Troubleshoot Food Ordering AI Agent errors

This document is for console users and for developers who integrate with the BidiProcessOrder streaming API. It explains common errors that you might see in the Food Ordering AI Agent console or from the streaming API, what causes them, and how to resolve them.

Console access errors

The Food Ordering AI Agent console uses the Identity and Access Management (IAM) roles that are granted on your Google Cloud project. The console reads data with your credentials, so each page shows only the data that your roles allow. The following predefined roles apply to Food Ordering AI Agent:

Role What the role lets you do
Food Ordering Agent Viewer (roles/foodorderingaiagent.viewer) View brands, stores, menus, menu ingestion logs, and order history.
Food Ordering Agent User (roles/foodorderingaiagent.agentUser) Process ordering sessions, including sessions in Agent preview, and run evaluations.
Food Ordering Agent Admin (roles/foodorderingaiagent.admin) Create, update, and delete Food Ordering AI Agent resources. This role includes the permissions of the Viewer and User roles.

Failed to load brands

After you select a project and a location, the console lists the brands in that location. If the console can't list the brands, it shows Failed to load brands with the message "You may not have permission, or the API is unavailable."

To resolve this error, check the following, and then click Retry:

  • Your account has a Food Ordering AI Agent role on the project. Listing brands requires the foodorderingaiagent.brands.list permission, which the Food Ordering Agent Viewer and Food Ordering Agent Admin roles include. If you don't have either role, ask a user who manages access for the project to grant you access.
  • Your group membership is active. If your role is granted to a group, confirm that you're a member of the group. If you were added to the group before you had a Google Account for your email address, you might have received an invitation instead of membership. After you create the Google Account, ask the group owner to add you to the group again.
  • You selected a location that your role allows. If your role is limited to specific locations, go back to the project selector and, in the Select your location list, select a location where your role applies. If your role is limited to specific resource types, the role's condition must allow the Location resource type.
  • Recent changes have taken effect. IAM changes can take several minutes to take effect. For more information, see Access change propagation.

If the console lists brands but not the brand that you expect, check that the selected location is the location of the brand.

Grant access to the console

To grant Food Ordering AI Agent roles, you need the resourcemanager.projects.setIamPolicy permission on the project. You can grant roles in the Google Cloud console or by using the Google Cloud CLI. For more information, see Manage access to projects, folders, and organizations.

For some brands, you can also grant roles on the Access management page of the Food Ordering AI Agent console:

  1. In the Food Ordering AI Agent console, select your project and brand.
  2. In the navigation menu, click Access management.

    The Access management item appears only if you have the resourcemanager.projects.getIamPolicy permission on the project.

  3. Click Add member.

  4. In the Member's email address field, enter one or more email addresses, separated by commas, and then click Add email.

    The console adds addresses that end with @googlegroups.com as groups, service account addresses as service accounts, and all other addresses as individual users. To add a group that has a different address, enter the address with the group: prefix, for example group:ordering-team@example.com.

  5. In the Select Roles section, select one or more roles.

  6. In the Access Scope Boundary section, select the scope of the roles:

    • Global: the roles apply to all locations and resources in the project.
    • Regional: the roles apply only in the locations that you select.
    • Resource: the roles apply only to the resource types that you select, such as Brand, Location, SessionRecord, Store, Menu, or ChannelConfig.
    • Custom: the roles apply only to the selected resource types in the selected locations.
  7. Click Add.

A Regional, Resource, or Custom scope adds an IAM condition to the role binding. Each console page needs permissions on the resources that it reads. For example, listing brands checks the Location resource, and listing the stores of a brand checks the Brand resource. If you limit the resource types, include the types for every page that the member needs to use.

Streaming API errors

When the Food Ordering AI Agent API ends a BidiProcessOrder stream with an error, gRPC clients receive the error's status code and message. WebSocket clients receive a close frame. The close code depends on the error's status code, and the close reason contains the error message. If the connection is rejected before the WebSocket handshake completes, your client receives an HTTP error response instead of a close frame.

The following table shows how status codes map to WebSocket close codes:

WebSocket close code gRPC status codes Meaning
1000 OK, CANCELLED The stream closed without an error.
1007 INVALID_ARGUMENT, FAILED_PRECONDITION, OUT_OF_RANGE The API rejected a message or the session configuration. Fix the request before you retry.
1008 UNAUTHENTICATED, PERMISSION_DENIED, NOT_FOUND, ALREADY_EXISTS, ABORTED, UNIMPLEMENTED The API couldn't authenticate or authorize the request, or a resource that the request refers to doesn't exist.
1011 INTERNAL, UNKNOWN, DATA_LOSS, UNAVAILABLE, DEADLINE_EXCEEDED, RESOURCE_EXHAUSTED A server error or a temporary condition. For internal errors, the message doesn't include details.

Session ends before the agent greets the customer

Before the agent starts the conversation, the API validates the initial Config message and loads the store's menu. If this step fails, the stream ends with one of the following errors.

