Fine-tuning or RAG for my documents?
The question worth answering before preparing a dataset. Both outputs are prepared here.
About this tool
Many people searching for "train a model on my PDFs" actually want RAG, and would skip the training entirely. It is worth separating before spending time.
Short rule: if you want the model to **know** something specific from your documents and cite it properly, RAG. You retrieve the passage at the time and hand it over; it updates by changing a file, and the model can say where the answer came from. If you want it to **write differently** — a fixed format, a tone, a report structure — fine-tuning.
The uncomfortable part: fine-tuning is not a reliable way to insert facts. A model learns the style of the material long before its content, and when a fact is missing it invents one in the right tone, which is worse than not knowing.
Both are prepared here. For RAG, PDF to Markdown with chunking at headings. For fine-tuning, this tool.
How it works
- 1
Decide whether you want the model to know something or to write differently.
- 2
For RAG, use PDF to Markdown and its chunks.
- 3
For fine-tuning, drop the PDFs here and pick the format.
Questions
- Can I do both?
- Yes, and that is the usual production setup: fine-tuning for the answer format and RAG for the facts. Both sets come from the same documents.
- Does RAG need anonymization too?
- If the index or the model is in the cloud, yes. If everything runs on your machine with a local model, it is optional; anonymization can be switched off, or the real data restored into the answer afterwards.
- What if my documents change every week?
- A strong argument for RAG: reindexing takes minutes, while retraining has to be redone in full.