414 lines
11 KiB
Go
414 lines
11 KiB
Go
package agent
|
||
|
||
import (
|
||
"context"
|
||
"encoding/json"
|
||
"fmt"
|
||
"sort"
|
||
"strings"
|
||
|
||
"laodingbot/internal/knowledge"
|
||
"laodingbot/internal/llm"
|
||
"laodingbot/internal/logger"
|
||
"laodingbot/internal/memory"
|
||
"laodingbot/internal/tools"
|
||
)
|
||
|
||
type Orchestrator struct {
|
||
llm llm.Client
|
||
store *memory.SQLiteStore
|
||
tools *tools.Registry
|
||
soul string
|
||
skills []knowledge.Skill
|
||
skillsDoc string
|
||
reactMaxStep int
|
||
log *logger.Logger
|
||
}
|
||
|
||
func NewOrchestrator(
|
||
llmClient llm.Client,
|
||
store *memory.SQLiteStore,
|
||
registry *tools.Registry,
|
||
soul string,
|
||
skills []knowledge.Skill,
|
||
skillsDoc string,
|
||
reactMaxStep int,
|
||
log *logger.Logger,
|
||
) *Orchestrator {
|
||
if reactMaxStep <= 0 {
|
||
reactMaxStep = 4
|
||
}
|
||
return &Orchestrator{
|
||
llm: llmClient,
|
||
store: store,
|
||
tools: registry,
|
||
soul: soul,
|
||
skills: skills,
|
||
skillsDoc: skillsDoc,
|
||
reactMaxStep: reactMaxStep,
|
||
log: log,
|
||
}
|
||
}
|
||
|
||
func (o *Orchestrator) HandleMessage(ctx context.Context, chatID, userID, text string) (string, error) {
|
||
if o.log != nil {
|
||
o.log.Infof("handle message chat_id=%s user_id=%s text_len=%d", chatID, userID, len(text))
|
||
o.log.Debugf("handle message text=%q", text)
|
||
}
|
||
if err := o.store.SaveMessage(chatID, userID, "user", text); err != nil {
|
||
if o.log != nil {
|
||
o.log.Errorf("save user message failed chat_id=%s err=%v", chatID, err)
|
||
}
|
||
return "", err
|
||
}
|
||
|
||
recent, err := o.store.LoadRecent(chatID, 16)
|
||
if err != nil {
|
||
if o.log != nil {
|
||
o.log.Errorf("load recent failed chat_id=%s err=%v", chatID, err)
|
||
}
|
||
return "", err
|
||
}
|
||
compressed := memory.CompressForPrompt(recent, 6000)
|
||
if o.log != nil {
|
||
o.log.Debugf("prompt context prepared chat_id=%s recent_count=%d compressed_len=%d", chatID, len(recent), len(compressed))
|
||
}
|
||
|
||
matchedSkills := o.matchSkills(ctx, compressed, text)
|
||
|
||
var response string
|
||
if len(matchedSkills) == 0 {
|
||
if o.log != nil {
|
||
o.log.Infof("no skill matched; use direct llm chat_id=%s", chatID)
|
||
}
|
||
response, err = o.runDirectLLM(ctx, compressed, text)
|
||
} else {
|
||
if o.log != nil {
|
||
names := make([]string, 0, len(matchedSkills))
|
||
for _, s := range matchedSkills {
|
||
names = append(names, s.Name)
|
||
o.log.Infof("skill selected name=%s source=%s", s.Name, s.Source)
|
||
o.log.Debugf("skill selected content name=%s content=%q", s.Name, s.Content)
|
||
}
|
||
o.log.Infof("skills matched chat_id=%s skills=%s", chatID, strings.Join(names, ","))
|
||
}
|
||
response, err = o.runReAct(ctx, compressed, text, matchedSkills)
|
||
}
|
||
if err != nil {
|
||
if o.log != nil {
|
||
o.log.Errorf("message generation failed chat_id=%s err=%v", chatID, err)
|
||
}
|
||
return "", err
|
||
}
|
||
|
||
