Introduction to AI Chatbots with OpenAI + WOZTELL

A practical guide to understanding GPT Assistant v1.3.0, Responses API, prompts, Knowledge Bases, tools, and channels in WOZTELL

This guide explains how the new AI chatbots in WOZTELL work using OpenAI and the GPT Assistant v1.3.0 extension.

With this extension, you can create AI assistants connected to OpenAI to respond automatically across channels such as WhatsApp Business, WhatsApp Cloud API, Instagram, Facebook Messenger, and Webchat. You can also use AI to support human agents inside the WOZTELL Inbox through reply suggestions.

Important: Migration from Assistants API

Previously, WOZTELL used configurations based on OpenAI’s Assistants API.

OpenAI has announced the deprecation of Assistants API, and its shutdown is scheduled for August 26, 2026. For this reason, legacy configurations must be migrated to the new chatbots based on Responses API before that date.

The new version of GPT Assistant in WOZTELL uses Responses API, which is OpenAI’s recommended standard for new implementations of assistants, conversational agents, and AI tools.

1. What is an AI chatbot in WOZTELL?

An AI chatbot in WOZTELL is an assistant connected to OpenAI that can interpret user messages, generate automatic responses, and use additional tools to provide better answers.

Depending on the configuration, the assistant can:

• Automatically reply to customers on WhatsApp, Instagram, Facebook Messenger, or Webchat.
• Search documents from a Knowledge Base before responding.
• Search for up-to-date information on the web.
• Execute external functions through Function Call.
• Support human agents with reply suggestions inside the WOZTELL Inbox.

This makes it possible to automate repetitive conversations, improve response speed, support the sales team, and deliver more consistent answers across all conversational channels.

2. What is GPT Assistant v1.3.0?

GPT Assistant v1.3.0 is the WOZTELL extension that allows you to connect an OpenAI account with WOZTELL’s conversational channels.

The main configuration is divided into three key steps:

1. API Key Setup: connects the OpenAI account using an API Key.
2. Configuration: configures the model, prompt, tools, Knowledge Base, and suggestions.
3. Connect to Channel: connects the assistant to one or more WOZTELL channels.

Once configured, the assistant can be activated on the available channels within the WOZTELL account.

3. Which API does WOZTELL use now?

Responses API

The new version of GPT Assistant uses OpenAI’s Responses API.

Responses API is the API recommended by OpenAI for new AI assistant implementations. It allows you to work with models, instructions, context, and tools such as File Search, Web Search, and Function Call.

In WOZTELL, conversational context is managed by WOZTELL. This means WOZTELL decides how many previous messages are sent to the model based on the assistant’s configuration.

Why is this migration important?

The migration is important because legacy configurations based on Assistants API will no longer be supported by OpenAI starting on August 26, 2026.

For this reason, customers still using previous configurations must migrate their assistants to the new configuration based on Responses API.

4. Basic OpenAI concepts

API Key

The API Key is the credential that allows WOZTELL to connect with the customer’s OpenAI account.

Without a valid API Key, WOZTELL cannot send requests to the model or correctly activate the assistant configuration.

Model

The Model is the OpenAI model that generates the assistant’s responses.

The list of available models is retrieved dynamically from the connected OpenAI account. This means not all customers will see exactly the same models.

Token

A token is the unit of text processed by the model. It can be a word, part of a word, a number, a space, or a punctuation mark.

OpenAI costs are usually calculated based on input and output tokens.

Input Tokens

These are the tokens sent to the model. They may include:

• The user’s message.
• The assistant’s instructions.
• The conversation history.
• Fragments retrieved from File Search.
• Web Search results.
• Data returned by Function Call.

Output Tokens

These are the tokens generated by the model in its response.

Context

Context is all the information sent to the model so it can respond correctly.

It may include instructions, previous messages, Knowledge Base information, tool results, and external data.

Context Window

The context window is the maximum amount of information a model can process in a single request.

The more context is sent, the higher the cost may be and, in some cases, the slower the response may become.

5. How to choose the right model

The choice of model depends on the use case, conversation volume, response complexity, and available budget.

Before publishing prices or creating a commercial proposal, it is recommended to check OpenAI’s official pricing page, as models and costs may change.

6. How to configure the assistant’s behavior

Instructions / Prompt

Instructions, also known as the base prompt or system prompt, define how the assistant should behave.

A good prompt should include:

• The assistant’s role.
• Communication tone.
• Topics it can answer.
• Topics it should avoid.
• Information it should request from the user.
• Human escalation rules.
• Expected response format.

Base prompt example

You are a presales assistant for WOZTELL.

Your goal is to help companies understand how WOZTELL can help them automate conversations across WhatsApp, Instagram, Facebook Messenger, and Webchat.

Be friendly, concise, and professional.

Only answer questions related to WOZTELL, WhatsApp Business Platform, AI chatbots, integrations, automation, and conversational channels.

If the user asks about something unrelated, politely explain that you can only help with WOZTELL-related topics.

Before qualifying the lead, ask for their name, company, website, and main use case.

Negative Constraints

Negative constraints are rules that indicate what the assistant should not do.

Do not invent prices.
Do not promise implementation timelines unless confirmed.
Do not answer questions outside the scope of WOZTELL.
Do not provide legal, financial or medical advice.
Do not say you are an AI unless the user asks directly.

Human Handover / Escalation

Human Handover, or escalation to a human, happens when the assistant needs to transfer the conversation to an agent.

Escalation is recommended when:

• The user asks to speak with a person.
• There is a pricing question.
• There is a technical issue.
• There are billing questions.
• The assistant does not have enough information.
• The conversation requires human validation.

