This article covers the most common issues teams run into while using Copilot and what to do about them. If you’re hitting a problem not listed here, contact support through the standard SqlDBM support channel.
Before you troubleshoot: Confirm that Copilot is enabled for your account and for your user. Many “Copilot isn’t working” tickets resolve to “Copilot is not yet enabled.” Ask your administrator if you’re unsure: see “Enabling Copilot for your account and users.”
The AI missed relationships I expected
Symptom: you ran Generate with AI on a requirements document and Copilot created the tables, but didn’t wire up the foreign keys the way you expected.
Why this happens: Copilot infers relationships from the information in your source document. If the document describes entities without stating how they relate, the AI has to guess from naming conventions, which doesn’t always match your intent.
What to do:
- Describe the relationships explicitly in a follow-up prompt: “Add foreign key relationships between orders, order_lines, and products.”
- Or edit your source document to make the relationships explicit and re-run Generate with AI.
- Or use the pre-prompt to state a standard assumption (“primary keys are always named <table>_ID; foreign keys follow the same pattern”) so Copilot can infer relationships consistently.
Copilot created a duplicate of an object that already exists
Symptom: you asked Copilot to generate a customer_dim table, and it created a new one instead of noticing that customer_dim already exists.
Why this happens: Copilot’s context is bounded to what it fetched for the current prompt. If your project is very large, Copilot may not have pulled the portion of the schema that contains the duplicate.
What to do:
- Point Copilot at the existing object explicitly: “Review customer_dim before creating a new one.”
- Add a pre-prompt rule: “Before creating a new entity, check whether a similar one already exists in the project and propose reuse instead of creating a duplicate.”
- For cross-project duplicates (a table exists in a different project), use global modeling references to make the shared object available in the current project. Cross-project Copilot awareness is planned for a future release.
My Excel file is too large or was rejected
Symptom: uploading an Excel file to Generate with AI returns a file-too-large error or takes an unusually long time to process.
Why this happens: Excel files carry zipped internal content that can inflate file size well beyond the useful data. A 20 MB Excel file often contains 500 KB of actual data and 19.5 MB of Excel overhead.
What to do:
- Open the Excel file.
- Save as CSV (one sheet per CSV, if multiple sheets matter).
- Upload the CSV instead. Processing is faster and the underlying data is identical.
I lost my chat when I switched pages
Symptom: you navigated to another screen in SqlDBM and returned to the Copilot panel to find your conversation gone.
Why this happens: in the current version, navigating away from the diagram resets the Copilot panel. Conversation history (if enabled on your account) preserves the prompt and response but the live chat state resets.
What to do:
- If your account has conversation history enabled, open the conversation list and re-open the prior chat to continue where you left off.
- If you’re mid-workflow and can’t afford to lose the chat, minimize the Copilot panel rather than navigating away. The minimize-and-return flow preserves state.
- For important responses you need to keep, copy them out of the panel into a document.
Copilot created output that mostly works but has one consistent error
Symptom: the proposal is 90 percent correct but there’s a pattern of mistakes: wrong naming convention on one prefix, a data type chosen for all date columns, something repeated.
Why this happens: Copilot follows the combination of your prompt, your pre-prompt, and its own defaults. If a default is winning over what you wanted, it’s usually because the pre-prompt isn’t specific enough.
What to do:
- Examine the pattern. Is there a specific rule that would prevent it? Add that rule to your pre-prompt.
- For one-off fixes, use a bulk governance prompt to fix the specific mistake across all affected objects: “Change the prefix on every table that starts with f_ to FCT_.”
- If the error is egregious, reject the whole batch, update the pre-prompt, and re-run.
The response seems incomplete: Copilot stopped midway
Symptom: Copilot’s response ends abruptly or doesn’t cover everything you asked about.
Why this happens: very long responses can hit limits. Also, if you asked about a very large portion of the project, Copilot may fetch only a subset of the schema per request.
What to do:
- Ask a more scoped follow-up: “Continue from where you stopped, focusing on the orders schema.”
- Break the question into pieces: instead of “summarize this entire 500-table project,” ask “summarize the customer schema” and then “summarize the orders schema” separately.
- For bulk actions on very large projects, run the action in batches scoped by schema or naming pattern.
The prompt worked yesterday, and today it gives different output
Symptom: a prompt that previously returned a good proposal is now returning something different, even though you didn’t change anything.
Why this happens: the model has some natural variation in output even for identical prompts. Also, the project context may have changed — a teammate may have added, renamed, or removed an object that changes what Copilot sees.
What to do:
- For predictability, tighten your prompt. More specific prompts produce more consistent output.
- Check the project’s revision history to see whether someone else made changes that would affect the answer.
- If you need the exact prior response, check your conversation history (if enabled). The prior response is preserved there.
I want to write better prompts: where to learn
General guidance:
- Be specific. “Create a customer dimension” is vague. “Create a customer dimension with standard SCD Type 2 columns: load_date, load_end_date, is_current_flag, and hash_diff” is clear.
- Use examples. If you have an existing table that represents the pattern you want, point at it: “Create a similar dimension for product, matching the structure of customer_dim.”
- Move standing rules to the pre-prompt. If you find yourself repeating the same context in every prompt, it belongs in the pre-prompt. See “Configuring Copilot settings and pre-prompts.”
- Iterate, don’t start over. If the first response isn’t quite right, refine with a follow-up rather than restating the whole prompt. Copilot keeps context.
Contacting support
If the issue you’re hitting isn’t covered here:
- Note the exact prompt, the response, and what you expected instead.
- If possible, include a screenshot of the proposal review screen.
- Contact support through the normal SqlDBM support channel.
For Copilot-specific issues, support may ask whether conversation history is enabled on your account. If it is, they can review the exact interaction; if it’s not, they’ll ask you to reproduce the issue so the exchange can be captured.
Related articles
- Getting started with Copilot
- Configuring Copilot settings and pre-prompts
- Reviewing and approving Copilot-proposed changes
- Managing conversation history and the admin toggle