if err := o.store.SaveMessage(chatID, userID, "assistant", response); err != nil {
|
||
if o.log != nil {
|
||
o.log.Errorf("save assistant response failed chat_id=%s err=%v", chatID, err)
|
||
}
|
||
return "", err
|
||
}
|
||
if o.log != nil {
|
||
o.log.Infof("message handled chat_id=%s response_len=%d", chatID, len(response))
|
||
}
|
||
return response, nil
|
||
}
|
||
|
||
func (o *Orchestrator) runDirectLLM(ctx context.Context, compressedContext, userInput string) (string, error) {
|
||
systemPrompt := strings.Join([]string{
|
||
"你是一个个人自动化助手，必须遵循如下人格设定并保持一致：",
|
||
o.soul,
|
||
"",
|
||
"如果当前问题没有匹配到已定义技能，请直接回答用户。",
|
||
"当你判断必须依赖外部工具结果才能可靠回答时，请明确告知用户需要进一步操作信息。",
|
||
}, "\n")
|
||
|
||
userPrompt := strings.Join([]string{
|
||
"历史上下文：",
|
||
compressedContext,
|
||
"",
|
||
"用户问题：",
|
||
userInput,
|
||
}, "\n")
|
||
|
||
return o.llm.Generate(ctx, systemPrompt, userPrompt)
|
||
}
|
||
|
||
type reactDecision struct {
|
||
Thought string `json:"thought"`
|
||
Action string `json:"action"`
|
||
ActionInput string `json:"action_input"`
|
||
Final string `json:"final"`
|
||
}
|
||
|
||
func (o *Orchestrator) runReAct(ctx context.Context, compressedContext, userInput string, selectedSkills []knowledge.Skill) (string, error) {
|
||
selectedSkillsDoc := formatSkills(selectedSkills)
|
||
toolDoc := o.formatToolDoc()
|
||
if o.log != nil {
|
||
names := make([]string, 0, len(selectedSkills))
|
||
for _, s := range selectedSkills {
|
||
names = append(names, s.Name)
|
||
}
|
||
o.log.Infof("react start steps=%d skills=%s", o.reactMaxStep, strings.Join(names, ","))
|
||
o.log.Debugf("react selected_skills_doc=%q", selectedSkillsDoc)
|
||
o.log.Debugf("react tools_doc=%q", toolDoc)
|
||
}
|
||
|
||
systemPrompt := strings.Join([]string{
|
||
"你是一个个人自动化助手，必须遵循如下人格设定并保持一致：",
|
||
o.soul,
|
||
"",
|
||
"已匹配到的 skills（只可按下列技能执行）：",
|
||
selectedSkillsDoc,
|
||
"",
|
||
"可用工具：",
|
||
toolDoc,
|
||
"",
|
||
"你必须使用 ReAct 模式做决策。",
|
||
"只有当技能明确需要工具能力时才调用工具。",
|
||
"如果问题可直接回答，不要调用工具。",
|
||
"你的输出必须是 JSON，对象字段为 thought, action, action_input, final。",
|
||
"规则：",
|
||
"1) 当需要调工具时：final 置空，action 必须是可用工具之一，action_input 为工具输入。",
|
||
"2) 当可以最终回答时：action 置 none，action_input 置空，final 填最终回复。",
|
||
"3) 不要输出 JSON 之外内容。",
|
||
}, "\n")
|
||
|
||
scratchpad := ""
|
||
for step := 1; step <= o.reactMaxStep; step++ {
|
||
if o.log != nil {
|
||
o.log.Infof("react step start step=%d/%d", step, o.reactMaxStep)
|
||
o.log.Debugf("react scratchpad_before step=%d content=%q", step, scratchpad)
|
||
}
|
||
prompt := strings.Join([]string{
|
||
"历史上下文：",
|
||
compressedContext,
|
||
"",
|
||
"用户问题：",
|
||
userInput,
|
||
"",
|
||
"当前推理记录（按时间顺序）：",
|
||
scratchpad,
|
||
"",
|
||
fmt.Sprintf("请输出下一步 JSON 决策。当前步骤: %d/%d", step, o.reactMaxStep),
|
||
}, "\n")
|
||
|
||
raw, err := o.llm.Generate(ctx, systemPrompt, prompt)
|
||
if err != nil {
|
||
return "", err
|
||
}
|
||
if o.log != nil {
|
||
o.log.Infof("react step llm output step=%d raw=%q", step, raw)
|
||
}
|
||
decision, err := parseDecision(raw)
