feat: add option to strip thinking from output (#100)

Signed-off-by: mudler <mudler@localai.io>

core/agent/agent.go

@@ -5,6 +5,8 @@ import (
 	"errors"
 	"fmt"
 	"os"
+	"regexp"
+	"strings"
 	"sync"
 	"time"
 
@@ -788,6 +790,24 @@ func (a *Agent) consumeJob(job *types.Job, role string) {
 	a.reply(job, role, conv, actionParams, chosenAction, reasoning)
 }
 
+func stripThinkingTags(content string) string {
+	// Remove content between <thinking> and </thinking>
+	content = regexp.MustCompile(`<thinking>.*?</thinking>`).ReplaceAllString(content, "")
+	// Remove content between <think> and </think>
+	content = regexp.MustCompile(`<think>.*?</think>`).ReplaceAllString(content, "")
+	// Clean up any extra whitespace
+	content = strings.TrimSpace(content)
+	return content
+}
+
+func (a *Agent) processPrompt(content string) string {
+	if a.options.stripThinkingTags {
+		content = stripThinkingTags(content)
+	}
+	// Future post-processing options can be added here
+	return content
+}
+
 func (a *Agent) reply(job *types.Job, role string, conv Messages, actionParams types.ActionParams, chosenAction types.Action, reasoning string) {
 	job.Result.Conversation = conv
 
@@ -854,7 +874,7 @@ func (a *Agent) reply(job *types.Job, role string, conv Messages, actionParams t
 
 		msg := openai.ChatCompletionMessage{
 			Role:    "assistant",
-			Content: replyResponse.Message,
+			Content: a.processPrompt(replyResponse.Message),
 		}
 
 		conv = append(conv, msg)
@@ -883,8 +903,10 @@ func (a *Agent) reply(job *types.Job, role string, conv Messages, actionParams t
 
 		msg = openai.ChatCompletionMessage{
 			Role:    "assistant",
-			Content: replyResponse.Message,
+			Content: a.processPrompt(replyResponse.Message),
 		}
+	} else {
+		msg.Content = a.processPrompt(msg.Content)
 	}
 
 	conv = append(conv, msg)

core/agent/options.go

@@ -25,6 +25,7 @@ type options struct {
 	randomIdentity                                                                               bool
 	userActions                                                                                  types.Actions
 	enableHUD, standaloneJob, showCharacter, enableKB, enableSummaryMemory, enableLongTermMemory bool
+	stripThinkingTags                                                                            bool
 
 	canStopItself         bool
 	initiateConversations bool
@@ -377,3 +378,8 @@ func WithObserver(observer Observer) Option {
 		return nil
 	}
 }
+
+var EnableStripThinkingTags = func(o *options) error {
+	o.stripThinkingTags = true
+	return nil
+}

core/state/config.go

@@ -67,6 +67,7 @@ type AgentConfig struct {
 	LongTermMemory        bool   `json:"long_term_memory" form:"long_term_memory"`
 	SummaryLongTermMemory bool   `json:"summary_long_term_memory" form:"summary_long_term_memory"`
 	ParallelJobs          int    `json:"parallel_jobs" form:"parallel_jobs"`
+	StripThinkingTags     bool   `json:"strip_thinking_tags" form:"strip_thinking_tags"`
 }
 
 type AgentConfigMeta struct {
@@ -292,6 +293,14 @@ func NewAgentConfigMeta(
 				HelpText:     "Script to prepare the MCP box",
 				Tags:         config.Tags{Section: "AdvancedSettings"},
 			},
+			{
+				Name:         "strip_thinking_tags",
+				Label:        "Strip Thinking Tags",
+				Type:         "checkbox",
+				DefaultValue: false,
+				HelpText:     "Remove content between <thinking></thinking> and <think></think> tags from agent responses",
+				Tags:         config.Tags{Section: "ModelSettings"},
+			},
 		},
 		MCPServers: []config.Field{
 			{

core/state/pool.go

@@ -526,6 +526,10 @@ func (a *AgentPool) startAgentWithConfig(name string, config *AgentConfig, obs O
 		opts = append(opts, EnableForceReasoning)
 	}
 
+	if config.StripThinkingTags {
+		opts = append(opts, EnableStripThinkingTags)
+	}
+
 	if config.KnowledgeBaseResults > 0 {
 		opts = append(opts, EnableKnowledgeBaseWithResults(config.KnowledgeBaseResults))
 	}