Interactive tool

A model's memory is frozen. Retrieval lets it open a book.

Everything a model "knows" from training got baked into its weights on a specific day — like a long-term memory that can never be updated without retraining the whole brain. Retrieval-Augmented Generation (RAG) works around that: before answering, the model looks up a live, specific, or private document and reads it as part of the question — open-book, instead of relying on memory alone. See both styles answer the same question below.

01 — Frozen memory

Training ends on a specific day. Everything after that is a blind spot.

A model is trained once on a huge snapshot of text, then "frozen" — its weights don't change again just because time keeps passing. Anything that happened, changed, or was written after that cutoff simply isn't in there.

What's actually frozen

The model's parameters — the billions of learned numbers that encode everything it picked up from training data — stop changing the moment training ends. Ask it anything within that snapshot and it answers from memory, instantly and fluently.

What falls in the gap

Last week's news, a price that changed yesterday, your company's private files, a teammate who started this morning — none of it existed when the model was trained. Memory alone can't know it, and worse, the model often won't tell you that.

02 — Ask the same question, two ways

Memory guesses. Retrieval looks it up first.

Pick a question below, then run both models. The memory-only model answers straight from its frozen training — confidently, even when it's stale or wrong. The RAG-enabled model first fetches a relevant snippet, "reads" it, and only then answers, grounded in what it just retrieved.

Memory only

No lookup
Click "Run memory-only" to see it answer from training alone.

RAG-enabled

Retrieve, then answer
Click "Run with retrieval" to watch it fetch a document first.
Run both columns to compare how each one arrived at its answer.
03 — Tradeoffs

RAG isn't free, and it isn't magic.

Retrieval solves real problems that pure memory can't — but it introduces new ones of its own. Knowing both sides is what separates "we added RAG" from "we added RAG and it actually helped."

What retrieval solves

  • Staleness — it can pull facts from today, not just from the training snapshot.
  • Private or proprietary data — your internal docs were never in training data and never will be, but retrieval can surface them at answer time.
  • Traceability — a good RAG answer can point to the exact passage it used, which pure memory can never do.
  • Narrow or niche facts — long-tail details a general model never memorized well.

What it costs you

  • Garbage in, garbage out — if the retrieved snippet is wrong, outdated, or irrelevant, the model confidently repeats that mistake.
  • Latency — searching and reading a document takes real time before generation even starts.
  • Cost — every retrieved snippet adds tokens the model has to process, on top of running the search itself.
  • Retrieval quality is its own hard problem — a great model fed a bad snippet still produces a bad answer.
04 — Quick check

Does it click?

Five short questions. Pick an answer to see if you're right, then move to the next one.

05 — Key takeaways

The four things worth remembering.

06 — Go deeper

"Just add RAG" is the start of the conversation, not the end.

Which documents get indexed, how they're chunked, how "relevant" is decided, and what happens when retrieval comes back empty or wrong — those design choices determine whether RAG actually fixes staleness or just adds latency to the same old guesses. That's the kind of architecture review we do with teams before they ship.

Related tools

Talk to us about your AI roadmap