Skip to content

Open-source LLM knowledge platform: turn raw documents into a queryable RAG, an autonomous reasoning agent, and a self-maintaining Wiki.

fix(agent): 拒绝将断流摘要保存为记忆检查点 (#3994)

诉说晚秋fa4f85d3dd1e
diff --git a/internal/agent/compaction/compactor.go b/internal/agent/compaction/compactor.go
index 2cc85a354..474213eda 100644
--- a/internal/agent/compaction/compactor.go
+++ b/internal/agent/compaction/compactor.go
@@ -306,6 +306,8 @@ func validateSummary(content, finishReason string) error {
return errors.New("empty response from LLM")
}
switch strings.ToLower(strings.TrimSpace(finishReason)) {
+ case types.FinishReasonIncomplete:
+ return errors.New(types.StreamEndedEarlyError)
case "length", "max_tokens", "max_output_tokens":
return errors.New("generation hit the token cap and the summary is incomplete")
}
diff --git a/internal/agent/compaction/stream_http_test.go b/internal/agent/compaction/stream_http_test.go
new file mode 100644
index 000000000..a0e4acaad
--- /dev/null
+++ b/internal/agent/compaction/stream_http_test.go
@@ -0,0 +1,68 @@
+package compaction
+
+import (
+ "context"
+ "fmt"
+ "net/http"
+ "net/http/httptest"
+ "sync/atomic"
+ "testing"
+
+ "github.com/Tencent/WeKnora/internal/models/api"
+ "github.com/Tencent/WeKnora/internal/models/api/openaicompletions"
+ "github.com/stretchr/testify/assert"
+ "github.com/stretchr/testify/require"
+)
+
+func TestCompactionHTTPStreamCheckpoint(t *testing.T) {
+ for _, tc := range []struct {
+ name string
+ failures int32
+ wantCalls int32
+ degraded bool
+ }{
+ {name: "complete", wantCalls: 2},
+ {name: "retry after EOF", failures: 1, wantCalls: 3},
+ {name: "persistent EOF", failures: 4, wantCalls: 4, degraded: true},
+ } {
+ t.Run(tc.name, func(t *testing.T) {
+ t.Setenv("SSRF_WHITELIST", "127.0.0.1")
+ var calls atomic.Int32
+ server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+ writeSummarySSE(w, calls.Add(1) <= tc.failures)
+ }))
+ defer server.Close()
+ model := openaicompletions.New(openaicompletions.Config{Endpoint: api.Endpoint{
+ BaseURL: server.URL, Model: "summary-test", Client: server.Client(),
+ }})
+ c := New(model, newEstimator(t), testSettings())
+ messages := withStoredHistory(reactTurn(12), storedTurn("stored-turn", 1, "/workspace/a.txt"))
+ result, err := c.Compact(context.Background(), messages, ReasonThreshold)
+ require.NoError(t, err)
+ require.NotNil(t, result.Checkpoint)
+ assert.Equal(t, tc.wantCalls, calls.Load())
+ assert.Equal(t, tc.degraded, result.Degraded)
+ assert.Equal(t, tc.degraded, result.Checkpoint.Degraded)
+ assert.NotContains(t, result.Summary, "INCOMPLETE_SUMMARY")
+ assert.NotContains(t, result.Checkpoint.Summary, "INCOMPLETE_SUMMARY")
+ if tc.degraded {
+ assert.Contains(t, result.Checkpoint.Summary, "Raw conversation archive")
+ assert.Contains(t, result.Checkpoint.Summary, "question stored-turn")
+ } else {
+ assert.Contains(t, result.Checkpoint.Summary, "COMPLETE_SUMMARY")
+ }
+ })
+ }
+}
+
+// The real provider must classify a clean HTTP EOF
+// without finish_reason as incomplete before the compactor decides to commit.
+func writeSummarySSE(w http.ResponseWriter, incomplete bool) {
+ w.Header().Set("Content-Type", "text/event-stream")
+ if incomplete {
+ _, _ = fmt.Fprint(w, `data: {"choices":[{"index":0,"delta":{"content":"INCOMPLETE_SUMMARY"}}]}`+"\n\n")
+ } else {
+ _, _ = fmt.Fprint(w, `data: {"choices":[{"index":0,"delta":{"content":"COMPLETE_SUMMARY"},`+
+ `"finish_reason":"stop"}]}`+"\n\n")
+ }
+}