400 prompt token count of 100701 exceeds the limit of 90000

[jimed99]

Made my first AI request of the day and that's the error I was met with. Any ideas anyone?
Ask mode; Claude Sonnet 3.7
Gemini 2.5 went through

[brad]

I have no idea how tokens work, but maybe your request was too long? Just guessing here. No idea how that works.

[jimed99]

68 words, 410 characters - and Gemini responded without hesitation - so I'm betting it's Claude.
But that whole token counting has got me lost too..

[Cheese]

Lets cheat and use AI to explain it...

Token Calculation by Model

1. OpenAI GPT-4 / GPT-4o

• Tokenizer: Uses tiktoken, a byte pair encoding (BPE) tokenizer.
• Token size:
  • 1 token ≈ 4 characters (English), 0.75 words.
  • Emojis, punctuation, and whitespace are all tokenized.
• How it's calculated: You can use tiktoken to tokenize and count tokens.
• Example:
  "I love AI." → 5 tokens: ["I", " love", " AI", ".", ""]

---

2. Anthropic Claude (Opus, Sonnet, Haiku)

• Tokenizer: Custom tokenizer, similar to sentencepiece.
• Token size:
  • 1 token ≈ 3–4 characters on average.
• Notable Feature: Claude's tokenizer is more optimized for English than OpenAI's in some ways, meaning fewer tokens for the same sentence.
• How it's calculated: Anthropic has not open-sourced its tokenizer, but 3rd party tools estimate similar token counts to OpenAI's.

---

3. Google Gemini (formerly Bard)

• Tokenizer: Based on SentencePiece, often with BPE or Unigram LM.
• Token size:
  • 1 token ≈ 3–4 characters (varies significantly by language).
• Multilingual support: Highly optimized for multilingual input.
• How it's calculated: Use Google’s open-source SentencePiece models to replicate.

Let's break down what 100,701 tokens roughly equates to.

Since token-to-word ratios vary by language and complexity, here's a range:

Language Complexity	Words per Token	Estimated Words for 100,701 Tokens
Simple English (average)	~0.75	~75,525 words
Complex English (technical/legal)	~0.66	~66,460 words

So, 100,701 tokens ≈ 66,000–75,000 words.

---

Web Page Size Estimate

1. By Word Count

Based on common webpage word counts:

Page Type	Avg. Word Count	100,701 Tokens Can Fit...
Blog Post	~1,000	~66–75 pages
Marketing Landing Page	~500	~132–150 pages
Technical Docs Page	~2,000	~33–38 pages

---

2. By Character Count / File Size

• 1 token ≈ 4 characters → 100,701 tokens ≈ 402,804 characters
• That’s about 400 KB of raw text (uncompressed).
• In terms of HTML page size, with basic styling/markup, that would be about 500–700 KB.

For comparison:
A typical full-featured web article (with light media and CSS) is 200–500 KB.

Summary:

• 100,701 tokens = about 66,000–75,000 words
• That's like:
  • A short novel (e.g., The Great Gatsby)
  • A manual or guidebook
  • 60–75 blog posts
  • A medium-sized website's text content

[jimed99]

LOL - showoff -

but how do I get an error on the first paragraph of the day

[Cheese]

@jimed99

LOL - showoff

Hahahahaha! I asked AI!

@jimed99

but how do I get an error on the first paragraph of the day

Ask Claude. I'm sure it will apologise in a very nice way.

In all seriousness I don't know Jimed. Seems very strange though.

[George]

See also my explanation about the context size and tokens usage as we supply Wappler specific knowledge to the AI models:

My Head Programmer thinks there is an AppConnect issue

Do note @Cheese and @Apple that within Wappler we supply the AI models with extensive knowledge of the used App Connect and Server Connect frameworks working and syntax when calling them in their system prompt. So the AI model can then translate the generic knowledge they have to Wappler specific solutions and use the App Connect expressions syntax and available components for the frontend as well generate or edit Server Connect actions with the available actions and syntax. Also project speci…

So on larger pages it is advisable to use models with larger context size.

We will be adding soon also the Google Gemini as direct provider which goes up to 1MB on tokens context size and should be plenty

[Cedric_Sobey]

I get a similar error:
400 Invalid 'messages[10].tool_calls': empty array. Expected an array with minimum length 1, but got an empty array instead.

It always occurs on a follow up chat. First message goes through no problem.
Tried OpenAi 40 and 4.1

Happens on Wappler 7.01 and 7.1.0

Another issue I have is if you have a very long prompt if fulls the entire vertical height of the AI window so it's impossible to see the response.

[George]

in Wappler 7.1 now we have also added Google Gemini as a new AI provider, so you can now use Gemini 2.5 Pro or Gemini Flash 2.5 with their 1MB tokens large context window and reasoning capabilities.

So this should be able to plenty for you w large files.

[patrick]

The tokens used are more than only the message you type.

Each request will contain:

• System Prompt: this is generated by Wappler and is dynamic. It contains the instructions for the AI on how it should react and depending on the editor information about the used frameworks. If you are in the html editor that uses App Connect the system prompt will include a lot of documentation about it.
• Tools: information about the tools the AI can use.
• Tool Result: when a tool was called the result is send back to the AI in a new request.
• User Prompt: the message you typed.
• Previous Messages: each request will include all previous messages and tool results; the AI itself has no memory so the whole conversation is being send each time.

Most tokens are used in the HTML editor for the App Connect documentation, the AI doesn't have enough knowledge about App Connect to generate correct code and the extra instructions are needed for correct code generation.