I asked an AI to tweak one image. It came back with a completely different picture. I went and figured out why — and it is more interesting than "the AI got confused".

It turns out that platforms like ChatGPT, Claude and Perplexity do not have real memory. Every time you send a message, the entire conversation gets resent to the model — from scratch. It rereads everything, every single turn, before it replies. That is also why the token count keeps climbing the longer a chat runs.

Here is the part that actually surprised me. As a conversation grows, the model does not treat every part of it equally. There is a documented behaviour called "lost in the middle" — it pays the most attention to the very start and to whatever is most recent, and quietly deprioritises everything in between.

Nothing gets deleted. It is all still there. The model just stops weighing it as strongly. So an instruction from message 4 can be functionally forgotten by message 20 — even though it is technically still sitting in the chat.

Nothing gets deleted. The model just stops weighing it as strongly.

Asking for a "small tweak" is not like Photoshop. There is no file being edited. The model rereads the whole chat, writes a brand-new prompt based on it, and hands that to a completely separate image model — which paints a new image from scratch. Every image is a fresh repaint. Guessed. Every single time.

I found this out the hard way making a 3-slide LinkedIn carousel: slide 1 came out perfect, slide 2 had content from slide 1 leaking in, and slide 3 drifted to a completely different theme. Same conversation. Same intent. Three very different results.

The interactive walkthrough at the top of this page lets you watch all of this unfold step by step — from the growing transcript, through the "lost in the middle" effect, to the 3-slide case study. Play with it, then come back here.

The fix is not a better model — it is understanding how the context window actually works, and restating your key instructions instead of relying on the model to "remember" them.

Context window — interactive walkthrough

I asked an AI to tweak one image and got back a completely different picture. I went and figured out why — and it is more interesting than "the AI got confused".

I asked AI to tweak one image. It repainted the whole thing.

Everyone tunes the model and the prompt. Far fewer people tune the chunking — and yet it often decides more about answer quality than the model choice does.

A chunk is the unit your retriever can return. Too large and you drown the model in irrelevant context; too small and you cut ideas in half so no single chunk is actually useful.

Retrieval quality is a data problem before it is a model problem.

from langchain.text_splitter import RecursiveCharacterTextSplitter

splitter = RecursiveCharacterTextSplitter(
    chunk_size=700,
    chunk_overlap=100,
    separators=["\n\n", "\n", ". ", " "],
)
chunks = splitter.split_documents(docs)

This is a starting point, not a law. Measure retrieval hit-rate on your own questions and adjust — the right numbers depend on your documents.

How chunk size, overlap and semantic splitting quietly decide whether your RAG chatbot answers well — and the settings I actually ship.

Research.

I asked AI to tweak one image. It repainted the whole thing.

Chunking strategies for production RAG