Embeddings: fail fast when embedding empty string (#53787)

The OpenAI API fails when given an input array with any empty strings:

```
❯ curl https://api.openai.com/v1/embeddings \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer xxxxx" \
  -d '{
    "input": ["a", ""],
    "model": "text-embedding-ada-002"
  }'
{
  "error": {
    "message": "'$.input' is invalid. Please check the API reference: https://platform.openai.com/docs/api-reference.",
    "type": "invalid_request_error",
    "param": null,
    "code": null
  }
}
```

When using Cody Gateway, we have three layers of retries: 20 retries
from ExternalDoer in Cody Gateway, 20 retries from ExternalDoer in
`embeddings`, and 3 retries from `GetEmbeddingsWithRetry` for a total of
_1200 retries_. Instead of doing that, this just lets us fail fast. We
should separately fix the massive number of retries.

enterprise/cmd/cody-gateway/internal/httpapi/embeddings/openai.go

@@ -27,6 +27,14 @@ type openaiClient struct {
 const apiURL = "https://api.openai.com/v1/embeddings"
 
 func (c *openaiClient) GenerateEmbeddings(ctx context.Context, input codygateway.EmbeddingsRequest) (*codygateway.EmbeddingsResponse, int, error) {
+	for _, s := range input.Input {
+		if s == "" {
+			// The OpenAI API will return an error if any of the strings in texts is an empty string,
+			// so fail fast to avoid making tons of retryable requests.
+			return nil, 0, response.NewHTTPStatusCodeError(http.StatusBadRequest, errors.New("cannot generate embeddings for an empty string"))
+		}
+	}
+
 	openAIModel, ok := openAIModelMappings[input.Model]
 	if !ok {
 		return nil, 0, response.NewHTTPStatusCodeError(http.StatusBadRequest, errors.Newf("no OpenAI model found for %q", input.Model))

enterprise/cmd/cody-gateway/internal/httpapi/embeddings/openai_test.go

@@ -0,0 +1,19 @@
+package embeddings
+
+import (
+	"context"
+	"testing"
+
+	"github.com/sourcegraph/sourcegraph/internal/codygateway"
+	"github.com/stretchr/testify/require"
+)
+
+func TestOpenAI(t *testing.T) {
+	t.Run("errors on empty embedding string", func(t *testing.T) {
+		client := NewOpenAIClient("")
+		_, _, err := client.GenerateEmbeddings(context.Background(), codygateway.EmbeddingsRequest{
+			Input: []string{"a", ""}, // empty string is invalid
+		})
+		require.ErrorContains(t, err, "empty string")
+	})
+}

enterprise/internal/embeddings/embed/client/openai/client.go

@@ -48,6 +48,14 @@ func (c *openaiEmbeddingsClient) GetModelIdentifier() string {
 // In case of failure, it retries the embedding procedure up to maxRetries. This due to the OpenAI API which
 // often hangs up when downloading large embedding responses.
 func (c *openaiEmbeddingsClient) GetEmbeddingsWithRetries(ctx context.Context, texts []string, maxRetries int) ([]float32, error) {
+	for _, text := range texts {
+		if text == "" {
+			// The OpenAI API will return an error if any of the strings in texts is an empty string,
+			// so fail fast to avoid making tons of retryable requests.
+			return nil, errors.New("cannot generate embeddings for an empty string")
+		}
+	}
+
 	embeddings, err := c.getEmbeddings(ctx, texts)
 	if err == nil {
 		return embeddings, nil

enterprise/internal/embeddings/embed/client/openai/client_test.go

@@ -0,0 +1,18 @@
+package openai
+
+import (
+	"context"
+	"testing"
+
+	"github.com/sourcegraph/sourcegraph/internal/conf/conftypes"
+	"github.com/stretchr/testify/require"
+)
+
+func TestOpenAI(t *testing.T) {
+	t.Run("errors on empty embedding string", func(t *testing.T) {
+		client := NewClient(&conftypes.EmbeddingsConfig{})
+		invalidTexts := []string{"a", ""} // empty string is invalid
+		_, err := client.GetEmbeddingsWithRetries(context.Background(), invalidTexts, 10)
+		require.ErrorContains(t, err, "empty string")
+	})
+}