secert updated

utils/embedding_test.go

@@ -3,6 +3,7 @@ package utils
 import (
 	"context"
 	"encoding/json"
+	"errors"
 	"net/http"
 	"net/http/httptest"
 	"strings"
@@ -73,16 +74,70 @@ func TestGeminiEmbedderSendsTheRequestTheAPIExpects(t *testing.T) {
 }
 
 func TestEmbedderSurfacesProviderErrors(t *testing.T) {
+	// Used to assert this with a 429, which is now the one status that is
+	// deliberately NOT passed through — see the two tests below. The rule it was
+	// written for still holds for every other failure: a provider that explains
+	// itself should not have that explanation swallowed.
 	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-		w.WriteHeader(429)
-		w.Write([]byte(`{"error":{"message":"Rate limit reached"}}`))
+		w.WriteHeader(400)
+		w.Write([]byte(`{"error":{"message":"input exceeds the maximum token count"}}`))
 	}))
 	defer srv.Close()
 
 	e, _ := NewEmbedder(config.EmbeddingConfig{Provider: "openai", Model: "m", APIKey: "k", BaseURL: srv.URL})
 	_, err := e.Embed(context.Background(), "x")
-	if err == nil || !strings.Contains(err.Error(), "429") || !strings.Contains(err.Error(), "Rate limit") {
-		t.Fatalf("want a 429 with the provider's message, got %v", err)
+	if err == nil || !strings.Contains(err.Error(), "400") || !strings.Contains(err.Error(), "maximum token count") {
+		t.Fatalf("want a 400 with the provider's message, got %v", err)
+	}
+}
+
+func TestARateLimitIsWaitedOutRatherThanReported(t *testing.T) {
+	// The common case on a free tier: the quota clears in under a second, so
+	// the right answer is a slightly slower success rather than an error.
+	var calls int
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		calls++
+		if calls == 1 {
+			w.WriteHeader(429)
+			w.Write([]byte(`{"error":{"message":"Rate limit reached. Please try again in 20ms."}}`))
+			return
+		}
+		w.Write([]byte(`{"data":[{"embedding":[0.5,0.25]}]}`))
+	}))
+	defer srv.Close()
+
+	e, _ := NewEmbedder(config.EmbeddingConfig{Provider: "openai", Model: "m", APIKey: "k", BaseURL: srv.URL})
+	vector, err := e.Embed(context.Background(), "x")
+	if err != nil {
+		t.Fatalf("a 20ms quota wait became an error: %v", err)
+	}
+	if len(vector) != 2 {
+		t.Fatalf("the retry did not return the answer: %v", vector)
+	}
+	if calls != 2 {
+		t.Fatalf("expected one retry, saw %d calls", calls)
+	}
+}
+
+func TestAPersistentRateLimitIsReportedWithoutTheProvidersDetails(t *testing.T) {
+	// When waiting does not help, the person is told to try again — not handed
+	// our organisation id and the tokens-per-minute arithmetic behind it.
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(429)
+		w.Write([]byte(`{"error":{"message":"Rate limit reached for model X in organization org_01m38 on tokens per minute (TPM): Limit 8000. Please try again in 20ms."}}`))
+	}))
+	defer srv.Close()
+
+	e, _ := NewEmbedder(config.EmbeddingConfig{Provider: "openai", Model: "m", APIKey: "k", BaseURL: srv.URL})
+	_, err := e.Embed(context.Background(), "x")
+
+	if !errors.Is(err, ErrBusy) {
+		t.Fatalf("a persistent quota was not reported as busy: %v", err)
+	}
+	for _, leaked := range []string{"org_01m38", "TPM", "8000"} {
+		if strings.Contains(err.Error(), leaked) {
+			t.Fatalf("%q reached the caller: %s", leaked, err.Error())
+		}
 	}
 }