Catalogued glossary terms across document sections and organized by source part

"Enumerate" is exactly the right word. Here's the full list, grouped by canonical source page — the page where each term actually gets its deepest explanation, per the source-of-truth rule.

Hardware Terms — Part 2

    GPU (built)
    TPU
    VRAM
    Interconnect
    Cluster

Model Terms — Part 4
6. Tokenization (canonical here — has the worked examples, even though it's also step 2 of the Part 3/5 pipelines)
7. Transformer
8. Parameter
9. Context Window
10. Self-Attention (built — this is also "Forward Pass," step 4 of the Part 5 pipeline)

Training Terms — Part 3
11. RLHF (dedicated section already exists)
12. Pre-training / Fine-tuning — leaning toward folding these into a single Training Pipeline entry (the whole 8-step diagram) rather than two separate terms, since Part 3 explains them at pipeline-step depth, not standalone-concept depth. Your call.

Inference Terms — Part 5
13. Temperature (this is "Token Sampling," step 5 of the pipeline)
14. Batching
15. KV Cache
16. Autoregressive Generation

Interface Terms — Part 6
17. System Prompt
18. Prompt Engineering

Application Terms — Part 7
19. RAG (Retrieval-Augmented Generation)

That's 19 terms total — 2 built, 17 to go. A few worth a decision before building starts:

    #12 — one combined "Training Pipeline" entry (whole diagram) vs. two separate terms? I'd lean combined, matching how the pipeline steps discussion went for Part 5.
    #10 — "Self-Attention" and "Forward Pass" are the same underlying concept from two different pipeline steps (Part 4 vs. Part 5). One entry, cross-referenced from both places, rather than two entries saying the same thing?

If those two calls make sense, that trims it to 17 real entries to build out. Want me to start building, or lock the list first?