Every term behind the WWG AI Dashboard, explained for non-technical readers. The first time an acronym appears, its full name is spelled out. For a print-ready version with diagrams, download the PDF.
↓ Download the illustrated PDF (14 pages)A program trained to understand and produce human language — it reads a question and writes a sensible answer.
Like: a very well-read new employee who knows nothing about our company until we show them our data.
The billions of internal numbers that hold a model’s knowledge, baked into a set of files. “14B” = 14 billion of them.
A model whose files we can download, keep, and run ourselves (Qwen, Phi, Llama) — versus a proprietary one we can only rent access to.
Why we care: we own a permanent copy that keeps working even if the maker disappears.
A version trained to follow instructions and answer helpfully — the kind you actually talk to.
A word-sized chunk of text — the unit models read, write, and bill by. ~1,000 tokens ≈ 750 words.
How much text a model can hold in mind at once (e.g. 32,000 tokens) — “the size of its desk.”
Building a model from scratch on much of the internet — done by tech giants, never by us. We start from a finished model.
Giving a finished model a focused extra course on our examples so a new skill or style sticks. A last resort, not a starting point.
An efficient fine-tuning shortcut that produces a small, portable “adapter” file holding just the new skill.
The small downloadable file from LoRA fine-tuning that clips onto the base model. Moves freely between systems.
Look up relevant info from our own files and hand it to the model before it answers — an “open-book exam.” Keeps answers grounded and current.
Turning a piece of text into a list of numbers that captures its meaning, so the computer can search by meaning, not just keywords.
The list of numbers (vector) and how many values it has (dimensions, e.g. 1,024). More dimensions capture finer nuance.
Finding the stored notes whose meaning is closest to the question. SQL Server can store and search these vectors itself.
Cutting long documents into focused passages before embedding them, so search returns a tight snippet, not a whole file.
Translating a plain-English question into a database query for exact numbers — the counterpart to RAG for structured data.
Using a finished model to answer (constant, light) vs. building or teaching it (occasional, heavy). Most running cost is inference.
The chip that runs AI fast, and its onboard memory the model must fit into — the main hardware constraint and cost driver.
Compressing a model to run ~4× smaller and faster with ~1–3% quality loss — like a high-quality JPEG. GGUF/AWQ are common formats.
A 12–14B model at 4-bit on a single 24 GB GPU — smart enough for manager Q&A, since RAG supplies the facts.
A sealed, portable package that runs identically anywhere. We encrypt it and can lift it from our servers to rented hardware and back.
Running on hardware we own. The simplest guarantee data stays put — now one option among several, not a requirement.
Renting compute from Amazon (AWS, Amazon Web Services), Microsoft (Azure), or Google Cloud.
Running on someone else’s hardware while keeping data secret even from them — encrypted even while being processed. Available on AI GPUs in 2026.
Like: renting an office but working inside a windowless safe only you can open.
We hold the encryption keys, not the host — so even the provider’s staff can’t read our data.
Keeping data in a controlled location/condition — for us, sealed in our encrypted container wherever it runs.
A private, walled-off section of a cloud reserved just for us.
Fully disconnected from the internet, so data physically cannot leak. The strongest posture; available but not strictly required here.
Over-dependence on one supplier. Our open-weight + container + own-keys design deliberately avoids it.