pjandro/OmniRoute
1
0
Fork 0
mirror of https://github.com/diegosouzapw/OmniRoute.git synced 2026-08-25 08:32:11 +03:00
Code Issues Packages Projects Releases Wiki Activity

fix(models): normalize media endpoint metadata (#11397)

Browse Source 
Validado em lote combinado (batch-0824g) contra o tip de release/v3.8.50: typecheck:core limpo, gates estáticos OK, 62/62 testes focados passando (endpoint/parser/schema/static-model + catálogo).

Canonicaliza metadados de endpoint legados (video/audio) para IDs específicos por operação, mantendo compatibilidade retroativa via `normalizeModelSupportedEndpoints` (valores antigos `audio`/`video` continuam válidos como entrada e são normalizados na escrita). Obrigado pela contribuição, primeira PR bem-vinda!
This commit is contained in:
Marcelo Karval
2026-08-24 17:23:48 -03:00
committed by GitHub
parent f93fecd86b
commit 11cbd7d4e0
11 changed files with 233 additions and 55 deletions
Download Patch File Download Diff File Expand all files Collapse all files

71
src/app/(dashboard)/dashboard/providers/[id]/components/CustomModelsSection.tsx
View File

@@ -11,6 +11,10 @@
import React, { useState, useEffect, useCallback, useMemo } from "react";
import { useTranslations } from "next-intl";
import { Button } from "@/shared/components";
import {
normalizeModelSupportedEndpoints,
type ModelSupportedEndpoint,
} from "@/shared/constants/modelSupportedEndpoints";
import { useNotificationStore } from "@/store/notificationStore";
import {
buildCompatMap,
@@ -51,6 +55,29 @@ function targetFormatLabel(value: string, t: (key: string) => string): string {
return key ? t(key) : value;
}
const MODEL_ENDPOINT_OPTIONS: ModelSupportedEndpoint[] = [
"chat",
"embeddings",
"rerank",
"images",
"videos",
"audio-speech",
"audio-transcriptions",
];
function endpointLabel(endpoint: ModelSupportedEndpoint, t: (key: string) => string): string {
const labels: Partial<Record<ModelSupportedEndpoint, string>> = {
chat: `💬 ${t("supportedEndpointChat")}`,
embeddings: `📐 ${t("supportedEndpointEmbeddings")}`,
rerank: providerText(t, "rerankEndpoint", "Rerank"),
images: `🖼️ ${t("supportedEndpointImages")}`,
videos: "🎬 Video",
"audio-speech": `🔊 ${t("audioSpeech")}`,
"audio-transcriptions": `🎙️ ${t("audioTranscriptions")}`,
};
return labels[endpoint] || endpoint;
}
/**
* #4125: parse the free-text "Context Window Override" field. Blank → no override
* (`value: null`, not an error). A non-empty value must be a positive whole number of
@@ -232,7 +259,7 @@ export default function CustomModelsSection({
setEditingApiFormat(model.apiFormat || "chat-completions");
setEditingEndpoints(
Array.isArray(model.supportedEndpoints) && model.supportedEndpoints.length
? model.supportedEndpoints
? normalizeModelSupportedEndpoints(model.supportedEndpoints)
: ["chat"]
);
setEditingTargetFormat(model.targetFormat || "");
@@ -428,6 +455,7 @@ export default function CustomModelsSection({
<option value="audio-transcriptions">{t("audioTranscriptions")}</option>
<option value="audio-speech">{t("audioSpeech")}</option>
