feat: add Vercel AI Gateway provider with pricing support

[naaa760]

• adds Vercel AI Gateway as a provider to support chat, transcription, and speech models.

Changes:

• Added VercelAIGatewayProvider class
• Integrated into provider factory with automatic detection for provider-prefixed models
• Added transcription and speech endpoint support
• Uses Vercel AI Gateway pricing API for accurate cost calculation
• Added SDK support (TypeScript & React)

fix: #573

[vercel]

@naaa760 is attempting to deploy a commit to the Merit Systems Team on Vercel.

A member of the Team first needs to authorize it.

[vercel]

The handleChatCompletionResponse method will throw an UnknownModelError when processing responses for provider-prefixed models (e.g., "openai/gpt-4", "anthropic/claude-3") because it attempts to look up pricing for the full prefixed model name, which doesn't exist in the pricing database.

View Details

📝 Patch Details

diff --git a/packages/app/server/src/providers/VercelAIGatewayProvider.ts b/packages/app/server/src/providers/VercelAIGatewayProvider.ts
index aefc3062..39bf3e5f 100644
--- a/packages/app/server/src/providers/VercelAIGatewayProvider.ts
+++ b/packages/app/server/src/providers/VercelAIGatewayProvider.ts
@@ -10,6 +10,19 @@ import { Decimal } from '@prisma/client/runtime/library';
 export class VercelAIGatewayProvider extends BaseProvider {
   private readonly VERCEL_AI_GATEWAY_BASE_URL = 'https://ai-gateway.vercel.sh/v1';
 
+  /**
+   * Extract base model name from provider-prefixed model (e.g., "openai/gpt-4" -> "gpt-4")
+   * Returns the original model if no prefix is found
+   */
+  private getBaseModelName(): string {
+    const model = this.getModel();
+    if (model.includes('/')) {
+      const [, ...modelParts] = model.split('/');
+      return modelParts.join('/');
+    }
+    return model;
+  }
+
   getType(): ProviderType {
     return ProviderType.VERCEL_AI_GATEWAY;
   }
@@ -81,7 +94,7 @@ export class VercelAIGatewayProvider extends BaseProvider {
     }
 
     const cost = getCostPerToken(
-      this.getModel(),
+      this.getBaseModelName(),
       prompt_tokens,
       completion_tokens
     );
@@ -110,17 +123,18 @@ export class VercelAIGatewayProvider extends BaseProvider {
     let cost = new Decimal(0);
     let metadata: LlmTransactionMetadata;
     const model = this.getModel();
+    const baseModel = this.getBaseModelName();
 
-    const modelPrice = getModelPrice(model);
+    const modelPrice = getModelPrice(baseModel);
     
     if (endpointType === 'transcription') {
       try {
         const transcriptionData = JSON.parse(data);
         const text = transcriptionData.text || '';
         
-        if (modelPrice && isValidModel(model)) {
+        if (modelPrice && isValidModel(baseModel)) {
           const textTokens = Math.ceil(text.length / 4);
-          cost = getCostPerToken(model, 0, textTokens);
+          cost = getCostPerToken(baseModel, 0, textTokens);
         } else {
           cost = new Decimal(0.01);
         }
@@ -135,7 +149,7 @@ export class VercelAIGatewayProvider extends BaseProvider {
         };
       } catch (error) {
         logger.error(`Error parsing transcription response: ${error}`);
-        cost = modelPrice && isValidModel(model) ? new Decimal(0) : new Decimal(0.01);
+        cost = modelPrice && isValidModel(baseModel) ? new Decimal(0) : new Decimal(0.01);
         metadata = {
           providerId: 'transcription',
           provider: this.getType(),
@@ -149,9 +163,9 @@ export class VercelAIGatewayProvider extends BaseProvider {
       const inputText = (requestBody?.input as string) || '';
       const characterCount = inputText.length;
       
-      if (modelPrice && isValidModel(model)) {
+      if (modelPrice && isValidModel(baseModel)) {
         const inputTokens = Math.ceil(characterCount / 4);
-        cost = getCostPerToken(model, inputTokens, 0);
+        cost = getCostPerToken(baseModel, inputTokens, 0);
       } else {
         const costPerCharacter = new Decimal(0.000015);
         cost = costPerCharacter.mul(characterCount);
@@ -166,7 +180,7 @@ export class VercelAIGatewayProvider extends BaseProvider {
         totalTokens: characterCount,
       };
     } else {
-      cost = modelPrice && isValidModel(model) ? new Decimal(0) : new Decimal(0.01);
+      cost = modelPrice && isValidModel(baseModel) ? new Decimal(0) : new Decimal(0.01);
       metadata = {
         providerId: 'audio',
         provider: this.getType(),
@@ -184,4 +198,3 @@ export class VercelAIGatewayProvider extends BaseProvider {
     };
   }
 }
-

Analysis

VercelAIGatewayProvider throws UnknownModelError for provider-prefixed models

What fails: VercelAIGatewayProvider.handleChatCompletionResponse() calls getCostPerToken(this.getModel(), ...) with provider-prefixed model names like "openai/gpt-4", but pricing database only contains base model names like "gpt-4"

How to reproduce:

// Create provider with prefixed model and process a chat completion response
const provider = new VercelAIGatewayProvider(false, 'openai/gpt-4');
const mockResponse = JSON.stringify({
  usage: { prompt_tokens: 10, completion_tokens: 10, total_tokens: 20 }
});
provider.handleBody(mockResponse); // Throws UnknownModelError

Result: getCostPerToken() calls isValidModel("openai/gpt-4") which returns false, then throws UnknownModelError: Invalid model: openai/gpt-4

Expected: Should extract base model "gpt-4" for pricing lookup per Vercel AI Gateway docs which confirm "creator/model-name" format is correct

Note: Same issue affects handleAudioResponse() method for transcription and speech endpoints