Glossary

One plain sentence per term. The last column is the week that teaches it properly; until then, this sentence is enough.

Term	What it means	Week
Abstention	The system declining to answer because nothing it retrieved is good enough, instead of guessing. Also called declining.	3
Agent	A system where the model chooses its own next step, such as which tool to call, rather than following steps you wrote.	1, 4
Agent loop	The cycle of model decides, tool runs, result goes back to the model, repeated until done or stopped by a step or cost limit.	4
Approval	A person confirming an action before it happens. Enforced in code: the tool returns "needs approval" and the loop stops.	4
Blast radius	The worst thing that can happen if a defence fails, given what the system is able to do.	5
Chunk, chunking	Cutting documents into pieces small enough to search and to show the model. How you cut changes what is found.	2
Citation	The document and section an answer came from, so the user can check it.	3
Context, context window	Everything the model is shown on one request. The window is the most it can take; filling it is not the same as helping it.	2
Context engineering	Deciding exactly what the model is shown on each request, and what to leave out.	2
Cost attribution	Knowing what each step of a request cost, so you can say what a ticket costs and where the money goes.	1, 5
Decision record	Your one-page weekly note: the decision, the options, the numbers, what you rejected and the risk you accepted.	All
Decline rate	How often the system declines. Measured separately on questions that have no answer and on questions that do.	3
Embedding, dense search	Turning text into numbers that capture meaning, then finding the pieces closest in meaning to the question.	2
Evaluation, test set	A fixed set of questions with known answers, run on every change, so you know whether a change helped or hurt.	3
Faithfulness	Whether every claim in an answer is supported by the text the system retrieved.	3
Function calling, tool schema	The model asking for a tool to be run by returning structured arguments. The schema describes the tool and its arguments, and the description is effectively a prompt.	1
Grounding	Making the model answer only from the text you gave it, not from what it learned in training.	3
Hidden test set	Questions the programme holds and you never see. It reports totals on your pull request, so you cannot tune to it.	3
Hybrid search	Combining keyword search with meaning-based search, because each misses what the other finds. Reciprocal rank fusion merges the two lists.	2
Idempotency	Safe to repeat: the same request twice does not do the work, or charge you, twice.	1
Judge	A model used to grade answers, for example for faithfulness. Trusted only after you check how often it agrees with your own labels.	3
Keyword search, BM25	Finding pieces that share the question's words. Strong on product codes, error numbers and setting names.	2
Least agency	Giving the system only the abilities it needs, so a successful attack can do little.	4, 5
MCP	Model Context Protocol: a standard way to serve tools to a model. It carries the calls; it does not decide what is allowed.	4
Metadata filtering	Searching only inside pieces with a given label, such as version 3. The fix for the version trap.	2
Observability, tracing, span	Recording what happened on each request so it can be explained afterwards. A span is one timed, costed step in that record.	5
Personal-data masking	Replacing names, emails and phone numbers before text reaches the model, and restoring them only in approved output.	5
Precision, recall, hit rate, MRR	Four ways of scoring whether retrieval found the right pieces. See worked example A.	2
Prompt injection	Instructions hidden in content the system reads, a document or a customer's ticket, aimed at making it do something it should not.	5
Regression gate	The check that blocks a change from merging if it makes the scores worse.	3
Retrieval	Finding the few relevant pieces of your documents and showing them to the model with the question.	2
Routing table	Your week 4 decision, per task type: plain code, a fixed workflow or an agent, and a cheap or frontier model, with the numbers behind it.	4
Structured output	Asking the model for data in a fixed shape and validating it, here with Pydantic, instead of trusting free text.	1
Trajectory	The path an agent took: which tools, in which order, how many steps, and whether it acted without approval.	3, 4
Version trap	The version 2 and version 3 pages read almost the same, so similarity alone finds the wrong one and the answer is confidently wrong.	2