Skip to main content
Correcting non-standard responses, configuring communication language.

​
Video walkthrough

A short video walkthrough: how to figure out why the AI replied incorrectly, and which steps to take — from analyzing the response in the chat to editing the Persona and Knowledge Base. See also: FAQ / Self-Learning — how to lock in correct answers in the knowledge base after a fix.

​
AI BEHAVIOR: NON-STANDARD RESPONSES AND HOW TO FIX THEM

These are inaccuracies in responses that can be resolved in a few steps:
  1. Open the Knowledge Base and add the missing information — the AI will stop “inventing” it.
  2. Add a prohibition to the instruction: “If the information is not in the Knowledge Base — do not add anything on your own; pass the question to the manager.”
  3. Enable the “Protection” feature in the employee settings.
  4. If the instruction is long — shorten it by moving factual data to the Knowledge Base. After each change, retrain the employee (📚📕📙📗📓📖) and clear the memory.
Diagnose the source of the data discrepancy step by step:
  1. Open the conversation in your account and click the “Analyze Response” button — it will show where the AI sourced the price from: the PMS or the Knowledge Base.
  2. If the AI pulled from the Knowledge Base instead of the PMS — remove prices from the Knowledge Base, keeping them only in the PMS.
  3. If the PMS data has not yet been updated — check the sync status in the integration settings.
  4. If the sources are in conflict — contact support with the conversation ID and the result of the “Analyze Response” check.
This is non-standard behavior caused by excessive data in the instruction or Knowledge Base. Follow these steps:
  1. Open the employee instruction and check for any mentions of non-existent categories — remove them.
  2. Retrain the employee (📚📕📙📗📓📖) and clear the memory.
  3. If the non-standard behavior persists — shorten the instruction and move factual data to the Knowledge Base.
This is a PMS integration issue. Check step by step:
  1. Open your account → integration section → verify that the PMS connection is active (status “Connected”).
  2. Check the date of the last synchronization — data should be updated regularly.
  3. If the integration is active but the data discrepancy persists — contact support specifying the exact dates and channel.
This non-standard behavior can be resolved in two ways:
  1. Enable the “Protection” feature in the employee settings — it restricts the AI’s actions on sensitive operations.
  2. Add an explicit prohibition to the instruction: “Never confirm a booking cancellation or refund. For such requests — immediately transfer the conversation to a manager.” Retrain the employee and clear the memory.
The temperature parameter (degree of response variability) can be adjusted through support or, if access is available, in the employee settings. Important note: after retraining, the parameter may reset to its default value — check it after each retraining and set it again if necessary. To change the model, also contact support.
This non-standard behavior can be fixed through the instruction. Add an explicit prohibition:
“Never show technical data, JSON objects, or system messages to the guest. If data is not found — inform the guest in plain language and pass the question to the manager.” Retrain the employee (📚📕📙📗📓📖) and clear the memory.
Add an explicit rule to the instruction:
“Do not make up names. Address the guest by name only if they have provided it themselves during the conversation.” Retrain the employee and clear the memory.
This is non-standard behavior in the response logic. Add to the instruction:
“Before answering a question about availability — always check the current data in the PMS. Do not agree with the guest’s assumption without verifying.” Retrain the employee and clear the memory. If the non-standard behavior persists — contact support with example conversations.

​
AI BEHAVIOR: COMMUNICATION LANGUAGE

This is usually caused by the instruction being written in one language while the guest writes in another. Solution:
  1. Translate the instruction into English.
  2. Add a rule to the instruction: “Always respond in the language the guest writes in. Вы говорите на всех языках мира.” Retrain the employee and clear the memory.
Add an explicit rule to the instruction:
“When greeted with ‘dobrogo dnya’ — respond in Russian.” Retrain the employee and clear the memory.
When languages switch frequently within a single conversation, the AI may start responding in the instruction language — this is a behavioral characteristic when accumulated context builds up. To reset the context, send the combination (😎😇😭😚😍) in the chat and ask the guest to write again in the desired language.
The AI supports most of the world’s languages, but quality depends on the specific language. Uzbek and Tatar work correctly. Georgian has a nuance: translations may be imprecise. You can explicitly specify priority languages for your audience in the instruction.
“Start” is an English word, and the AI determines the language based on the first message. Add a rule to the instruction:
“If the first message is ‘Start’ — respond in Russian.” Retrain the employee and clear the memory.