<option value="images-generations">{t("imagesGenerations")}</option>
<option value="video">Video</option>
</select>
</div>
<div className="w-48">
@@ -454,7 +482,7 @@ export default function CustomModelsSection({
{t("supportedEndpointsLabel")}
</span>
<div className="flex items-center gap-3">
{["chat", "embeddings", "rerank", "images", "audio"].map((ep) => (
{MODEL_ENDPOINT_OPTIONS.map((ep) => (
<label
key={ep}
className="flex items-center gap-1.5 text-xs text-text-main cursor-pointer"
@@ -471,15 +499,7 @@ export default function CustomModelsSection({
}}
className="rounded border-border"
/>
{ep === "chat"
? `💬 ${t("supportedEndpointChat")}`
: ep === "embeddings"
? `📐 ${t("supportedEndpointEmbeddings")}`
: ep === "rerank"
? providerText(t, "rerankEndpoint", "Rerank")
: ep === "images"
? `🖼️ ${t("supportedEndpointImages")}`
: `🔊 ${t("supportedEndpointAudio")}`}
{endpointLabel(ep, t)}
</label>
))}
</div>
@@ -594,6 +614,22 @@ export default function CustomModelsSection({
{`🔊 ${t("audioShortLabel")}`}
</span>
)}
{(model.supportedEndpoints?.includes("videos") ||
model.supportedEndpoints?.includes("video")) && (
<span className="text-[10px] px-1.5 py-0.5 rounded-full bg-red-500/15 text-red-400 font-medium">
🎬 Video
</span>
)}
{model.supportedEndpoints?.includes("audio-speech") && (
<span className="text-[10px] px-1.5 py-0.5 rounded-full bg-green-500/15 text-green-400 font-medium">
{`🔊 ${t("audioSpeech")}`}
</span>
)}
{model.supportedEndpoints?.includes("audio-transcriptions") && (
<span className="text-[10px] px-1.5 py-0.5 rounded-full bg-cyan-500/15 text-cyan-400 font-medium">
{`🎙️ ${t("audioTranscriptions")}`}
</span>
)}
{anyNormalizeCompatBadge(model.id!, customMap, overrideMap) && (
<span
className="text-[10px] px-1.5 py-0.5 rounded-full bg-slate-500/15 text-slate-400 font-medium"
@@ -639,6 +675,7 @@ export default function CustomModelsSection({
<option value="audio-transcriptions">{t("audioTranscriptions")}</option>
<option value="audio-speech">{t("audioSpeech")}</option>
<option value="images-generations">{t("imagesGenerations")}</option>
<option value="video">Video</option>
</select>
</div>
<div className="w-[11rem] shrink-0 min-w-0">
@@ -697,7 +734,7 @@ export default function CustomModelsSection({
{t("supportedEndpointsLabel")}
</span>
<div className="flex flex-wrap items-center gap-x-2 sm:gap-x-3 gap-y-1 min-w-0">
{["chat", "embeddings", "rerank", "images", "audio"].map((ep) => (
{MODEL_ENDPOINT_OPTIONS.map((ep) => (
<label
key={ep}
className="flex items-center gap-1.5 text-xs text-text-main cursor-pointer whitespace-nowrap"
@@ -716,15 +753,7 @@ export default function CustomModelsSection({
}}
className="rounded border-border"
/>
{ep === "chat"
? `💬 ${t("supportedEndpointChat")}`
: ep === "embeddings"
? `📐 ${t("supportedEndpointEmbeddings")}`
: ep === "rerank"
? providerText(t, "rerankEndpoint", "Rerank")
: ep === "images"
? `🖼️ ${t("supportedEndpointImages")}`
: `🔊 ${t("supportedEndpointAudio")}`}
{endpointLabel(ep, t)}
</label>
))}
</div>

