Enable reasoning separately

action/intention.go

@@ -36,6 +36,7 @@ func (a *IntentAction) Definition() ActionDefinition {
 			},
 			"tool": {
 				Type:        jsonschema.String,
+				Description: "The tool you want to use",
 				Enum:        a.tools,
 			},
 		},

agent/actions.go

@@ -52,6 +52,7 @@ func (a Actions) Find(name string) Action {
 type decisionResult struct {
 	actionParams action.ActionParams
 	message      string
+	actioName    string
 }
 
 // decision forces the agent to take one of the available actions
@@ -85,7 +86,7 @@ func (a *Agent) decision(
 		return nil, err
 	}
 
-	return &decisionResult{actionParams: params}, nil
+	return &decisionResult{actionParams: params, actioName: msg.ToolCalls[0].Function.Name}, nil
 }
 
 type Messages []openai.ChatCompletionMessage
@@ -130,11 +131,16 @@ func (a *Agent) generateParameters(ctx context.Context, pickTemplate string, act
 		}, conversation...)
 	}
 
-	return a.decision(ctx,
-		append(conversation, openai.ChatCompletionMessage{
+	cc := conversation
+	if a.options.forceReasoning {
+		cc = append(conversation, openai.ChatCompletionMessage{
 			Role:    "system",
 			Content: fmt.Sprintf("The agent decided to use the tool %s with the following reasoning: %s", act.Definition().Name, reasoning),
-		}),
+		})
+	}
+
+	return a.decision(ctx,
+		cc,
 		a.systemInternalActions().ToTools(),
 		openai.ToolChoice{
 			Type:     openai.ToolTypeFunction,
@@ -187,6 +193,24 @@ func (a *Agent) prepareHUD() PromptHUD {
 func (a *Agent) pickAction(ctx context.Context, templ string, messages []openai.ChatCompletionMessage) (Action, string, error) {
 	c := messages
 
+	if !a.options.forceReasoning {
+		// We also could avoid to use functions here and get just a reply from the LLM
+		// and then use the reply to get the action
+		thought, err := a.decision(ctx,
+			messages,
+			a.systemInternalActions().ToTools(),
+			nil)
+		if err != nil {
+			return nil, "", err
+		}
+		// Find the action
+		chosenAction := a.systemInternalActions().Find(thought.actioName)
+		if chosenAction == nil {
+			return nil, "", fmt.Errorf("no action found for intent:" + thought.actioName)
+		}
+		return chosenAction, thought.message, nil
+	}
+
 	var promptHUD *PromptHUD
 	if a.options.enableHUD {
 		h := a.prepareHUD()
@@ -241,7 +265,7 @@ func (a *Agent) pickAction(ctx context.Context, templ string, messages []openai.
 	params, err := a.decision(ctx,
 		append(c, openai.ChatCompletionMessage{
 			Role:    "system",
-			Content: "The assistant thought: " + reason,
+			Content: "Given the assistant thought, pick the relevant action: " + reason,
 		}),
 		Actions{intentionsTools}.ToTools(),
 		intentionsTools.Definition().Name)

agent/options.go

@@ -23,6 +23,7 @@ type options struct {
 
 	canStopItself         bool
 	initiateConversations bool
+	forceReasoning        bool
 	characterfile         string
 	statefile             string
 	context               context.Context
@@ -78,6 +79,11 @@ var EnableHUD = func(o *options) error {
 	return nil
 }
 
+var EnableForceReasoning = func(o *options) error {
+	o.forceReasoning = true
+	return nil
+}
+
 var EnableKnowledgeBase = func(o *options) error {
 	o.enableKB = true
 	o.kbResults = 5

agent/templates.go

@@ -84,12 +84,12 @@ const reSelfEvalTemplate = pickSelfTemplate + `
 We already have called other tools. Evaluate the current situation and decide if we need to execute other tools.`
 
 const pickActionTemplate = hudTemplate + `
-You can take any of the following tools: 
+When deciding what to do, consider that you can take any of the following tools: 
 
 {{range .Actions -}}
 - {{.Name}}: {{.Description }}
 {{ end }}
-To answer back to the user, use the "reply" tool.
+To answer back to the user, use the "reply" or the "answer" tool.
 Given the text below, decide which action to take and explain the detailed reasoning behind it. For answering without picking a choice, reply with 'none'.
 
 {{if .Reasoning}}Reasoning: {{.Reasoning}}{{end}}

example/webui/agentconfig.go

@@ -24,6 +24,7 @@ type AgentConfig struct {
 	PeriodicRuns          string `json:"periodic_runs" form:"periodic_runs"`
 	PermanentGoal         string `json:"permanent_goal" form:"permanent_goal"`
 	EnableKnowledgeBase   bool   `json:"enable_kb" form:"enable_kb"`
+	EnableReasoning       bool   `json:"enable_reasoning" form:"enable_reasoning"`
 	KnowledgeBaseResults  int    `json:"kb_results" form:"kb_results"`
 	CanStopItself         bool   `json:"can_stop_itself" form:"can_stop_itself"`
 	SystemPrompt          string `json:"system_prompt" form:"system_prompt"`

example/webui/agentpool.go

@@ -265,6 +265,9 @@ func (a *AgentPool) startAgentWithConfig(name string, config *AgentConfig) error
 	if config.EnableKnowledgeBase {
 		opts = append(opts, EnableKnowledgeBase)
 	}
+	if config.EnableReasoning {
+		opts = append(opts, EnableForceReasoning)
+	}
 
 	if config.KnowledgeBaseResults > 0 {
 		opts = append(opts, EnableKnowledgeBaseWithResults(config.KnowledgeBaseResults))

example/webui/views/create.html

@@ -86,6 +86,8 @@
         <div class="mb-4">
             <label for="enable_kb" class="block text-lg font-medium text-gray-400">Enable Knowledge Base</label>
             <input type="checkbox" name="enable_kb" id="enable_kb" class="mt-1 focus:ring-indigo-500 focus:border-indigo-500 block w-full shadow-sm sm:text-lg border-gray-300 rounded-md bg-gray-700 text-white">  
+            <label for="enable_reasoning" class="block text-lg font-medium text-gray-400">Enable Reasoning</label>
+            <input type="checkbox" name="enable_reasoning" id="enable_reasoning" class="mt-1 focus:ring-indigo-500 focus:border-indigo-500 block w-full shadow-sm sm:text-lg border-gray-300 rounded-md bg-gray-700 text-white">  
             <div class="mb-6">
                 <label for="kb_results" class="block text-lg font-medium text-gray-400">Knowledge base results</label>
                 <input type="text" name="kb_results" id="kb_results" class="mt-1 focus:ring-indigo-500 focus:border-indigo-500 block w-full shadow-sm sm:text-lg border-gray-300 rounded-md bg-gray-700 text-white" placeholder="3">