7. Available tools in GPT Assistant

File Search

File Search allows the assistant to search for information inside a Knowledge Base before responding.

It is useful when the assistant needs to answer based on official documentation, FAQs, product guides, internal processes, or commercial materials.

Web Search

Web Search allows the assistant to search for up-to-date information on the internet.

It is recommended when the answer depends on information that may change, such as external documentation, prices, regulations, news, or recent public information.

Domain Filtering

Domain Filtering allows web searches to be limited to specific domains.

Correct example:

woztell.com
support.woztell.com
openai.com

Incorrect example:

https://woztell.com/

Function Call

Function Call allows the assistant to call an external function to retrieve data or execute an action.

Use case examples:

• Checking order status.
• Verifying an appointment.
• Looking up customer data.
• Creating a ticket.
• Sending information to a CRM.
• Validating availability.

8. Knowledge Base and Vector Stores

Knowledge Base

The Knowledge Base is the set of documents the assistant can consult to answer questions.

It may include:

• FAQs.
• Installation guides.
• Product documentation.
• Sales playbooks.
• Internal policies.
• Support materials.

Vector Store

The Vector Store is the technical container where processed documents are stored so the assistant can search for relevant information.

A simple way to understand it:

• Knowledge Base: the document library.
• Vector Store: the technical container where documents are indexed.
• File Search: the tool that searches inside that library.

Retrieval

Retrieval is the process through which the assistant retrieves relevant fragments from the Knowledge Base and uses them as additional context to respond.

Best practices

• Use clear and updated documents.
• Remove outdated information.
• Avoid duplicated documents.
• Do not mix contradictory versions.
• Separate technical, commercial, and support documentation when necessary.
• Review the Knowledge Base periodically.

9. AI Chatbot vs AI Inbox Replies vs SuggestGPT

AI Chatbot

The AI Chatbot replies automatically to end users on channels such as WhatsApp, Instagram, Facebook Messenger, and Webchat.

AI Inbox Replies

AI Inbox Replies helps human agents by generating reply suggestions inside the WOZTELL Inbox. The agent can review, edit, and send the reply manually.

10. Channel connection

Channel

The Channel is the communication channel where the assistant will reply to users.

Supported channels:

• WhatsApp Business
• WhatsApp Cloud API
• Instagram
• Facebook Messenger
• Webchat

Environment

The Environment is a space within a channel that allows you to separate testing, development, and production.

If no additional environment has been created, the default environment will be used.

Prefix Name

The Prefix Name is the identifying name of the assistant configuration connected to a specific channel or environment.

Examples:

AI-CHATBOT-WHATSAPP-PRESALES
AI-CHATBOT-INSTAGRAM-SUPPORT
AI-CHATBOT-WEBCHAT-FAQ
AI-CHATBOT-FACEBOOK-LEADS

Connect

Connect links the assistant to the selected channel and activates it.

Disconnect

Disconnect unlinks the assistant from the selected channel.

Go Live

Go Live is the moment when the assistant becomes active on a real channel and starts replying to end users.

11. Costs and usage control

The cost of using OpenAI mainly depends on the model, the tokens processed, and the tools used.

Factors that can increase cost

• Prompts that are too long.
• Very large conversation history.
• Very long responses.
• Frequent use of File Search.
• Frequent use of Web Search.
• More advanced models.
• Long audio messages that require transcription.
• Large or poorly structured Knowledge Bases.

Cost Control best practices

• Use efficient models for high-volume use cases.
• Limit the number of previous messages sent as context.
• Keep responses concise.
• Enable Web Search only when necessary.
• Keep the Knowledge Base clean and updated.
• Escalate to a human when the conversation is complex.
• Review logs and usage periodically.

12. Quality, accuracy, and prevention of invented answers

Hallucination

A hallucination is a response generated by the model that may sound correct but is not based on real or confirmed information.

How to reduce hallucinations

• Use clear instructions.
• Limit the assistant’s scope.
• Enable File Search when official documentation exists.
• Keep the Knowledge Base updated.
• Tell the assistant not to invent information.
• Define clear human escalation rules.

If the answer is not available in the Knowledge Base or confirmed documentation, do not invent it. Ask the user to contact the WOZTELL team or escalate to a human agent.

Grounding

Grounding means basing the assistant’s responses on specific sources, such as Knowledge Base documents, Web Search results, or data returned by a Function Call.

13. Logs and troubleshooting

Logs

Logs are technical records that show events, requests, and errors generated during the assistant’s operation.

They are used to diagnose problems related to:

• API Key.
• Models.
• Incompatible parameters.
• Channels.
• File Search.
• Web Search.
• Function Call.
• OpenAI errors.

14. Checklist before activating the assistant

• Valid API Key saved.
• Model correctly selected.
• Clear and tested prompt.
• Escalation rules defined.
• Appropriate Temperature.
• Limited chat history.
• Format configured as Text for regular chatbots.
• Updated Knowledge Base if File Search is used.
• Web Search limited to approved domains if applicable.
• Function Call tested if an external integration is used.
• Correct channel connected.
• Correct Environment selected.
• Clear Prefix Name.
• Testing completed before Go Live.
• Logs reviewed after the first conversations.

Summary

The new WOZTELL AI chatbots, powered by OpenAI and Responses API, make it possible to create more flexible assistants that can operate across multiple channels and use tools such as File Search, Web Search, and Function Call.

This new architecture progressively replaces the previous configurations based on Assistants API. For this reason, customers who are still using legacy assistants must migrate before August 26, 2026.

Correctly configuring the model, prompt, Knowledge Base, tools, channel, and escalation rules is essential to create an assistant that is reliable, useful, and cost-controlled.