How to do simple AI text analysis? LLM, fine-tuning, BERT, or embeddings?

Few-shot with 100 examples helps a little. Few-shot with 1000 examples actually confuses small models and stretches context so much that the result stops making sense.

That is why few-shot 1000 is practically nonsense: dramatically more expensive and worse. Input-token caching helps for few-shot scenarios, but it does not change the rules.

Sequential processing comes out at roughly 0.64 to 1.06 seconds per sample. The 1000-example variant is completely off, three seconds per sample with the mini model.

No training, no pipeline, no specialized model. I change the prompt and get different behavior. One endpoint, one library, one runtime model.

Fine-tuning helped dramatically. GPT-4.1-mini reached 80.0%. Interestingly, few-shot stops helping and actually hurts. My guess is that 100 randomly picked examples are not representative against the 25,000 used in training.

Fine-tuning itself cost $40. In Azure OpenAI Service the fine-tuned model uses the same token prices as the regular version, but hosting is billed at $1.7/h. With OpenAI you do not pay hosting, but tokens are more expensive. Sometimes one is cheaper, sometimes the other.

I tested in Azure where you pay hosting. This should not be taken as an official guarantee, but I measured 2x to 3x lower latency than the regular variant.

It is a specialized model, but still generative. Small instruction tweaks or in-context examples work. Bigger changes mean another fine-tuning run.

A simple and briefly trained variant ended up clearly above the plain LLMs and similar to logistic regression on embeddings. I believe BERT could be pushed further.

We need a GPU machine. I counted two older NVIDIA T4 nodes on Azure VM at $0.56/h. Training ran for 500 seconds, so almost nothing. But inferencing costs compute and that brings BERT into GPT-4.1-nano territory.

BERT ran locally on Azure VM, so it did not have the same cloud round-trip as the other variants. Even if 0.09 s became 0.12 s in a more realistic setup, it is still excellent.

You need code or AutoML; the model is single-purpose and you cannot tune it via a prompt. Plus I started from BERT, which does not speak Czech, so Czech and multilingual scenarios are harder.

Training on the full dataset puts the result clearly above the plain LLM and close to BERT.

The cost of logistic regression is negligible. What matters is the embedding token price and that is very low.

Embeddings come back from the cloud quickly and the local LR model is instant.

Like BERT, the final classifier is fixed. Unlike BERT, modern embeddings typically speak Czech and other languages, which is a big practical advantage.