18
src/app/api/v1/models/catalog.ts
View File

@@ -69,6 +69,7 @@ import {
import { createModelCapabilityResolutionSnapshot } from "@/lib/modelCapabilityResolutionSnapshot";
import { getModelsDevPricing, getSyncedCapability } from "@/lib/modelsDevSync";
import { getModelSpec } from "@/shared/constants/modelSpecs";
import { classifyModelSupportedEndpoints } from "@/shared/constants/modelSupportedEndpoints";
import { getModelsCatalogPrefixMode } from "@/shared/utils/featureFlags";
import { buildReservedPrefixes, selectCompatibleNodeForPrefix } from "@/lib/providerNodePrefixes";
import { applyCatalogPostFilters, finalizeCatalogResponse } from "./catalogResponse";
@@ -1152,18 +1153,15 @@ async function buildUnifiedModelsResponseCore(
const aliasId = `${alias}/${displayModelId}`;
const endpoints = Array.isArray(sm.supportedEndpoints) ? sm.supportedEndpoints : ["chat"];
const apiFormat = typeof sm.apiFormat === "string" ? sm.apiFormat : "chat-completions";
let modelType: string | undefined;
if (endpoints.includes("embeddings")) modelType = "embedding";
else if (endpoints.includes("rerank")) modelType = "rerank";
else if (endpoints.includes("images")) modelType = "image";
else if (endpoints.includes("audio")) modelType = "audio";
const classification = classifyModelSupportedEndpoints(endpoints);
const modelType = classification.type;
// Same owned_by the alias/canonical entries below will carry — computed once
// so the effort_tiers exclusion (codex/glm/kimi) and the entries agree.
const syncedOwnedBy = resolvePublicOwnerId(providerId, canonicalProviderId);
const syncedFields = {
...(modelType ? { type: modelType } : {}),
...(apiFormat !== "chat-completions" ? { api_format: apiFormat } : {}),
...(modelType === "audio" ? { subtype: "transcription" } : {}),
...(classification.subtype ? { subtype: classification.subtype } : {}),
...(sm.inputTokenLimit ? { context_length: sm.inputTokenLimit } : {}),
...(typeof sm.outputTokenLimit === "number"
? { max_output_tokens: sm.outputTokenLimit }
@@ -1604,11 +1602,8 @@ async function buildUnifiedModelsResponseCore(
: ["chat"];
const apiFormat =
typeof model.apiFormat === "string" ? model.apiFormat : "chat-completions";
let modelType: string | undefined;
if (endpoints.includes("embeddings")) modelType = "embedding";
else if (endpoints.includes("rerank")) modelType = "rerank";
else if (endpoints.includes("images")) modelType = "image";
else if (endpoints.includes("audio")) modelType = "audio";
const classification = classifyModelSupportedEndpoints(endpoints);
const modelType = classification.type;
if (
modelType &&
hasEquivalentSpecialtyModel(canonicalProviderId, modelId, modelType, aliasId)
@@ -1631,6 +1626,7 @@ async function buildUnifiedModelsResponseCore(
parent: null,
custom: true,
...(modelType ? { type: modelType } : {}),
...(classification.subtype ? { subtype: classification.subtype } : {}),
...(apiFormat !== "chat-completions" ? { api_format: apiFormat } : {}),
...(endpoints.length > 1 || !endpoints.includes("chat")
? { supported_endpoints: endpoints }

3
src/lib/db/models.ts
View File

@@ -190,7 +190,8 @@ export async function addCustomModel(
| "rerank"
| "audio-transcriptions"
| "audio-speech"
| "images-generations" = "chat-completions",
| "images-generations"
| "video" = "chat-completions",
supportedEndpoints: string[] = ["chat"],
// #2905: optional per-model wire format override (e.g. "claude" for an
// opencode-go custom model). When unset, routing falls back to the provider

20
src/lib/providerModels/geminiModelsParser.ts
View File

@@ -3,11 +3,15 @@
*
* Each model's `supportedGenerationMethods` is mapped to OmniRoute endpoints:
* - generateContent / generateAnswer → "chat"
* - predictLongRunning → "video" (Veo video generation)
* - predict → "images" (Imagen image generation)
* - predictLongRunning → "videos" (Veo video generation)
* - embedContent → "embeddings"
* - bidiGenerateContent → "audio" (Live real-time audio)
*
* Model-id heuristics ensure Veo models remain in the video bucket.
* Model-id heuristics refine the long-running bucket because Google exposes both
* Imagen and Veo via long-running methods on the same endpoint:
* - id contains "veo" → ensure "videos"
* - id contains "imagen" → force "images" (never "videos")
*
* Note: `gemini-*-image` models (e.g. gemini-3-pro-image) generate images via the
* regular `generateContent` path, so they stay "chat" (image output is a chat
@@ -21,7 +25,8 @@
const METHOD_TO_ENDPOINT: Record<string, string> = {
generateContent: "chat",
embedContent: "embeddings",
predictLongRunning: "video",
predict: "images",
predictLongRunning: "videos",
bidiGenerateContent: "audio",
generateAnswer: "chat",
};
@@ -63,9 +68,14 @@ export function parseGeminiModelsList(data: any): GeminiDiscoveryModel[] {
const id = ((m.name as string) || (m.id as string) || "").replace(/^models\//, "");
const lowerId = id.toLowerCase();
// Keep Veo models in the video bucket even when the method list is incomplete.
// Google exposes Imagen (image) and Veo (video) via long-running methods; the
// method alone can't always distinguish them, so refine by model id.
if (lowerId.includes("veo")) {
endpoints.add("video");
endpoints.add("videos");
}
if (lowerId.includes("imagen")) {
endpoints.delete("videos");
endpoints.add("images");
}
if (endpoints.size === 0) endpoints.add("chat");

