Making your own Chatbot

How easy is it to make your own chatbot using a pre-determined set of training resources?

The answer probably ranges from very easily and cheaply to, with some effort and at great cost of manhours and subscriptions.

Larry Ferlazzo created five using ChatGPT. When I asked him how, this was his response:

It was easy with ChatGPT. You can create your own custom chatbot. I just told each one to use the list of resources and experts I gave it as its knowledge base, and nothing else. I had to repeat those instructions a few times, but they seem to work fine.

An alternative, to provide support for a specific course, is to build from scratch as follows:

“The AI Companion will be based on a Retrieval-Augmented Generation (RAG) architecture, ensuring pedagogical grounding and minimisation of hallucinations. It will include the following elements:
A Learner Interface: Learners will interact through a simple chatbot window integrated into the e-learning platform. They will be able to:

• Ask natural language questions
• Interact with course modules
• Request clarifications
• Explore digital entrepreneurship-related topics
  The interface will be accessible, intuitive, and designed specifically for users of low digital-literacy (NEETs).
  An Entrepreneurial Knowledge Base: To ensure accuracy and alignment with course objectives, the AI Companion will rely on a curated and project-controlled knowledge base instead of unrestricted internet queries.”

I don’t yet know how different these two alternatives are in reality, especially in terms of quality of response and interaction. Asking Co-Pilot resulted in the following table:

Feature	General GPT	Larry’s Chatbots	RAG Companion
Knowledge source	Entire model	Entire model + embedded resource list	Curated external knowledge base
Retrieval	None	None	Yes (vector search)
Hallucination control	Low	Medium	High
Updateability	None	Manual prompt edits	Continuous KB updates
Pedagogical alignment	None	Light	Strong
Integration with platform	None	None	Full (modules, progress, scaffolding)
Accessibility for low‑literacy learners	None	None	Designed for this
EU compliance (data control, transparency)	Weak	Weak–medium	Strong
Development cost	None	Very low	Medium–high
Maintenance cost	None	Low	Medium
Suitability as a learning companion	Low	Medium	High

There are intermediate possibilities between the Ferlazzo prompt model and the full RAG implentation so how do they stack up?

Approach	Cost	Technical complexity	Hallucination control	Pedagogical alignment	Suitable for NEET learners?
Ferlazzo-style prompt	Very low	Very low	Low–medium	Low	No
Chunked KB in prompt	Low	Low	Medium	Medium	Maybe
Lookup-table “toolformer”	Low–medium	Medium	Medium–high	Medium–high	Yes, with UI work
Lightweight RAG	Medium	Medium	High	High	Yes
Full RAG	Medium–high	High	Very high	Very high	Yes

So RAG, creating your own carefully curated trove of materials, is obviously a good idea but even here there are options for strength of implementation.