fix(llm): only query llm with necessary lint result

Author: CarlJi

deploy/reviewbot.yaml

@@ -34,7 +34,7 @@ spec:
             - -kube-config=/root/.kube/config
             - -llm.provider=ollama
             - -llm.server-url=$(LLM_SERVER_URL)
-            - -llm.model=qwen2.5-coder:7
+            - -llm.model=qwen2.5-coder:7b
           env:
             - name: GITHUB_WEBHOOK_SECRET
               valueFrom:

internal/lint/agent.go

@@ -124,43 +124,56 @@ func (a *Agent) processOutput(ctx context.Context, output LinterOutput, ref conf
 	return newOutput, true
 }
 
-// ApplyTypedMessageByIssueReferences applies the issue references to the lint results with the typed message.
-func (a *Agent) ApplyTypedMessageByIssueReferences(ctx context.Context, lintResults map[string][]LinterOutput) map[string][]LinterOutput {
-	log := util.FromContext(ctx)
+// EnrichWithIssueReferences applies the issue references to the lint results with the typed message.
+func (a *Agent) EnrichWithIssueReferences(ctx context.Context, lintResults map[string][]LinterOutput) map[string][]LinterOutput {
 	msgFormat := getMsgFormat(a.LinterConfig.ReportType)
-	newLintResults := make(map[string][]LinterOutput, len(lintResults))
+	enriched := make(map[string][]LinterOutput, len(lintResults))
 
 	// Process each file's outputs
 	for file, outputs := range lintResults {
-		newOutputs := make([]LinterOutput, 0, len(outputs))
+		newOutputs := append([]LinterOutput(nil), outputs...)
 
 		// Apply each reference pattern
-		for _, output := range outputs {
-			processed := false
+		for i, output := range outputs {
 			for _, ref := range a.IssueReferences {
 				if newOutput, ok := a.processOutput(ctx, output, ref, msgFormat); ok {
-					newOutputs = append(newOutputs, newOutput)
-					processed = true
+					newOutputs[i] = newOutput
 					break
 				}
 			}
-			if !processed {
+		}
+
+		if len(newOutputs) > 0 {
+			enriched[file] = newOutputs
+		}
+	}
+	return enriched
+}
+
+// EnrichWithLLM enriches the lint results with LLM generated content.
+func (a *Agent) EnrichWithLLM(ctx context.Context, lintResults map[string][]LinterOutput) map[string][]LinterOutput {
+	log := util.FromContext(ctx)
+	enriched := make(map[string][]LinterOutput, len(lintResults))
+
+	for file, outputs := range lintResults {
+		newOutputs := append([]LinterOutput(nil), outputs...)
+
+		for i, output := range outputs {
+			// only enrich the output that is not typed
+			if output.TypedMessage == "" {
 				resp, err := llm.QueryForReference(ctx, a.ModelClient, output.Message)
 				if err != nil {
 					log.Errorf("failed to query LLM server: %v", err)
 				} else {
-					output.TypedMessage = output.Message + fmt.Sprintf(ReferenceFooter, resp)
+					newOutputs[i].TypedMessage = newOutputs[i].Message + fmt.Sprintf(ReferenceFooter, resp)
 				}
-				newOutputs = append(newOutputs, output)
 			}
 		}
-
 		if len(newOutputs) > 0 {
-			newLintResults[file] = newOutputs
+			enriched[file] = newOutputs
 		}
 	}
-
-	return newLintResults
+	return enriched
 }
 
 const ReferenceFooter = `

internal/lint/agent_test.go

@@ -200,7 +200,7 @@ func TestApplyTypedMessageByIssueReferences(t *testing.T) {
 			ctx := context.Background()
 
 			// Call the method
-			result := agent.ApplyTypedMessageByIssueReferences(ctx, tc.lintResults)
+			result := agent.EnrichWithIssueReferences(ctx, tc.lintResults)
 
 			// Compare results
 			if !reflect.DeepEqual(result, tc.expectedOutput) {

internal/lint/lint.go

@@ -194,8 +194,7 @@ func Report(ctx context.Context, a Agent, lintResults map[string][]LinterOutput)
 
 	log.Infof("[%s] found %d files with valid %d linter errors related to this PR %d (%s) \n", linterName, len(lintResults), countLinterErrors(lintResults), num, orgRepo)
 
-	lintResults = a.ApplyTypedMessageByIssueReferences(ctx, lintResults)
-
+	lintResults = a.EnrichWithIssueReferences(ctx, lintResults)
 	if len(lintResults) > 0 {
 		metric.IncIssueCounter(orgRepo, linterName, a.Provider.GetCodeReviewInfo().URL, a.Provider.GetCodeReviewInfo().HeadSHA, float64(countLinterErrors(lintResults)))
 	}

internal/lint/providergithub.go

@@ -343,6 +343,7 @@ func (g *GithubProvider) Report(ctx context.Context, a Agent, lintResults map[st
 
 		metric.NotifyWebhookByText(ConstructGotchaMsg(linterName, a.Provider.GetCodeReviewInfo().URL, ch.GetHTMLURL(), lintResults))
 	case config.GitHubPRReview:
+		lintResults = a.EnrichWithLLM(ctx, lintResults)
 		comments, err := g.ProcessComments(ctx, a, lintResults)
 		if err != nil {
 			log.Errorf("failed to process need to add comments: %v", err)
@@ -374,6 +375,7 @@ func (g *GithubProvider) Report(ctx context.Context, a Agent, lintResults map[st
 
 		// report top 10 lint results to pull request review comments at most
 		top10LintResults := listTop10LintResults(lintResults)
+		top10LintResults = a.EnrichWithLLM(ctx, top10LintResults)
 		comments, err := g.ProcessComments(ctx, a, top10LintResults)
 		if err != nil {
 			log.Errorf("failed to process need to add comments: %v", err)

internal/llm/llm.go

@@ -89,7 +89,7 @@ func QueryForReference(ctx context.Context, model llms.Model, linterOutput strin
 		}
 		return "", err
 	}
-	log.Infof("length of LLM respText: %d", len(respText))
+	log.Infof("promote linter output:%s, response:%s", linterOutput, respText)
 
 	return respText, nil
 }

server.go

@@ -429,6 +429,7 @@ func (s *Server) handleCodeRequestEvent(ctx context.Context, info *codeRequestIn
 }
 
 func (s *Server) withCancel(ctx context.Context, info *codeRequestInfo, fn func(context.Context) error) error {
+	log := util.FromContext(ctx)
 	prID := fmt.Sprintf("%s-%s-%s-%d", info.platform, info.org, info.repo, info.num)
 	if cancel, exists := prMap[prID]; exists {
 		log.Infof("Cancelling processing for Pull Request : %s\n", prID)