fix(gittar): optimize ai code-review (#6201) (#6212)

* limit_sync_group: fix the issue that s.Add blocked when added-items much than maxSize.

The behaviour should be same as sync.WaitGroup: you can add all items just once at first.

* optimize prompt and review result

* replace function-call by normal chat completion

* remove useless code

Co-authored-by: sfwn <[email protected]>

cmd/gittar/conf/i18n/i18n.yaml

@@ -1,4 +1,16 @@
 zh:
   mr.note.comment.cannot.be.empty: 评论不能为空
+  template.mr.ai.cr.tip.reach.max.limit: 这里仅展示前 %d 个文件的审查结果。对于其他未审查到的文件，请移至 "**变更**" 页里，手动点击 "**AI 审查**" 按钮，或者手动选择代码片段开启 "**AI 对话**"。
+  mr.ai.cr.title: AI 代码审查
+  file: 文件
+  snippet: 代码片段
+  mr.ai.cr.no.suggestion: 你的代码变更看起来不错，暂无审查建议。
+  template.mr.ai.cr.file.content.max.limit: 文件内容超长，本次审查至第 %d 行。对于其他未审查到的内容，请在 "**变更**" 页里手动选择代码片段开启 "**AI 对话**"。
 en:
   mr.note.comment.cannot.be.empty: Comment cannot be empty
+  template.mr.ai.cr.tip.reach.max.limit: Only the first %d files are displayed here. For other files that have not been reviewed, please manually click the "**AI Code Review**" button in the "**Changes**" tab, or manually select the code snippet to start the "**AI Conversation**".
+  mr.ai.cr.title: AI Code Review
+  file: File
+  snippet: Code Snippet
+  mr.ai.cr.no.suggestion: Your code changes look good, no review suggestions for now.
+  template.mr.ai.cr.file.content.max.limit: The file content is too long, only the first %d lines are reviewed. For other content that has not been reviewed, please manually select the code snippet in "Changes" tab to start the "AI Conversation".

internal/tools/gittar/ai/cr/impl/cr_mr/mr.go

@@ -15,16 +15,26 @@
 package cr_mr
 
 import (
+	"fmt"
+	"strings"
 	"sync"
 
+	"github.com/sirupsen/logrus"
+
+	"github.com/erda-project/erda-infra/providers/i18n"
 	"github.com/erda-project/erda/apistructs"
 	"github.com/erda-project/erda/internal/tools/gittar/ai/cr/impl/cr_mr_file"
+	"github.com/erda-project/erda/internal/tools/gittar/ai/cr/util/mdutil"
 	"github.com/erda-project/erda/internal/tools/gittar/ai/cr/util/mrutil"
 	"github.com/erda-project/erda/internal/tools/gittar/models"
 	"github.com/erda-project/erda/internal/tools/gittar/pkg/gitmodule"
 	"github.com/erda-project/erda/pkg/limit_sync_group"
 )
 
+const (
+	MaxDiffFileNum = 5
+)
+
 type mrReviewer struct {
 	req  models.AICodeReviewNoteRequest
 	repo *gitmodule.Repository
@@ -38,26 +48,32 @@ func init() {
 	})
 }
 
-func (r *mrReviewer) CodeReview() string {
+func (r *mrReviewer) CodeReview(i18n i18n.Translator, lang i18n.LanguageCodes) string {
 	diff := mrutil.GetDiffFromMR(r.repo, r.mr)
 
 	// mr has many changed files, we will review only the first ten files one by one. Then, combine the file-level suggestions.
 
 	var changedFiles []models.FileCodeReviewer
+	var reachMaxDiffFileNumLimit bool
 	for _, diffFile := range diff.Files {
-		if len(changedFiles) >= 10 {
+		if len(changedFiles) >= MaxDiffFileNum {
+			reachMaxDiffFileNumLimit = true
 			break
 		}
 		if len(diffFile.Sections) > 0 {
-			fr := cr_mr_file.NewFileReviewer(diffFile, r.user, r.mr)
+			fr, err := cr_mr_file.NewFileReviewer(diffFile.Name, r.repo, r.mr, r.user)
+			if err != nil {
+				logrus.Warnf("failed to create file reviewer for file %s, err: %v", diffFile.Name, err)
+				continue
+			}
 			changedFiles = append(changedFiles, fr)
 		}
 	}
 
 	// parallel do file-level cr
 	var fileOrder []string
 	fileSuggestions := make(map[string]string)
-	wg := limit_sync_group.NewSemaphore(5) // parallel is 5
+	wg := limit_sync_group.NewSemaphore(MaxDiffFileNum) // parallel is 5
 	var mu sync.Mutex
 
 	wg.Add(len(changedFiles))
@@ -66,19 +82,33 @@ func (r *mrReviewer) CodeReview() string {
 		go func(file models.FileCodeReviewer) {
 			defer wg.Done()
 
