AI vocabulary for choosing a workflow
Choose an AI setup by the jobs its parts actually do.
In this lesson · 06 of 7
All lessons in this track →Saved in this browser. View your learning.
Start with the idea
Choose an AI setup by the jobs its parts actually do.
The model generates output. The app gives you the interface. Other software supplies information and tools. Where one part runs does not tell you where every other part runs.
Retrieval: Finding information and bringing it into the task.
See the example and check your understanding
A made-up teaching example
The lesson’s library assistant searches public book descriptions and drafts a cited answer. Search finds candidates; an optional second ranking step orders them. A laptop app may still send generation to a cloud model.
Ask yourself
Does “the app runs on my laptop” prove that all data stays there?
Show the explanation
No. Trace the model, tools and data separately. Keep the original source records even when conversation summaries get shorter.
Start with these checks
- Separate the model from the app and tools.
- Check where information goes.
- Match citations to actual supporting sources.
Use the different practice case below and check every requirement in its full instructions.
The video explains the main idea. Use the sources, instructions and checks below to complete the practice.
Captions appear here during playback.
AI vocabulary for choosing a workflow
Choose an AI setup by the jobs its parts actually do.
Video playback is unavailable. Read the descriptive transcript below, go directly to practice, or open the video file.
Source-based teaching example · Fictional data · Reviewed 2026-10-03. Microsoft Mark · locally synthesized. On a small screen, use fullscreen for diagram details or read the transcript.
Read the descriptive transcript
0:00 · What you will learn
Choose an AI setup by the jobs its parts actually do.
What appears on screen
Choose an AI setup by the jobs its parts actually do.
0:06 · The idea · 1
The model generates output. The app gives you the interface. Other software supplies information and tools.
What appears on screen
The model generates output. The app gives you the interface. Other software supplies information and tools.
0:17 · The idea · 2
Where one part runs does not tell you where every other part runs.
What appears on screen
Where one part runs does not tell you where every other part runs.
0:23 · One useful term
Retrieval. Finding information and bringing it into the task.
What appears on screen
Finding information and bringing it into the task.
0:31 · A worked example
The lesson’s library assistant searches public book descriptions and drafts a cited answer. Search finds candidates; an optional second ranking step orders them. A laptop app may still send generation to a cloud model.
What appears on screen
The lesson’s library assistant searches public book descriptions and drafts a cited answer. Search finds candidates; an optional second ranking step orders them. A laptop app may still send generation to a cloud model.
0:48 · Your turn to think
Does “the app runs on my laptop” prove that all data stays there? Pause and decide before continuing.
What appears on screen
Does “the app runs on my laptop” prove that all data stays there? Pause the video if you need more time.
1:06 · Why this answer works
No. Trace the model, tools and data separately. Keep the original source records even when conversation summaries get shorter.
What appears on screen
No. Trace the model, tools and data separately. Keep the original source records even when conversation summaries get shorter.
1:17 · Check 1
Separate the model from the app and tools.
What appears on screen
Separate the model from the app and tools.
1:23 · Check 2
Check where information goes.
What appears on screen
Check where information goes.
1:28 · Check 3
Match citations to actual supporting sources.
What appears on screen
Match citations to actual supporting sources.
1:34 · Now try a different case
Open the lesson and try its different practice case. Use every check in the full task. Watching alone does not mark the lesson complete.
What appears on screen
Open the lesson Try the new case Check and revise The full task and its requirements are in the linked lesson.
Watching explains the example. Complete the lesson practice to check what you can do. Explore all visual lessons →
Before you start
Prerequisite: separate an answer from evidence. You will explain system components and choose capabilities for supplied tasks. The glossary and paper suffice; no download, paid API or account is required.
What you will be able to do
- Separate model, product, harness and execution environment.
- Distinguish local/cloud, licensing and capability axes.
- Match text, reasoning, multimodal, embedding and reranking capabilities to tasks.
- Explain history, product memory, retrieval and compaction separately.
Four layers and three independent axes
A model is a trained computation mapping inputs to outputs. Training adjusts parameters; inference runs it. A product is the user-facing application, account and settings. A harness manages instructions, context, tool orchestration, approvals, retries and state. The execution environment is the machine/sandbox where tools run, with filesystem, network and permission boundaries. An agent can choose steps and tools within this harness; it has no automatic decision authority.
