alex/LocalAGI
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: improve action picking

Browse Source 
Signed-off-by: mudler <mudler@localai.io>
This commit is contained in:
mudler
2025-04-11 20:43:48 +02:00
parent 9dad2b0ba4
commit e4271b4d2f
6 changed files with 200 additions and 141 deletions
Download Patch File Download Diff File Expand all files Collapse all files

49
core/action/goal.go Normal file
View File

@@ -0,0 +1,49 @@
package action
import (
"context"
"github.com/mudler/LocalAGI/core/types"
"github.com/sashabaranov/go-openai/jsonschema"
)
// NewGoal creates a new intention action
// The inention action is special as it tries to identify
// a tool to use and a reasoning over to use it
func NewGoal(s ...string) *GoalAction {
return &GoalAction{tools: s}
}
type GoalAction struct {
tools []string
}
type GoalResponse struct {
Goal string `json:"goal"`
Achieved bool `json:"achieved"`
}
func (a *GoalAction) Run(context.Context, types.ActionParams) (types.ActionResult, error) {
return types.ActionResult{}, nil
}
func (a *GoalAction) Plannable() bool {
return false
}
func (a *GoalAction) Definition() types.ActionDefinition {
return types.ActionDefinition{
Name: "goal",
Description: "Check if the goal is achieved",
Properties: map[string]jsonschema.Definition{
"goal": {
Type: jsonschema.String,
Description: "The goal to check if it is achieved.",
},
"achieved": {
Type: jsonschema.Boolean,
Description: "Whether the goal is achieved",
},
},
Required: []string{"goal", "achieved"},
}
}

