fix(openai): reject an unknown audio response_format as a 400 before the backend runs

Transcription and diarization checked response_format only after the
backend had run, and returned a plain error for an unknown value.
Failover counts a plain error as a target failure, so one request with
a bad response_format ran the backend on every target of a chain and
tripped all of them.

Check the format before the backend runs and answer 400.

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Assisted-by: Claude:claude-opus-5-5 [Claude Code]

core/http/endpoints/openai/audio_upload_test.go

@@ -50,6 +50,22 @@ var _ = Describe("audio upload endpoints reject bad uploads as client errors", f
 		Expect(he.Code).To(Equal(http.StatusBadRequest))
 	}
 
+	// An unknown response_format is the caller's fault too, and must be
+	// rejected before the backend runs: a failover chain would otherwise
+	// transcribe on every target and count the error against each of them.
+	for name, ec := range map[string]endpointCase{
+		"transcription": cases["transcription"],
+		"diarization":   cases["diarization"],
+	} {
+		It(name+": unknown response_format", func() {
+			body := "--xyz\r\nContent-Disposition: form-data; name=\"response_format\"\r\n\r\nbogus\r\n" +
+				"--xyz\r\nContent-Disposition: form-data; name=\"file\"; filename=\"a.wav\"\r\n\r\nRIFF\r\n--xyz--\r\n"
+			var err error
+			Expect(func() { err = run(ec, "multipart/form-data; boundary=xyz", body) }).NotTo(Panic(), "the backend must not be reached")
+			expectBadRequest(err)
+		})
+	}
+
 	for name, ec := range cases {
 		It(name+": multipart content type without a boundary", func() {
 			expectBadRequest(run(ec, "multipart/form-data", ""))

core/http/endpoints/openai/diarization.go

@@ -1,7 +1,6 @@
 package openai
 
 import (
-	"errors"
 	"fmt"
 	"io"
 	"net/http"
@@ -75,6 +74,13 @@ func DiarizationEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, ap
 		if responseFormat == "" {
 			responseFormat = schema.DiarizationResponseFormatJson
 		}
+		switch responseFormat {
+		case schema.DiarizationResponseFormatJson, schema.DiarizationResponseFormatJsonVerbose, schema.DiarizationResponseFormatRTTM:
+		default:
+			// Checked before the backend runs, for the same reason as in
+			// TranscriptEndpoint.
+			return echo.NewHTTPError(http.StatusBadRequest, "invalid response_format (expected: json, verbose_json, rttm)")
+		}
 
 		file, err := uploadedFile(c, "file")
 		if err != nil {
@@ -126,7 +132,7 @@ func DiarizationEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, ap
 		case schema.DiarizationResponseFormatJsonVerbose:
 			return c.JSON(http.StatusOK, result)
 		default:
-			return errors.New("invalid response_format (expected: json, verbose_json, rttm)")
+			return echo.NewHTTPError(http.StatusBadRequest, "invalid response_format (expected: json, verbose_json, rttm)")
 		}
 	}
 }

core/http/endpoints/openai/transcription.go

@@ -2,7 +2,6 @@ package openai
 
 import (
 	"encoding/json"
-	"errors"
 	"fmt"
 	"io"
 	"mime/multipart"
@@ -117,6 +116,13 @@ func TranscriptEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, app
 			}
 		}
 
+		// Reject an unknown format before the backend runs: the backend work
+		// would be wasted, and a failover chain would count the error
+		// against every target.
+		if !stream && !validTranscriptionResponseFormat(responseFormat) {
+			return echo.NewHTTPError(http.StatusBadRequest, "invalid response_format")
+		}
+
 		// retrieve the file data from the request
 		file, err := uploadedFile(c, "file")
 		if err != nil {
@@ -230,11 +236,21 @@ func TranscriptEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, app
 			}
 			return c.JSON(http.StatusOK, trs)
 		default:
-			return errors.New("invalid response_format")
+			return echo.NewHTTPError(http.StatusBadRequest, "invalid response_format")
 		}
 	}
 }
 
+func validTranscriptionResponseFormat(f schema.TranscriptionResponseFormatType) bool {
+	switch f {
+	case "", schema.TranscriptionResponseFormatLrc, schema.TranscriptionResponseFormatText,
+		schema.TranscriptionResponseFormatSrt, schema.TranscriptionResponseFormatVtt,
+		schema.TranscriptionResponseFormatJson, schema.TranscriptionResponseFormatJsonVerbose:
+		return true
+	}
+	return false
+}
+
 // streamTranscription emits OpenAI-format SSE events for a transcription
 // request: one `transcript.text.delta` per backend chunk, a final
 // `transcript.text.done` with the assembled text, and `[DONE]`. Backends that