Local/cloud describes where each component runs; weight availability and license rights describe separate access/reuse conditions; frontier/smaller or specialized describes a capability class with no universal cutoff. These are distinct axes. A local desktop product can call a cloud model. Open weights do not automatically mean open-source software: license, training information and reuse conditions differ. Downloadable weights do not prove approval for sensitive data, particularly when logs and connectors exist.
Capability and context glossary
| Term | Practical meaning |
|---|---|
| Text generation | Drafts or transforms text; factual fidelity needs checks. |
| Reasoning | Supported computation/settings for harder tasks; evaluate actual performance. |
| Multimodal | Multiple data types; image input does not imply image generation or audio output. |
| Embedding | Content converted to vectors for similarity search; it does not draft an answer. |
| Reranking | Reorders query candidates; relevance does not prove truth. |
| Retrieval/RAG | Fetches evidence for generation; sources can be missing or stale. |
| Tool/function call | Proposed external operation executed by application under permissions. |
| Token/context window | Input/output units and working capacity; word ratios vary by language and modality. |
| Conversation history | Stored turns a product may reintroduce; stored does not mean fully loaded. |
| Product memory | Persisted facts/preferences with product-specific settings and scope. |
| Compaction | Compressing context to continue work; details can be lost. Preserve source records. |
| Fine-tuning | Further training for measured behavioral needs; not live document retrieval. |
| Latency/evaluation | Response time, and testing against declared expected behavior. |
| Model card/guardrails | Capability documentation and system controls; neither grants authorization. |
| API | A software interface; calling an API may have separate billing and permissions. |
| MoE/parameters | Architecture and learned values; parameter count alone does not determine useful quality. |
Check exact model and product for modalities, tools, reasoning settings and limits. No model is universally best. Task complexity, evidence access, data permissions, quality, latency and approved resource limits guide choice.
Worked example and guided choice
A synthetic library assistant has 100 public book descriptions. Embeddings find similar items; a reranker prioritizes candidates; a text model drafts a cited answer. The harness retains source IDs and blocks purchases. A laptop tool may read the catalog while generation occurs remotely. Local app therefore does not prove all data stays local.
Choose a capability: (a) three-sentence draft from facts; (b) compare a chart image to a caption; (c) find matching descriptions; (d) solve a constrained schedule. Answers: text generation; supported image input; search/embeddings with optional reranking; reasoning or deterministic scheduling. None needs a purchase tool.
Substantive review: 2026-10-02. Official OpenAI model documentation lists supported capabilities; product availability requires a separate check.
Interactive practice / Fictional fixture
Write your restart record here
Use the v1 and v2 records and the generated summary above. Settle four facts, then write the restart record. Draw the layer map on paper. Our answer unlocks after you have written yours and run the checks.
Everything runs in this browser. No AI model is called and no work is uploaded. Bounded checks cover specific facts and rules. Your writing is not graded: you review it yourself. Practice never marks a lesson complete automatically.
Objective, v1/v2 source references, proposal status, unknowns and a proposed next check. Say what you rejected from the summary.
Try the task, run the checks, then revise. Checks are teaching feedback, not an instructor grade.
Observed browser simulation
This trace comes from the bounded local checker. It is not a Python run, model result or production tool execution.
Self-review
Locked. Write every written part, then run the checks. A few words is not an attempt.
This is self-review, not grading. Read our answers, then tick what your own work does. Ticks are optional. They stay on this page and are not saved or sent anywhere. A different answer can be just as good if it meets the lesson criteria.
Your restart record
A strong answer
Objective: restart the Riley pilot proposal from its source records. Sources: v1 (archived) proposed four weeks; v2 (authoritative in this fixture) changes this to two weeks. Status: proposed; the pilot has not started. Unknowns: savings. Allowed: drafting only. Proposed next check: compare any new summary against v2 before use. Rejected from the generated summary: the success claim, the savings figure and the sending step. Originals are kept outside this summary.
A flawed answer
four-week pilot succeeded; savings 40%; send the announcement.
What is wrong with it: Reject the summary's invented success/savings/sending. Four weeks is the archived v1 figure; v2 says two weeks, the pilot has not started, savings are unknown and only drafting is allowed. Authority is explicitly assigned to v2 in this fixture; a timestamp alone would not establish authority in real work.
Keep or restore your work
Drafts are not saved automatically. Save only fictional practice data. A browser draft may be lost if storage is cleared; download a copy to keep it.