Stuck?
Find it here

Usually this means no payment method has been registered.

Go to console.upstage.ai → Dashboard → Billing, register a card, and try again. While you have credits, those are used first.

Your key is not being passed correctly. Check in this order.

1. Confirm Bearer comes before the key.
2. Check that no spaces or quotation marks were copied with the key.
3. Confirm the key starts with up_.
4. Create a fresh key in the console and swap it in.

Check your Colab secret settings.

1. Under the key (🔑) icon on the left, confirm the name is exactly UPSTAGE_API_KEY, including case and underscores.
2. Confirm notebook access is turned on for that secret.
3. After turning it on, re-run your first cell for it to take effect.

A VPN or ad blocker sometimes interferes. Turn both off and try again.

If you are on an institutional network, check the firewall policy as well.

You have gone over the rate limit. This is not a failure. Tier 0 allows 100 requests and 50,000 tokens per minute.

• Wait a few seconds instead of resending immediately. If it repeats, wait longer.
• Reduce how many run at the same time.
• In a classroom, split into groups and run in turns.

The 429 response includes X-Upstage-RateLimit-Retry-After-Requests. You may retry after that time.

Document Parse allows 1 request per second with the synchronous API. Sending them together exceeds that.

Leave about a second between requests, and record how far you got so you can resume from the failures. See the loop example in Part 3.

With 100 requests per minute, 30 students running several times each fills the limit quickly.

• Split into groups and run with a time gap.
• Running it once yourself before class gives you a feel for the pace.
• If students use separate accounts, the limit applies per account.

Purchasing additional credits raises your tier and the limits with it. You can check the limits for your new tier under Billing → Commitment → Rate limit in the console.

For institution-scale needs, contact Upstage support.

There are three possible causes.

First, you requested a model outside the covered set. The covered language models are solar-pro2, solar-pro3, and solar-pro4. Calling anything else, such as solar-mini, gives you this error.

Second, the address is wrong. Each product has its own address. The message points at the model, but the real cause is often the URL.

Third, you called solar-pro2 after mid-October 2026. Support ends then and the model can no longer be called. Change it to solar-pro4.

Change only the model name. Usage, speed, and pricing are the same, so nothing else needs to change.

Do it before mid-October 2026. Solar Pro 2 support ends then, and after that the model cannot be called at all. Changing it now is easier than fixing it under pressure later.

Wording and structure may shift slightly, so if you have fixed output formats it is worth checking the results once.

You are most likely calling a model the program does not cover. It covers exactly four: solar-pro2, solar-pro3, solar-pro4, and document-parse.

The easily confused ones — information-extract, document-classify, and the embedding models (solar-embedding-2-query, solar-embedding-2-passage) — are all separately billed. They work fine, so there is no error; the credits just quietly go down.

Worth checking especially if a coding agent wrote the code. Asked to extract fields, agents often reach for information-extract. See Part 3 · What is outside the program for how to do the same job within the covered set.

This happens when a Solar example is lifted as-is. The two products have different response shapes.

Solar returns text at choices[0].message.content, but a Document Parse response has no choices at all. Read content.html, content.markdown, or elements[] instead.

The usage unit differs too: usage.total_tokens for Solar, usage.pages for Document Parse. There is a comparison table in Part 3 · Working with documents.

That is the nature of language models. To make results more consistent:

• Fix the output format. The approach in Part 3 pins field names and types.
• Fix the allowed values. Wrapping verdicts in an enum restricts the answer to your list.
• State criteria as numbers and conditions. "Hold if over 100,000" is far more stable than "judge appropriately".

Check the image quality and the processing mode.

1. Switch OCR to Force and run again. This helps most with scans and photos.
2. Confirm the document is at least 640 pixels wide.
3. Confirm the smallest text is at least 2.5% of the image height.
4. If tables and figures are complex, try Enhanced mode.

Check these three things.

• Format: JPEG · PNG · BMP · PDF · TIFF · HEIC · DOCX · PPTX · XLSX · HWP · HWPX are supported.
• Size: must be 50MB or less.
• Filename: very long names can cause problems, so try shortening it.

The synchronous API processes up to 100 pages per file; anything beyond that is not included.

Split the document, or use the asynchronous API which handles up to 1,000 pages.

If tables continue across pages, turn on Merge Multipage Tables.

If numbers inside charts matter, check Chart recognition as well.

You can see the table structure in content.html.

Usually this means you did not include the material the answer should be based on.

Put the document or rules into the prompt. Adding "if it is not in this material, say 'not found in the material' rather than guessing" makes a large difference.

Check two things.

First, whether the information is actually in the document. If it is not there, an empty value is correct.

Second, whether your description is specific enough. "The final payment amount including tax" finds the value far better than "amount". Saying where in the document it appears helps too.

Check finish_reason before parsing.

stop means it completed. length means it hit the token limit and is cut off mid-object, so it cannot be parsed. Increase max_tokens or retry.

Also confirm you set strict to true and additionalProperties to false.

Ask about facts the model does not have and it may produce something plausible but wrong.

• Include the source material in your request.
• Require the basis for the answer — a page number or quotation — so it is easy to check.
• Tell it to say when it does not know.
• Have a person verify important values such as amounts, dates, and names against the original.

Usually the description is too vague. The model decides when to call a tool from its description.

Write what the tool is for: "Look up budget execution status by department and year" rather than "data lookup".

Note that not calling a tool for something like a greeting is correct behaviour. Test with a question that genuinely needs the tool.