141
core/agent/actions.go
View File

@@ -5,7 +5,6 @@ import (
"encoding/json"
"fmt"
"os"
"strings"
"github.com/mudler/LocalAGI/core/action"
"github.com/mudler/LocalAGI/core/types"
@@ -80,6 +79,15 @@ func (m Messages) ToOpenAI() []openai.ChatCompletionMessage {
return []openai.ChatCompletionMessage(m)
}
func (m Messages) RemoveIf(f func(msg openai.ChatCompletionMessage) bool) Messages {
for i := len(m) - 1; i >= 0; i-- {
if f(m[i]) {
m = append(m[:i], m[i+1:]...)
}
}
return m
}
func (m Messages) String() string {
s := ""
for _, cc := range m {
@@ -412,27 +420,26 @@ func (a *Agent) pickAction(ctx context.Context, templ string, messages []openai.
actionsID = append(actionsID, m.Definition().Name.String())
}
thoughtPromptStringBuilder := strings.Builder{}
thoughtPromptStringBuilder.WriteString("You have to pick an action based on the conversation and the prompt. Describe the full reasoning process for your choice. Here is a list of actions: ")
for _, m := range a.availableActions() {
thoughtPromptStringBuilder.WriteString(
m.Definition().Name.String() + ": " + m.Definition().Description + "\n",
)
}
// thoughtPromptStringBuilder := strings.Builder{}
// thoughtPromptStringBuilder.WriteString("You have to pick an action based on the conversation and the prompt. Describe the full reasoning process for your choice. Here is a list of actions: ")
// for _, m := range a.availableActions() {
// thoughtPromptStringBuilder.WriteString(
// m.Definition().Name.String() + ": " + m.Definition().Description + "\n",
// )
// }
thoughtPromptStringBuilder.WriteString("To not use any action, respond with 'none'")
// thoughtPromptStringBuilder.WriteString("To not use any action, respond with 'none'")
//thoughtPromptStringBuilder.WriteString("\n\nConversation: " + Messages(c).String())
//thoughtPromptStringBuilder.WriteString("\n\nConversation: " + Messages(c).RemoveIf(func(msg openai.ChatCompletionMessage) bool {
// return msg.Role == "system"
//}).String())
thoughtPrompt := thoughtPromptStringBuilder.String()
//thoughtPrompt := thoughtPromptStringBuilder.String()
xlog.Debug("[pickAction] thought", "prompt", thoughtPrompt)
//thoughtConv := []openai.ChatCompletionMessage{}
thought, err := a.askLLM(ctx,
append(c, openai.ChatCompletionMessage{
Role: "user",
Content: thoughtPrompt,
}),
c,
maxRetries,
)
if err != nil {
@@ -440,71 +447,89 @@ func (a *Agent) pickAction(ctx context.Context, templ string, messages []openai.
}
originalReasoning := thought.Content
xlog.Debug("[pickAction] thought", "reason", originalReasoning)
thought, err = a.askLLM(ctx,
[]openai.ChatCompletionMessage{
{
Role: "system",
Content: "Your only objective is to return the string of the action to take based on the user input",
},
{
Role: "user",
Content: "Given the following sentence, answer with the only action to take: " + originalReasoning,
}},
maxRetries,
)
if err != nil {
return nil, nil, "", err
}
reason := thought.Content
xlog.Debug("[pickAction] filtered thought", "reason", reason)
// From the thought, get the action call
// Get all the available actions IDs
intentionsTools := action.NewIntention(append(actionsID, "none")...)
// by grammar, let's decide if we have achieved the goal
// 1. analyze response and check if goal is achieved
// NOTE: we do not give the full conversation here to pick the action
// to avoid hallucinations
params, err := a.decision(ctx,
[]openai.ChatCompletionMessage{
{
Role: "system",
Content: "You have to pick the correct action based on the user reasoning",
Content: "Extract an action to perform from the following reasoning: ",
},
{
Role: "user",
Content: reason,
},
},
types.Actions{intentionsTools}.ToTools(),
intentionsTools.Definition().Name, maxRetries)
Content: originalReasoning,
}},
types.Actions{action.NewGoal()}.ToTools(),
action.NewGoal().Definition().Name, maxRetries)
if err != nil {
return nil, nil, "", fmt.Errorf("failed to get the action tool parameters: %v", err)
}
actionChoice := action.IntentResponse{}
if params.actionParams == nil {
return nil, nil, params.message, nil
}
err = params.actionParams.Unmarshal(&actionChoice)
goalResponse := action.GoalResponse{}
err = params.actionParams.Unmarshal(&goalResponse)
if err != nil {
return nil, nil, "", err
}
if actionChoice.Tool == "" || actionChoice.Tool == "none" {
if goalResponse.Achieved {
xlog.Debug("[pickAction] goal achieved", "goal", goalResponse.Goal)
return nil, nil, "", nil
}
// Find the action
chosenAction := a.availableActions().Find(actionChoice.Tool)
if chosenAction == nil {
return nil, nil, "", fmt.Errorf("no action found for intent:" + actionChoice.Tool)
// if the goal is not achieved, pick an action
xlog.Debug("[pickAction] goal not achieved", "goal", goalResponse.Goal)
xlog.Debug("[pickAction] thought", "conv", c, "originalReasoning", originalReasoning)
// TODO: FORCE to select ana ction here
// NOTE: we do not give the full conversation here to pick the action
// to avoid hallucinations
params, err = a.decision(ctx,
[]openai.ChatCompletionMessage{
{
Role: "system",
Content: "Extract an action to perform from the following reasoning: ",
},
{
Role: "user",
Content: originalReasoning,
}},
a.availableActions().ToTools(),
nil, maxRetries)
if err != nil {
return nil, nil, "", fmt.Errorf("failed to get the action tool parameters: %v", err)
}
chosenAction := a.availableActions().Find(params.actioName)
// xlog.Debug("[pickAction] params", "params", params)
// if params.actionParams == nil {
// return nil, nil, params.message, nil
// }
// xlog.Debug("[pickAction] actionChoice", "actionChoice", params.actionParams, "message", params.message)
// actionChoice := action.IntentResponse{}
// err = params.actionParams.Unmarshal(&actionChoice)
// if err != nil {
// return nil, nil, "", err
// }
// if actionChoice.Tool == "" || actionChoice.Tool == "none" {
// return nil, nil, "", nil
// }
// // Find the action
// chosenAction := a.availableActions().Find(actionChoice.Tool)
// if chosenAction == nil {
// return nil, nil, "", fmt.Errorf("no action found for intent:" + actionChoice.Tool)
// }
return chosenAction, nil, originalReasoning, nil
}

