Two ways to teach a bot your business

A general model knows a lot about the world and nothing about your shop. Teams usually pick one of these:

• Fine-tuning — keep training the model on your own examples, so the facts are baked into it
• A knowledge base (RAG) — keep the facts in documents, search them when someone asks, and paste the matching bits into the question

RAG is the short name for retrieval-augmented generation. In plain terms: look it up, then answer.

What fine-tuning does

You take a model that already works and train it further on your material. Afterwards it "just knows" your business. That sounds tidy. It gets messy as soon as something changes.

• It costs real money and time. You prepare training examples, run the job, and check that it did not get worse.
• It goes stale. Change a price on Tuesday and the model is wrong until you train it again.
• It can still invent details. Training teaches patterns. It does not guarantee the model will quote your returns policy word for word.
• Your text ends up inside the model. That copy lives with whoever trained it, not as a document you can open and edit.

What a knowledge base does

The documents stay as documents. When someone asks a question, ChatNexus:

1. Turns the question into a numeric fingerprint (an embedding)
2. Finds the chunks whose fingerprints are closest
3. Sends those chunks to the model along with the question
4. The model writes the reply from that text

The model itself does not change. If the return window moves from 14 days to 30, you update the document. If that document is a Google Doc, you can sync it straight into the knowledge base so the bot picks up the edit on its own.

Side by side

Question	Fine-tuning	Knowledge base
Setup	Prepare examples and pay to train	Add the documents you already have
When prices change	Train again	Edit the source and let it re-index
Exact facts	It remembers the gist	It can quote the chunk it found
Made-up answers	Easier, because nothing is looked up	Harder, because the reply is tied to the text you supplied
Where the text lives	Inside the trained model	In the knowledge base, as documents you can edit or delete
Changing models	You are stuck with the one you trained	Switch models and keep the same documents

A lookup adds a short search before the reply starts. For a support bot, that trade is usually worth it.

How ChatNexus does the lookup

You do not set up your own vector database. On the agent's knowledge base you can:

• Paste text
• Upload a .txt, .md, or PDF
• Paste a normal web page URL and crawl it
• Connect a Google Doc or Google Sheet (share it with the address shown in the dashboard first)

Long documents are split into chunks of about 1,500 characters, with a small overlap so a sentence is not cut in half. When someone asks a question, ChatNexus pulls a few matching chunks (up to six, and no more than two from the same source) plus the nearby chunks from that same document. That stops one long file from crowding out a shorter page that is actually the right answer.

A pasted note or an uploaded file is ready after it finishes processing. A Google Doc or Sheet with auto-sync turned on is checked about once an hour. You can also sync that file yourself from the knowledge base page. Google links do not work in the normal URL crawl. Use the Google Doc tab for those.

The knowledge base is stored by ChatNexus, not baked into a custom model. You can edit or delete individual chunks. If a better model shows up, you can switch the agent to it and keep the same documents. ChatNexus models come from Groq, OpenAI, and xAI.

When retraining can still make sense

Fine-tuning is a poor fit for prices, policies, and product facts. It can help when you need a very particular writing style that a system instruction cannot hold, or when the model keeps misunderstanding specialist terms no matter how you write the documents. ChatNexus does not fine-tune models for you. Tone on ChatNexus is set in the system instruction, and the facts belong in the knowledge base.