4
src/lib/providers/staticModels.ts
View File

@@ -221,7 +221,7 @@ export function getStaticModelsForProvider(provider: string): LocalCatalogModel[
if (speechProvider) {
appendModels(speechProvider.models, {
apiFormat: "audio",
supportedEndpoints: ["audio"],
supportedEndpoints: ["audio-speech"],
});
}
@@ -229,7 +229,7 @@ export function getStaticModelsForProvider(provider: string): LocalCatalogModel[
if (transcriptionProvider) {
appendModels(transcriptionProvider.models, {
apiFormat: "audio",
supportedEndpoints: ["audio"],
supportedEndpoints: ["audio-transcriptions"],
});
}

54
src/shared/constants/modelSupportedEndpoints.ts Normal file
View File

@@ -0,0 +1,54 @@
export const MODEL_SUPPORTED_ENDPOINT_VALUES = [
"chat",
"embeddings",
"rerank",
"images",
"videos",
"audio-speech",
"audio-transcriptions",
"images-generations",
// Persisted legacy values remain valid input and normalize on write/edit.
"video",
"audio",
] as const;
export type ModelSupportedEndpoint = (typeof MODEL_SUPPORTED_ENDPOINT_VALUES)[number];
export function normalizeModelSupportedEndpoints(endpoints: readonly string[]): string[] {
const normalized: string[] = [];
const add = (endpoint: string) => {
if (!normalized.includes(endpoint)) normalized.push(endpoint);
};
for (const endpoint of endpoints) {
if (endpoint === "video") {
add("videos");
} else if (endpoint === "audio") {
add("audio-speech");
add("audio-transcriptions");
} else {
add(endpoint);
}
}
return normalized;
}
export function classifyModelSupportedEndpoints(endpoints: readonly string[]): {
type?: "embedding" | "rerank" | "image" | "video" | "audio";
subtype?: "speech" | "transcription";
} {
if (endpoints.includes("embeddings")) return { type: "embedding" };
if (endpoints.includes("rerank")) return { type: "rerank" };
if (endpoints.includes("images")) return { type: "image" };
if (endpoints.includes("videos") || endpoints.includes("video")) return { type: "video" };
const supportsSpeech = endpoints.includes("audio-speech");
const supportsTranscription =
endpoints.includes("audio-transcriptions") || endpoints.includes("audio");
if (!supportsSpeech && !supportsTranscription) return {};
if (supportsSpeech && !supportsTranscription) return { type: "audio", subtype: "speech" };
if (supportsTranscription && !supportsSpeech) {
return { type: "audio", subtype: "transcription" };
}
return { type: "audio" };
}

20
src/shared/validation/schemas/provider.ts
View File

@@ -6,6 +6,10 @@ import {
import { SUPPORTED_BATCH_ENDPOINTS } from "@/shared/constants/batchEndpoints";
import { MAX_REQUEST_BODY_LIMIT_MB, MIN_REQUEST_BODY_LIMIT_MB } from "@/shared/constants/bodySize";
import { COMBO_CONFIG_MODES } from "@/shared/constants/comboConfigMode";
import {
MODEL_SUPPORTED_ENDPOINT_VALUES,
normalizeModelSupportedEndpoints,
} from "@/shared/constants/modelSupportedEndpoints";
import { providerAllowsOptionalApiKey } from "@/shared/constants/providers";
import { HIDEABLE_SIDEBAR_ITEM_IDS } from "@/shared/constants/sidebarVisibility";
import {
@@ -238,22 +242,12 @@ export const providerModelMutationSchema = z.object({
"audio-transcriptions",
"audio-speech",
"images-generations",
"video",
])
.default("chat-completions"),
supportedEndpoints: z
.array(
z.enum([
"chat",
"embeddings",
"rerank",
"images",
"audio",
"audio-transcriptions",
"audio-speech",
"images-generations",
"videos",
])
)
.array(z.enum(MODEL_SUPPORTED_ENDPOINT_VALUES))
.transform(normalizeModelSupportedEndpoints)
.default(["chat"]),
// #2905: optional per-model wire format override for custom models (e.g. a
// custom opencode-go model that must use the Anthropic Messages shape).