19
core/agent/agent.go
View File

@@ -515,10 +515,21 @@ func (a *Agent) consumeJob(job *types.Job, role string) {
//job.Result.Finish(fmt.Errorf("no action to do"))\
xlog.Info("No action to do, just reply", "agent", a.Character.Name, "reasoning", reasoning)
conv = append(conv, openai.ChatCompletionMessage{
Role: "assistant",
Content: reasoning,
})
if reasoning != "" {
conv = append(conv, openai.ChatCompletionMessage{
Role: "assistant",
Content: reasoning,
})
} else {
xlog.Info("No reasoning, just reply", "agent", a.Character.Name)
msg, err := a.askLLM(job.GetContext(), conv, maxRetries)
if err != nil {
job.Result.Finish(fmt.Errorf("error asking LLM for a reply: %w", err))
return
}
conv = append(conv, msg)
reasoning = msg.Content
}
xlog.Debug("Finish job with reasoning", "reasoning", reasoning, "agent", a.Character.Name, "conversation", fmt.Sprintf("%+v", conv))
job.Result.Conversation = conv

80
core/agent/templates.go
View File

@@ -82,11 +82,7 @@ Current State:
- Short-term Memory: {{range .CurrentState.Memories}}{{.}} {{end}}{{end}}
Current Time: {{.Time}}`
const pickSelfTemplate = `Available Tools:
{{range .Actions -}}
- {{.Name}}: {{.Description }}
{{ end }}
const pickSelfTemplate = `
You are an autonomous AI agent with a defined character and state (as shown above).
Your task is to evaluate your current situation and determine the best course of action.
@@ -108,40 +104,21 @@ Remember:
- Keep track of your progress and state
- Be proactive in addressing potential issues
{{if .Reasoning}}Previous Reasoning: {{.Reasoning}}{{end}}
` + hudTemplate
const reSelfEvalTemplate = pickSelfTemplate + `
Previous actions have been executed. Evaluate the current situation:
1. Review the outcomes of previous actions
2. Assess progress toward your goals
3. Identify any issues or challenges
4. Determine if additional actions are needed
Consider:
- Success of previous actions
- Changes in the situation
- New information or insights
- Potential next steps
Make a decision about whether to:
- Continue with more actions
- Provide a final response
- Adjust your approach
- Update your goals or state`
const pickActionTemplate = hudTemplate + `
Available Tools:
{{range .Actions -}}
- {{.Name}}: {{.Description }}
{{ end }}
Task: Analyze the situation and determine the best course of action.
{{if .Reasoning}}Previous Reasoning: {{.Reasoning}}{{end}}
` + hudTemplate
const reSelfEvalTemplate = pickSelfTemplate
const pickActionTemplate = hudTemplate + `
Your only task is to analyze the situation and determine a goal and the best tool to use, or just a final response if we have fullfilled the goal.
Guidelines:
1. Review the current state and context
1. Review the current state, what was done already and context
2. Consider available tools and their purposes
3. Plan your approach carefully
4. Explain your reasoning clearly
@@ -159,38 +136,11 @@ Decision Process:
4. Explain your reasoning
5. Execute the chosen action
Available Tools:
{{range .Actions -}}
- {{.Name}}: {{.Description }}
{{ end }}
{{if .Reasoning}}Previous Reasoning: {{.Reasoning}}{{end}}`
const reEvalTemplate = pickActionTemplate + `
Previous actions have been executed. Let's evaluate the current situation:
1. Review Previous Actions:
- What actions were taken
- What were the results
- Any issues or challenges encountered
2. Assess Current State:
- Progress toward goals
- Changes in the situation
- New information or insights
- Current challenges or opportunities
3. Determine Next Steps:
- Additional tools needed
- Final response required
- Error handling needed
- Approach adjustments required
4. Decision Making:
- If task is complete: Use "reply" tool
- If errors exist: Address them appropriately
- If more actions needed: Explain why and which tools
- If situation changed: Adapt your approach
Remember to:
- Consider all available information
- Be specific about next steps
- Explain your reasoning clearly
- Handle errors appropriately
- Provide complete responses when done`
const reEvalTemplate = pickActionTemplate