+			fileSuggestion := file.CodeReview(i18n, lang)
+			if strings.TrimSpace(fileSuggestion) == "" {
+				fileSuggestion = i18n.Text(lang, models.I18nKeyMrAICrNoSuggestion)
+			}
 			mu.Lock()
-			fileSuggestions[file.GetFileName()] = file.CodeReview()
+			fileSuggestions[file.GetFileName()] = fileSuggestion
 			mu.Unlock()
 		}(file)
 	}
 	wg.Wait()
 
 	// combine result
-	var mrReviewResult string
-	mrReviewResult = "# AI Code Review\n"
+	var mrReviewResults []string
+	mrReviewResults = append(mrReviewResults, fmt.Sprintf("# %s", i18n.Text(lang, models.I18nKeyMrAICrTitle)))
+	if reachMaxDiffFileNumLimit {
+		tip := fmt.Sprintf(i18n.Text(lang, models.I18nKeyTemplateMrAICrTipForEachMaxLimit), MaxDiffFileNum)
+		tip = mdutil.MakeRef(mdutil.MakeItalic(tip))
+		mrReviewResults = append(mrReviewResults, tip)
+	}
+	mrReviewResults = append(mrReviewResults, "")
 	for _, fileName := range fileOrder {
-		mrReviewResult += "------\n## File: `" + fileName + "`\n\n" + fileSuggestions[fileName] + "\n"
+		mrReviewResults = append(mrReviewResults, "------")
+		mrReviewResults = append(mrReviewResults, fmt.Sprintf("## %s: `%s`", i18n.Text(lang, models.I18nKeyFile), fileName))
+		mrReviewResults = append(mrReviewResults, "")
+		mrReviewResults = append(mrReviewResults, fileSuggestions[fileName])
+		mrReviewResults = append(mrReviewResults, "")
 	}
 
-	return mrReviewResult
+	return strings.Join(mrReviewResults, "\n")
 }

internal/tools/gittar/ai/cr/impl/cr_mr_code_snippet/prompt.yaml

@@ -1,10 +1,10 @@
 messages:
+  - role: system
+    content: Please reply in {{.UserLang}}.
+
   - role: system
     content: |
       Please give a review suggestions for the selected code, use markdown title for each suggestion. Code examples can be provided when necessary.
       The first-level title is: AI Code Review.
       Code:
       {{.SelectedCode}}
-
-  - role: system
-    content: Please reply in Chinese.

internal/tools/gittar/ai/cr/impl/cr_mr_code_snippet/snippet.go

@@ -26,8 +26,10 @@ import (
 	"github.com/sashabaranov/go-openai"
 	"gopkg.in/yaml.v3"
 
+	"github.com/erda-project/erda-infra/providers/i18n"
 	"github.com/erda-project/erda/apistructs"
 	"github.com/erda-project/erda/internal/tools/gittar/ai/cr/util/aiutil"
+	"github.com/erda-project/erda/internal/tools/gittar/ai/cr/util/i18nutil"
 	"github.com/erda-project/erda/internal/tools/gittar/models"
 	"github.com/erda-project/erda/internal/tools/gittar/pkg/gitmodule"
 )
@@ -41,6 +43,7 @@ type CodeSnippet struct {
 	CodeLanguage string
 	SelectedCode string
 	Truncated    bool // if there are too many changes, we have to truncate the content according to the model context
+	UserLang     string
 
 	user *models.User
 }
@@ -82,15 +85,16 @@ func newSnippetCodeReviewer(codeLang, selectedCode string, truncated bool, user
 	return cs
 }
 
-func (cs CodeSnippet) CodeReview() string {
-	// invoke ai
-	req := cs.constructAIRequest()
+func (cs CodeSnippet) CodeReview(i18n i18n.Translator, lang i18n.LanguageCodes) string {
+	// construct AI request
+	req := cs.constructAIRequest(i18n, lang)
 
 	// invoke
 	return aiutil.InvokeAI(req, cs.user)
 }
 
-func (cs CodeSnippet) constructAIRequest() openai.ChatCompletionRequest {
+func (cs CodeSnippet) constructAIRequest(i18n i18n.Translator, lang i18n.LanguageCodes) openai.ChatCompletionRequest {
+	cs.UserLang = i18nutil.GetUserLang(lang)
 	msgs := deepcopy.Copy(promptStruct.Messages).([]openai.ChatCompletionMessage)
 
 	// invoke ai