cleaned).To(Equal("Text More"))
|
||||
})
|
||||
})
|
||||
})
|
||||
Reference in New Issue
Block a user