Choice of model
The model processes client messages, reads the system prompt, and decides when to trigger an action. A more powerful model usually handles complex rules better, but can last longer and cost more.
The list of models and their cost change. Topical options are always shown in the bot settings.
Where to choose a model
- Open My Bots.
- Press Change in front of the desired bot.
- On the Basic tab, open the Model field.
- Select the option and press Save changes.
In the list of models are divided into categories, for example Recommended, Fast and economical and Old. The relative level of value is displayed next to the name: the more $ characters, the more expensive the processing.
Which category to choose
For a new bot, start with a model from the Recommended category. Switch to a cheaper one only after checking that the quality remains acceptable.
What to check before choosing
The model must be correct:
- respect the role and limitations of the Prompt;
- Ask clarifying questions;
- Find information in the knowledge base;
- compile mandatory variables;
- call the necessary actions in the right order;
- Do not invent facts in the absence of data;
- Transfer a complex request to an employee.
An honest comparison of the two models
- Prepare 10-20 identical test dialogues.
- Save the current model and go through all scenarios.
- Write down errors, response time and cost.
- Select another model without changing the pump and other settings.
- Repeat the same set of tests.
- Leave the option that gives the right quality at an acceptable cost.
Don’t compare models on one question: a randomly successful answer doesn’t show stability.
What the cost depends on
The model doesn’t just handle the customer’s last message. The cost may include:
- system prompt;
- history of dialogue;
- found fragments of the knowledge base;
- input and output data of actions;
- model's answer.
To reduce the cost, remove repetitions from the prompt, do not connect unnecessary files and do not pass unnecessary data to the action.
When to change the model
Consider replacing if the bot regularly breaks the rules, confuses the sequence of actions, fails to cope with long context, or is too expensive. After any replacement, repeat the full testing before launching on real customers.
Do not choose an outdated model for a new project just because of a familiar name. Its functionality and support may be limited.