You train a chatbot on your own documents by grounding it in them rather than retraining the model: upload your files to a tool's project or knowledge feature so the chatbot retrieves relevant passages at answer time (this is called RAG, or retrieval-augmented generation), or use an API connection to a vector store if you're building your own system.
Fine-tuning is the other path, but it teaches style and format, not facts — so for most beginners with a pile of PDFs, retrieval is the right first move. The mechanism matters because it explains why the two approaches behave so differently. When you upload documents to a chat tool, the system splits each file into small chunks, converts every chunk into a list of numbers called an embedding, and stores those numbers in a searchable index.
When you ask a question, your question is embedded too, and the system pulls back the chunks whose numbers sit closest to it. Those chunks get pasted into the prompt as context, and the model writes an answer from them. Nothing about the model's weights changes.
That's why retrieval is fast to set up and easy to undo: delete the file, and the knowledge is gone. Fine-tuning works the opposite way. You feed the provider many examples of input-and-ideal-output pairs, and it adjusts the model's internal weights.
The model never sees your source documents at answer time — it has absorbed a pattern. That makes fine-tuning good at teaching a consistent tone, a fixed output structure, or a specialised format, and bad at storing facts, because facts change and retraining is slow and costly.
Here's a concrete worked example. Say you run a small accounting firm and want a chatbot that answers client questions about your own service agreements. You create a project in a chat tool — ChatGPT's Projects and Claude's Projects both accept uploaded files, and both appear in our AI tool database as chat-category tools with free tiers and paid plans (ChatGPT Plus at $20 per month, Claude Pro at $17 per month billed annually or $20 monthly).
You upload twelve PDFs of engagement letters and internal process notes. You ask, "What's our turnaround time for quarterly filings?" The chatbot retrieves the two chunks that mention turnaround times and answers from them.
If instead you wanted every reply to follow your firm's exact house style — a specific greeting, a fixed sign-off, no bullet points — that's a fine-tuning job, not a retrieval one. A useful decision rule: if the thing you want to change is a fact, use retrieval; if it's a behaviour, use fine-tuning; if it's both, do retrieval first and add fine-tuning later.
A few practical details separate a working setup from a frustrating one. Chunk size matters more than people expect: chunks that are too small lose the surrounding context, and chunks that are too large dilute the match with irrelevant text, so a document with clear headings and short sections retrieves better than one giant wall of prose.
File format counts too — clean text and well-structured PDFs index far more reliably than scanned images, because the system has to read the words before it can chunk them. If you're building your own pipeline rather than using a chat tool's upload button, you'll connect an API model to a vector database, and that's where the cost picture shifts: according to our AI tool database, DeepSeek's V4 Pro API runs at $0.435 per million input tokens and $0.87 per million output tokens, which is roughly twelve times cheaper than GPT-5.5, and long documents eat a lot of tokens.
That matters because retrieval sends several chunks with every question, so your token bill grows with how much context you attach.
The honest limits are worth stating plainly. Retrieval fails when your documents contradict each other — the chatbot will pick whichever chunk scores highest and sound confident doing it, even if the older policy says something different. It also fails on questions that require joining facts across many documents, because each answer only sees the handful of chunks that were retrieved.
Fine-tuning fails in a different way: it can make a model sound authoritative about facts it learned during training that are now out of date, and you can't easily correct a single wrong fact without retraining. Neither approach gives you a guarantee of accuracy, and neither is free at scale — retrieval costs tokens on every question, and fine-tuning costs money and time up front.
Pricing and plan limits for these tools change often, so check the vendor's own page before you commit. If your real problem is that the chatbot keeps skipping steps or drifting off your outline, that's a prompting issue rather than a grounding one, and the fix lives in a different workflow.