How is consumption measured in Botmaker's generative technologies?

The generative technologies in Botmaker are used to analyze and generate words. The platform uses tokens as a unit of measurement, a word can represent between 1 and 5 tokens, depending on the number of syllables and the language.

How to estimate token consumption?

Tokens are the basic units used by language models to process text. Their size and composition vary according to multiple factors:

• The language used;
• The context of the surrounding words;
• The special characters and punctuation marks;
• The specific AI model being used.

For example, in the sentence “Hello, how are you?!” which contains 19 characters, the model consumes 6 tokens. This happens because the model can:

• Separate whole words such as “Hello” into one token;
• Group punctuation marks such as “?!” into a single token;
• Splitting longer words into multiple tokens according to common language patterns.

It is important to understand that there is no fixed rule for calculating tokens, as each language model (LLM) has its own tokenization system based on how it learns to give meaning to words and their components (n-grams).Botmaker uses different LLM models, aggregates the amount of tokens and has a fixed price per block of 10,000 tokens consumed, which can be seen in this section.

What products consume tokens in generative AI?

Within the platform, the following functionalities consume tokens:

• Copilot: consumes tokens from the previous messages of the conversation for context (approximately 20 last messages), from the instructions given and from the messages generated as suggestions for the agent.
• AI agents: consume tokens from the previous messages of the conversation for context (approximately 20 last messages), from the instructions given, the queries to the content bases (according to the paragraphs of relevant content found) and from the result generated by the generative AI.
• Generative text responses: consume tokens from the previous messages of the conversation for context (approximately 20 last messages), from the instructions given, the queries to the content bases (according to the relevant content paragraphs found) and from the result generated by the generative AI.
• Content-based response: consumes tokens from the previous messages of the conversation for context (approximately 20 last messages), from the instructions given, the queries to the content bases (according to the relevant content paragraphs found) and from the result generated by the generative AI. In addition, recommended buttons that query the content bases are added.
• Generative natural languages: each message received by a bot with generative natural languages consumes tokens to verify if the user's message complies with any of the instructions of the configured natural languages.
• Training of content bases: the content bases of files or websites consume tokens for the extraction of relevant content from them.

Token consumption can be estimated as between 2.5 and 5 times the size of the text (or HTML code) processed in tokens.

The cost of the token varies according to the language model used. You can visualize the values by clicking here: Link