Message contains Status code (close code) Cause and resolution
no menus found for store "STORE" NOT_FOUND (1008) The API didn't find a menu for the store resource name in your Config message. For the causes, see Resolve "no menus found for store" errors.
store project ID ("PROJECT_A") does not match session project ID ("PROJECT_B") INVALID_ARGUMENT (1007) The session and store fields refer to different projects. Use the same project ID in both fields.
initial message must contain a config INVALID_ARGUMENT (1007) The first message on the stream wasn't a Config message. Send Config as the first message, and send it only once.
failed to extract location ID from path INVALID_ARGUMENT (1007) The store field isn't a full store resource name. Use the format projects/PROJECT/locations/LOCATION/brands/BRAND/stores/STORE.
is disabled FAILED_PRECONDITION (1007) The store's status is disabled. The message includes the brand and store IDs. Enable the store before you start sessions for it.

For gRPC clients, the no menus found for store and project mismatch errors also include an ErrorInfo detail with the domain foodorderingaiagent.googleapis.com and the reason MENU_NOT_FOUND or PROJECT_ID_MISMATCH.

If the stream ends with PERMISSION_DENIED (close code 1008), the principal that your client authenticates as doesn't have the foodorderingaiagent.sessions.process permission for the session. Grant the Food Ordering Agent User role (roles/foodorderingaiagent.agentUser) to the principal. For more information, see Authentication.

The API looks for the menu whose store exactly matches the store resource name in your Config message. This error occurs in the following cases:

  • No menu exists for the store yet. Create a menu for the store before you start sessions for it. In the Food Ordering AI Agent console, check that the store's menu appears on the Menus page.
  • The store ID doesn't exactly match the store ID of the menu. For example, a store ID that's missing a prefix or leading zeros doesn't match. Use the exact store resource name of the store's menu.
  • The location isn't the location where the menu was created. Use the same location, such as us, in the WebSocket URL and in the session and store fields. global isn't a Food Ordering AI Agent location. For the supported locations, see Food Ordering AI Agent locations.

Session ends during the conversation

Message or behavior Status code (close code) Cause and resolution
input audio size SIZE exceeds maximum allowed size of 25600 bytes INVALID_ARGUMENT (1007) An AudioInput message contained more than 25,600 bytes of audio. Send smaller audio chunks. If you enable echo cancellation, the limit applies to the combined customer and crew audio in each message.
received config message in the main loop INVALID_ARGUMENT (1007) The client sent a second Config message on the same stream. Send Config only as the first message.
internal error, or any error with close code 1011 INTERNAL, UNAVAILABLE, or another code that maps to 1011 A server error or a temporary condition. Reconnect and send the same Config message with the same session ID to resume the session. If the error persists, collect the session details and contact support.
The stream closes without an error before the order is complete. OK (1000) The API ends a session if it doesn't receive any message from your client for 60 seconds. For voice sessions, keep streaming audio input for as long as the session is active.

Errors in Agent preview

Agent preview in the Food Ordering AI Agent console starts sessions with the same API that your clients use, so it returns the same errors. If a session in Agent preview fails, the console shows Session ended abruptly due to an error: Server-side error during session: followed by the error message and its code. For the meaning of the message, see the tables in Session ends before the agent greets the customer and Session ends during the conversation.

To start sessions in Agent preview, you need the foodorderingaiagent.sessions.process permission, which the Food Ordering Agent User and Food Ordering Agent Admin roles include. The Food Ordering Agent Viewer role doesn't include this permission.

Find failed sessions in Order history

The Order history page shows an outcome for each session. The following outcomes indicate that a drive-thru session ended early or with an error:

Outcome Meaning
Disconnected immediately server error Before the agent took any action, the session ended with an INTERNAL or UNAVAILABLE error.
Disconnected immediately unknown Before the agent took any action, the session ended with a status code other than OK, CANCELLED, INTERNAL, or UNAVAILABLE, for example, NOT_FOUND or INVALID_ARGUMENT.
Disconnected immediately client cancellation The client cancelled the stream before the agent took any action.
Disconnected immediately normal The stream closed without an error before the agent took any action.
Order submission error The order failed to submit to the point-of-sale (POS) system.
Error The session ended because of an error, or the session doesn't match any other outcome.

In the outcome filter, the Error group includes Order submission error, Error, and Disconnected immediately server error. To investigate a session, click its row to open the session details, and use the session ID when you contact support.

Agent repeats a retry message and then ends the session

For each turn, the agent sends its instructions, the store's menu, and the conversation history to the model. If this input is larger than the model can accept, the agent can't generate a response. In this case, the customer hears a generic retry message, such as "Hmm, I'm having trouble with that. Do you want me to try again?" After repeated failures, the session ends. Depending on the agent's configuration, the stream can instead end with close code 1007 and a message that contains Input size exceeds the maximum model context window limit.

A menu that contains many entries that customers can't order can cause this issue. The limit applies to the total size of the menu and the instructions, not to a fixed number of items.

To reduce the size of the menu, create a new menu that includes only the items that customers can order in the channel. For example, remove delivery-only items, duplicate items, and entries that aren't food or drink products. If you create menus from another system, remove these entries from the source data before you create the menu. For more information about creating menus, see Integrating menu data using the API.

Information to collect before you contact support

If you can't resolve an error, collect the following information before you contact your Google representative:

  • The session resource name from the Config message, which includes the session ID.
  • The store resource name from the Config message.
  • The location in the WebSocket URL or gRPC request.
  • The approximate start time of the session, including the time zone.
  • The gRPC status code and message, or the WebSocket close code and close reason.