LLM Cheat Sheet
The vocabulary you need before the rest of this track makes sense — one page, no fluff.
Beginner
TopNotchNote
Notes that matter
core vocabulary
| Term | Means |
|---|---|
| Token | The unit a model actually reads/writes — usually a sub-word chunk, not a whole word. See Tokenization. |
| Context window | The maximum number of tokens (input + output combined) a model can attend to in one call. See Context Windows. |
| Temperature | Sampling randomness at generation time. 0 = deterministic/greedy; higher values flatten the next-token probability distribution. |
| System prompt | The instruction set the caller (not the end user) supplies to steer behavior before the conversation starts. |
| Embedding | A fixed-length vector representing meaning, used for similarity search. See Embeddings. |
| RAG | Retrieval-Augmented Generation — fetch relevant text at query time and put it in the prompt instead of relying on parametric memory. See RAG Architecture. |
| Fine-tuning | Updating a model's weights on task-specific data, as opposed to steering it purely through the prompt. See Fine-Tuning vs. RAG. |
| Hallucination | A fluent, confident output that is factually wrong or unsupported by any source the model was given. |
| Quantization | Storing model weights in lower precision (e.g. INT8/INT4 instead of FP16) to cut memory and increase throughput. See Quantization. |
| KV cache | Cached attention keys/values from prior tokens so generation doesn't recompute them every step. See KV Cache & Continuous Batching. |
the three ways to change model behavior
| Lever | Changes weights? | Latency to try | Best for |
|---|---|---|---|
| Prompting | No | Seconds | Format, tone, one-off tasks, quick iteration |
| RAG | No | Minutes-hours to build | Grounding on facts the model wasn't trained on, or that change often |
| Fine-tuning | Yes | Hours-days | Consistent style/format at scale, teaching a narrow skill, cutting per-call prompt length |
Full decision framework on Fine-Tuning vs. RAG vs. Prompting — in practice most production systems combine all three rather than picking one.
reading a model's spec sheet
# the numbers vendors publish, and what they actually bound
context_window = 200_000 # tokens; input + output share this budget
max_output_tokens = 8_192 # a separate, usually smaller cap within the window
training_cutoff = "2026-01" # the model has no knowledge of events after thisA large context window doesn't mean the model uses all of it equally well — see the "lost in the middle" effect covered in Context Windows before assuming you can just paste in everything.
where to go from here
The Transformer Architecture — what's actually running under these terms.
Prompt Engineering Fundamentals — the first lever, in practice.
RAG Architecture — the second lever, end to end.
Prompt Engineering Fundamentals — the first lever, in practice.
RAG Architecture — the second lever, end to end.