Model call errors

Every model call that fails in DataMind Oda comes back as one class, with its own message and, when there is one, a short code. This page lists the classes, the wording you may meet, and the fixes you can apply yourself. Model calls and failures covers how calls can end and how the temporary failures are retried for you.

The classes Oda reports

The code column names the class of the failure. Keep the message and its code together when you ask for help, so the same problem can be looked up wherever it appeared.

What Oda reports	Code	What it means	What to do
Nothing available to call	None	Oda found no way to reach the model, for example its provider was removed or was never set up.	Choose another model, or check the provider on Settings > Models.
A missing credential	`MISSING_CREDENTIAL`	The call needs a key or a saved sign-in, and none is in place.	Store the key the message names on Settings > Models; the message names where Oda looked, never the key.
A refused sign-in	`AUTH`	The service turned the key or the sign-in away.	Replace the key, or sign in to the provider again, on Settings > Models.
Rate limiting	`RATE_LIMIT`	The service is taking too many requests and asked Oda to slow down.	Oda retries it for you first; wait for the next attempt.
A problem on the service's side	`SERVER`	The service reported its own problem or answered with an error.	Oda retries it for you first; if it repeats, keep the message for your administrator.
A full context	`CONTEXT_WINDOW_EXCEEDED`	The request was larger than the model's context window.	Make the next request smaller; in a long chat Oda shortens the older part on its own (system prompt, context and compaction).
An unusable key	`INVALID_CREDENTIAL`	The stored key is blank, or holds characters Oda cannot send, so the call is refused before it goes out.	Enter the key again on Settings > Models; the message names where the key comes from, never the key.
A spent quota	`QUOTA`	The account behind the service has no quota, credit or budget left.	Check the account and top up (your Platform account).
An empty answer	`EMPTY_RESPONSE`	The service finished the call with no content.	Oda retries it for you first; if it comes back empty again, send the message once more.
An image the service cannot take	`INVALID_REQUEST`	A single image is larger than the service accepts.	Use a smaller copy of the image; attachments in model requests covers the limits.
Images that do not fit	`IMAGE_OFFLOAD_REQUIRED`	The request carries more images, or more image bytes, than the route's budgets allow.	Oda sets the oldest images aside first; if that is not enough, send fewer images.
A request that cannot be sent	`INVALID_REQUEST`	The request falls outside what the model accepts.	Fix what the message names; if it repeats, keep the message for your administrator.
A connection problem	`TRANSPORT`	The connection to the service failed or dropped.	Oda retries it for you first; if it keeps failing, check your connection.
A slow or silent reply	`TIMEOUT`	Oda waited for activity on the reply and stopped waiting after the route's stream idle timeout.	Send the message again; the wait can be raised in the route settings.
A cut-off stream	`STREAM_CLOSED`	The reply stopped before the service said it was finished.	Oda retries it for you first.
A malformed answer	`MALFORMED_RESPONSE`	The service's answer did not have the shape Oda reads.	Send the message again; if it repeats, keep the message for your administrator.
You stopped it	`ABORTED`	The call was cancelled, for example when you pressed Stop.	Nothing: a stopped turn is not a failure (failures and recovery).
Update required	None	Your DataMind server needs a newer version of DataMind Oda before it will answer.	Take the update from the notice; troubleshooting and update DataMind Oda have the steps.

Messages you may meet

A missing key or sign-in. When a call needs a key and none is in place, the failure names the exact name Oda looked for, such as no API key for the configured provider route or no credential for provider route, and where to put it: on Settings > Models, or in the environment Oda was started from. The key's value is never printed, and Oda does not fall back to another name, so a key saved under a different name does not help. Set up models and providers shows where to store it.
A provider Oda cannot read. When you read a provider's model list and the read fails, Oda explains what went wrong: it could not reach the address; the address answered with an error, with the message asking to check the API key; or the answer did not answer with JSON. When the answer holds no model list Oda can read, the message suggests you enter this provider's models by hand, which you do with Add model on the provider's entry. The list you had stays until a read succeeds.
A route that clashes with one already in place. If a provider you save conflicts with a route already set up, the change is refused and the previous setup keeps working; the message says configuration conflicts with an existing provider route. Nothing is saved until the clash is gone; check the provider you already have, or give a custom provider a different id.
A sign-in Oda cannot keep. Browser sign-in is offered for the providers that support it and that Oda can keep a saved sign-in for; where it cannot, a key is the way in. The sign-in steps and wording are with the sign-in flow in set up models and providers.

When an old answer cannot be reused

Every answer keeps, beside its text, the details its provider needs to read the answer back in a later request.
If those details no longer fit the answer, for example after a chat moves to a different model, or when the answer was written by an older version of Oda, Oda sends that one answer as plain content and carries on. The call is not failed.
The reasons stay inside Oda; nothing is shown for this and you do not need to do anything. The conversation is kept either way. Model selection and defaults covers switching the model for a chat.

Questions, answered

The key is stored. Why does the call say one is missing?

Oda reads the exact name the failure shows: the key saved under that name on Settings > Models, or the same name in the environment Oda was started from. It does not try other names, so a key saved under a different name is not used. Store the key under the name in the message, or point the provider's Credential setting at the name you used (set up models and providers).

Why does Oda ask me to `check the API key`?

That wording appears when you read a provider's model list and the address answered with an error for the key. Check the key on Settings > Models: enter it again if it was replaced or no longer works, then read the list again. Set up models and providers walks through the refresh.

What does `enter this provider's models by hand` mean?

The address answered without a model list Oda can read, so the message suggests entering the models yourself instead of reading them: choose Add model on the provider's entry and fill in each model's id, name, context window and limits.

Does a chat still work after I switch to another model?

Yes. Each request goes to the model the chat uses at the time. For an earlier answer whose saved details no longer fit the new model, Oda sends that one answer as plain content instead of failing the call, so the chat carries on. The conversation itself is kept either way; model selection and defaults covers switching.

Why does a failure show a code?

The code is the failure's class in a short form, such as AUTH or QUOTA, so the same problem reads the same whichever provider answered, and support can look it up. Keep the message and the code together when you ask for help; failures and recovery shows how a failed turn looks.

Why did a call finish with no answer?

The service ended the call without any content, which Oda reports as an empty answer with the code EMPTY_RESPONSE. Oda retries it for you first; if empty answers keep coming, send the message again, and keep the failure for your administrator if it still repeats.

Why did the model call time out?

Oda stops waiting when a streaming reply goes quiet for longer than the route's stream idle timeout, five minutes by default; the call then fails with the code TIMEOUT. Send the message again, and raise Stream idle timeout in the route settings if long replies regularly go quiet.