Instructions

Give your chatbot a purpose.

Think of Instructions as setting the ground rules and personality for your chatbot. You’re telling it how to behave and giving it basic, essential information.

Feel free to provide further details about your business, contact details, and any contextual data. This is particularly useful for behavior-related aspects, contact information, and contextual data.

AI Engine lets you use a placeholder like {CONTENT} in your instructions. This automatically inserts the text, from the current webpage (post, page, …) the user is on, to the chatbot’s instructions. 

This gives the chatbot immediate context about what the user is seeing. You can also build your own placeholders through dynamic context, we will explain that later.

Embeddings

Give your chatbot memory and knowledge 

If you want your chatbot to “know” a lot of specific information – like the content of your website, product manuals, books, or company documents – Embeddings are the way to go. This is often implemented using a technique called Retrieval Augmented Generation (RAG).

AI Engine lets you easily (by allowing you to sync your WordPress posts or upload PDF files) build your own textual knowledge base to store information that can then be added to the chatbot’s context so it knows what to respond with. But keep in mind: this is not a traditional database that the chatbot can search through.

Learn how to use Embeddings in AI Engine by reading this documentation.

Vector Store

Work with files

Embeddings are not the only way to build a knowledge base. You can also directly upload files into a Vector Store. This way, you can upload files like PDFs directly to a storage space in OpenAI that your Assistant or Chatbot can search through.

The OpenAI Assistants are being deprecated and will no longer be maintained by the OpenAI team.

Moving forward, AI Engine will rely on the Responses API, but you’ll still be able to use Vector Stores.

You can create a Vector Store directly in your OpenAI Dashboard and then link it to your chatbot so it can continue retrieving and using your stored data without relying on Assistants.

Dynamic Context

Let your Chatbot know in real time.

This is great for information that isn’t static, like product availability, today’s news, or user-specific data.

Dynamic Context allows your chatbot to get and use information that changes frequently or is specific to the immediate situation, during the conversation. This is not a “feature” per se but rather the concept of adding information to your chatbot context dynamically.

Filters

You can set up WordPress filters provided by AI Engine that let you modify the user’s query, change the chatbot’s planned reply, or even edit the context at query time. For instance, you saw earlier the {CONTENT} placeholder provided by the plugin for the content awareness feature. You can build your own placeholder, like {WEATHER}, and replace it with the current weather at the user’s location. Find all the filters here.

Function Calling

This is a more advanced method where the AI itself can decide it needs more information ( or run any process really, like sending a mail ) to answer a question. It can then “call a function” – essentially run a pre-defined piece of code ( that you make ) – to fetch data (e.g., search your website for a specific product, check current stock levels, or perform a web search). This new information is then fed back to the AI to help it answer. You can learn more about Function Calling here.

Implementing this usually requires some technical know-how. You can build everything yourself, but we at Meow Apps have developed add-ons that handle this for you, like the “Web Search add-on” or the “Better Links add-on” to recommend your posts and products to your clients. If you purchase them, you get access to premium support, and we ensure they are always up-to-date with AI Engine. Alternatively, you can also contact a freelancer to help you develop exactly what you are looking for.

Find all of our add-ons here: Meow Apps Add-ons.

Fine-Tuning

Turn your chatbot into a specialized tool

Fine-tuning is different from giving your chatbot knowledge (like with embeddings). Instead, it’s about teaching the AI model how to behave, how to respond in a particular style, or how to perform a specific task that it wasn’t originally trained for. Thinks of this as “re-training the ai model”.

If you want an AI to write marketing copy in a very specific brand voice, or to classify customer support tickets into predefined categories with high accuracy based on subtle cues, fine-tuning might be appropriate. For a technical support assistant, you might fine-tune it to follow a specific diagnostic questioning flow, rather than just pulling answers from a knowledge base. It learns how to approach problems.

This a complex process !

• Large Data Requirement: You typically need a substantial amount of high-quality training data (often 3,000 entries or more).
• Time and Cost: It’s a time-consuming and potentially expensive process.
• Uncertain Outcomes: The results might not always meet expectations, especially for general chatbot interactions.
• Not for Knowledge: It’s generally not the best way to teach an AI new facts or information. Embeddings (RAG) are usually better and more efficient for that. If you fine-tune a model on new information, it might “forget” some of its previous general knowledge or skills (this is called catastrophic forgetting).