4
tests/unit/gemini-models-parser.test.ts
View File

@@ -82,11 +82,11 @@ test("parseGeminiModelsList maps embedContent and bidiGenerateContent", () => {
]);
});
test("parseGeminiModelsList maps Veo predictLongRunning models to the video endpoint", () => {
test("parseGeminiModelsList maps Veo predictLongRunning models to the videos endpoint", () => {
const models = parseGeminiModelsList(SAMPLE);
const veo = models.find((m) => m.id === "veo-3.0-generate-001");
assert.ok(veo, "veo-3.0-generate-001 should be present");
assert.deepEqual(veo!.supportedEndpoints, ["video"]);
assert.deepEqual(veo!.supportedEndpoints, ["videos"]);
});
test("parseGeminiModelsList defaults to chat and tolerates empty/missing input", () => {

58
tests/unit/model-supported-endpoints.test.ts Normal file
View File

@@ -0,0 +1,58 @@
import assert from "node:assert/strict";
import test from "node:test";
import {
classifyModelSupportedEndpoints,
MODEL_SUPPORTED_ENDPOINT_VALUES,
normalizeModelSupportedEndpoints,
} from "../../src/shared/constants/modelSupportedEndpoints.ts";
test("normalizes legacy video and audio metadata to operation-specific endpoint ids", () => {
assert.deepEqual(normalizeModelSupportedEndpoints(["chat", "video", "audio"]), [
"chat",
"videos",
"audio-speech",
"audio-transcriptions",
]);
});
test("deduplicates canonical endpoint ids while preserving order", () => {
assert.deepEqual(
normalizeModelSupportedEndpoints([
"videos",
"video",
"audio-speech",
"audio",
"audio-transcriptions",
]),
["videos", "audio-speech", "audio-transcriptions"]
);
});
test("exports operation-specific values accepted by model metadata", () => {
assert.ok(MODEL_SUPPORTED_ENDPOINT_VALUES.includes("videos"));
assert.ok(MODEL_SUPPORTED_ENDPOINT_VALUES.includes("audio-speech"));
assert.ok(MODEL_SUPPORTED_ENDPOINT_VALUES.includes("audio-transcriptions"));
});
test("preserves endpoint ids introduced by external discovery", () => {
assert.deepEqual(normalizeModelSupportedEndpoints(["responses", "video"]), [
"responses",
"videos",
]);
});
test("classifies operation-specific media endpoints for the model catalog", () => {
assert.deepEqual(classifyModelSupportedEndpoints(["videos"]), { type: "video" });
assert.deepEqual(classifyModelSupportedEndpoints(["audio-speech"]), {
type: "audio",
subtype: "speech",
});
assert.deepEqual(classifyModelSupportedEndpoints(["audio-transcriptions"]), {
type: "audio",
subtype: "transcription",
});
assert.deepEqual(classifyModelSupportedEndpoints(["audio-speech", "audio-transcriptions"]), {
type: "audio",
});
});

16
tests/unit/provider-model-endpoint-schema.test.ts Normal file
View File

@@ -0,0 +1,16 @@
import assert from "node:assert/strict";
import test from "node:test";
import { providerModelMutationSchema } from "../../src/shared/validation/schemas/provider.ts";
test("provider model mutations accept video and persist canonical operation endpoints", () => {
const parsed = providerModelMutationSchema.parse({
provider: "example",
modelId: "media-model",
apiFormat: "video",
supportedEndpoints: ["video", "audio"],
});
assert.equal(parsed.apiFormat, "video");
assert.deepEqual(parsed.supportedEndpoints, ["videos", "audio-speech", "audio-transcriptions"]);
});

20
tests/unit/static-model-operation-endpoints.test.ts Normal file
View File

@@ -0,0 +1,20 @@
import assert from "node:assert/strict";
import test from "node:test";
import { getStaticModelsForProvider } from "../../src/lib/providers/staticModels.ts";
test("speech-only static models advertise the speech operation", () => {
const models = getStaticModelsForProvider("elevenlabs") || [];
assert.ok(models.length > 0);
assert.ok(models.every((model) => model.supportedEndpoints?.includes("audio-speech")));
assert.ok(models.every((model) => !model.supportedEndpoints?.includes("audio")));
});
test("transcription-only static models advertise the transcription operation", () => {
const models = getStaticModelsForProvider("gladia") || [];
assert.ok(models.length > 0);
assert.ok(models.every((model) => model.supportedEndpoints?.includes("audio-transcriptions")));
assert.ok(models.every((model) => !model.supportedEndpoints?.includes("audio")));
});