|
||
if err != nil {
|
||
if o.log != nil {
|
||
o.log.Warnf("react parse failed, use raw as final err=%v", err)
|
||
}
|
||
return strings.TrimSpace(raw), nil
|
||
}
|
||
if o.log != nil {
|
||
o.log.Infof("react step decision step=%d thought=%q action=%q action_input=%q final=%q", step, decision.Thought, decision.Action, decision.ActionInput, decision.Final)
|
||
}
|
||
|
||
action := strings.ToLower(strings.TrimSpace(decision.Action))
|
||
if action == "" {
|
||
action = "none"
|
||
}
|
||
|
||
if action == "none" {
|
||
finalText := strings.TrimSpace(decision.Final)
|
||
if finalText == "" {
|
||
finalText = "我已完成思考，但当前没有足够信息给出稳定结论。"
|
||
}
|
||
if o.log != nil {
|
||
o.log.Infof("react final step=%d final=%q", step, finalText)
|
||
}
|
||
return finalText, nil
|
||
}
|
||
|
||
tool, ok := o.tools.Get(action)
|
||
if !ok {
|
||
if o.log != nil {
|
||
o.log.Warnf("react step tool missing step=%d tool=%s", step, action)
|
||
}
|
||
scratchpad += fmt.Sprintf("Step %d Thought: %s\nStep %d Observation: tool %s 不存在\n", step, decision.Thought, step, action)
|
||
continue
|
||
}
|
||
|
||
toolOut, toolErr := tool.Call(ctx, decision.ActionInput)
|
||
if o.log != nil {
|
||
o.log.Infof("react step tool call step=%d tool=%s input=%q", step, action, decision.ActionInput)
|
||
}
|
||
obs := strings.TrimSpace(toolOut)
|
||
if obs == "" {
|
||
obs = "(empty output)"
|
||
}
|
||
if toolErr != nil {
|
||
obs = obs + "\nERROR: " + toolErr.Error()
|
||
}
|
||
if o.log != nil {
|
||
o.log.Infof("react step observation step=%d tool=%s observation=%q", step, action, obs)
|
||
}
|
||
if len(obs) > 2000 {
|
||
obs = obs[:2000]
|
||
}
|
||
scratchpad += fmt.Sprintf("Step %d Thought: %s\nStep %d Action: %s\nStep %d ActionInput: %s\nStep %d Observation: %s\n", step, decision.Thought, step, action, step, decision.ActionInput, step, obs)
|
||
}
|
||
|
||
return "我尝试了多轮思考与工具调用，但仍未得到稳定结论。请给我更具体的约束或允许我继续尝试。", nil
|
||
}
|
||
|
||
func (o *Orchestrator) matchSkills(ctx context.Context, compressedContext, userInput string) []knowledge.Skill {
|
||
if len(o.skills) == 0 {
|
||
return nil
|
||
}
|
||
|
||
type skillChoice struct {
|
||
Skills []string `json:"skills"`
|
||
}
|
||
|
||
systemPrompt := strings.Join([]string{
|
||
"你是技能路由器。",
|
||
"任务：根据用户问题，从候选技能中选择 0-2 个最相关技能名称。",
|
||
"输出必须是 JSON：{\"skills\":[\"name1\",\"name2\"]}",
|
||
"如果没有匹配技能，返回 {\"skills\":[]}。",
|
||
"不要输出 JSON 之外内容。",
|
||
}, "\n")
|
||
|
||
userPrompt := strings.Join([]string{
|
||
"候选技能：",
|
||
formatSkillCatalog(o.skills),
|
||
"",
|
||
"历史上下文：",
|
||
compressedContext,
|
||
"",
|
||
"用户问题：",
|
||
userInput,
|
||
}, "\n")
|
||
|
||
raw, err := o.llm.Generate(ctx, systemPrompt, userPrompt)
|
||
if err != nil {
|
||
if o.log != nil {
|
||
o.log.Warnf("skill match llm failed err=%v", err)
|
||
}
|
||
return nil
|
||
}
|
||
if o.log != nil {
|
||
o.log.Infof("skill router output raw=%q", raw)
|
||
}
|
||
|
||
raw = normalizeJSON(raw)
|
||
choice := skillChoice{}
|
||
if err := json.Unmarshal([]byte(raw), &choice); err != nil {
|
||
if o.log != nil {
|
||
o.log.Warnf("skill match parse failed err=%v", err)
|
||
}
|
||
return nil
|
||
}
|
||
|
||
picked := make([]knowledge.Skill, 0, 2)
|
||
seen := map[string]struct{}{}
|
||
for _, name := range choice.Skills {
|
||
name = strings.TrimSpace(strings.ToLower(name))
|
||
if name == "" {
|
||
continue
|
||
}
|
||
if _, ok := seen[name]; ok {
|
||
continue
|
||
}
|
||
for _, skill := range o.skills {
|
||
if strings.ToLower(strings.TrimSpace(skill.Name)) == name {
|
||
picked = append(picked, skill)
|
||
seen[name] = struct{}{}
|
||
break
|
||
}
|
||
}
|
||
if len(picked) >= 2 {
|
||
break
|
||
}
|
||
}
|
||
if o.log != nil {
|
||
names := make([]string, 0, len(picked))
|
||
for _, s := range picked {
|
||
names = append(names, s.Name)
|
||
}
|
||
o.log.Infof("skill router selected skills=%s", strings.Join(names, ","))
|
||
}
|
||
|
||
return picked
|
||
}
|
||
|
||
func parseDecision(raw string) (reactDecision, error) {
|
||
raw = normalizeJSON(raw)
|
||
start := strings.Index(raw, "{")
|
||
end := strings.LastIndex(raw, "}")
|
||
if start < 0 || end < start {
|
||
return reactDecision{}, fmt.Errorf("no json object found")
|
||
}
|
||
raw = raw[start : end+1]
|
||
|
||
var out reactDecision
|
||
if err := json.Unmarshal([]byte(raw), &out); err != nil {
|
||
return reactDecision{}, err
|
||
}
|
||
return out, nil
|
||
}
|
||
|
||
func normalizeJSON(raw string) string {
|
||
raw = strings.TrimSpace(raw)
|
||
raw = strings.TrimPrefix(raw, "```json")
|
||
raw = strings.TrimPrefix(raw, "```")
|
||
raw = strings.TrimSuffix(raw, "```")
|
||
return strings.TrimSpace(raw)
|
||
}
|
||
|
||
func formatSkills(skills []knowledge.Skill) string {
|
||
b := strings.Builder{}
|
||
for _, skill := range skills {
|
||
b.WriteString("## ")
|
||
b.WriteString(skill.Name)
|
||
b.WriteString("\n")
|
||
b.WriteString(skill.Content)
|
||
b.WriteString("\n\n")
|
||
}
|
||
return strings.TrimSpace(b.String())
|
||
}
|
||
|
||
func formatSkillCatalog(skills []knowledge.Skill) string {
|
||
b := strings.Builder{}
|
||
for _, skill := range skills {
|
||
summary := strings.ReplaceAll(skill.Content, "\n", " ")
|
||
summary = strings.TrimSpace(summary)
|
||
if len(summary) > 220 {
|
||
summary = summary[:220]
|
||
}
|
||
b.WriteString("- ")
|
||
b.WriteString(skill.Name)
|
||
if summary != "" {
|
||
b.WriteString(": ")
|
||
b.WriteString(summary)
|
||
}
|
||
b.WriteString("\n")
|
||
}
|
||
return strings.TrimSpace(b.String())
|
||
}
|
||
|
||
func (o *Orchestrator) formatToolDoc() string {
|
||
list := o.tools.List()
|
||
if len(list) == 0 {
|
||
return "(none)"
|
||
}
|
||
sort.Slice(list, func(i, j int) bool {
|
||
return list[i].Name() < list[j].Name()
|
||
})
|
||
b := strings.Builder{}
|
||
for _, t := range list {
|
||
b.WriteString("- ")
|
||
b.WriteString(t.Name())
|
||
b.WriteString(": ")
|
||
b.WriteString(t.Description())
|
||
b.WriteString("\n")
|
||
}
|
||
return strings.TrimSpace(b.String())
|
||
}
|