Context engineering basics

Question 1

A handbook is on disk, and one paragraph from it is inserted. How much of the rest of the handbook was in the window?

Accepted Answer

None of it. Text the model cannot see was never in the window

Question 2

A speaker says the model was told a rule. What must be true for that rule to be in the prompt?

Accepted Answer

The rule's words were included in the text the model sees

Question 3

A call contains one question and no retrieved pages. Which piece is the prompt?

Accepted Answer

The question text the model sees

Question 4

A neighbor uses the question's terms but states the opposite rule. Why can it still rank first?

Accepted Answer

The embedding matched similar text, and similar text can be wrong

Question 5

After a prompt-cache hit, what is the model reading at the front of the call?

Accepted Answer

The same prefix text that was cached

Question 6

A system line, a user question, and a worked example are all sent in one call. What do they have in common?

Accepted Answer

Each one is text the model sees

Question 7

A correct policy uses different words from the question, and a mistaken note uses the question's words. Which outcome can embedding retrieval produce?

Accepted Answer

It can return the mistaken note as a neighbor and miss the correct policy

Question 8

The inserted quote stops before a sentence that reverses the claim. The citation looks formally complete. What limits it?

Accepted Answer

It is only as good as the quote that was inserted, and that quote omits the reversal

Question 9

Truncation and a summary can both make an input shorter. What is particular to truncation?

Accepted Answer

It drops tokens, usually from a chosen end, instead of writing a replacement paragraph

Question 10

A summary says "follow the usual process," and the source listed three exceptions. What can the model read about those exceptions?

Accepted Answer

Only what the summary kept, and the summary can have dropped those constraints

Question 11

What is a retrieval score, relative to whether a chunk is true?

Accepted Answer

A rank for which text to insert, not a proof the text is true

Question 12

In embedding retrieval, what is a neighbor?

Accepted Answer

A chunk close in the vector space, not a chunk shown to be true

Question 13

Two citations appear, and only one source's paragraph was inserted. Which source text is in the window?

Accepted Answer

The paragraph that was inserted; the other citation has no matching text

Question 14

A constraint matters, and the summary is what will be inserted. When does the model see that constraint?

Accepted Answer

When the summary text includes it, because the summary replaces the source

Question 15

Candidates are filtered by date before any of them are pasted into the prompt. When does that filter run?

Accepted Answer

Before the model sees the set

Question 16

A tool returns a value, and that value is not written into the prompt. What did the model see of the value?

Accepted Answer

Nothing. What the model cannot see was never in the window

Question 17

The first thousand tokens are identical on every call, and the rest changes. What can a prompt cache hold from that pattern?

Accepted Answer

The repeated prefix, not the part that changes

Question 18

A grader marks the model wrong for missing a clause that was never pasted in. What is fair to say about that clause?

Accepted Answer

It was never in the window, so the model could not see it

Question 19

An old system line is already in the processed tokens, and a new system line is appended after it. What is true of the KV cache?

Accepted Answer

It can hold attention state for both lines, and it does not by itself refresh the old line

Question 20

A "must" clause and an "unless" clause end up in different chunks. Only the "must" chunk is inserted. What should you expect?

Accepted Answer

The model sees the requirement without the exception the boundary cut off

Question 21

A table's cells are never written into the prompt. What does the model see of those cells on this call?

Accepted Answer

Nothing from those cells, because what the model cannot see was never in the window

Question 22

A summary uses fewer tokens than the source it replaces. What was saved, and what can be lost?

Accepted Answer

Tokens were saved, and a constraint from the source can be lost

Question 23

A value that changes on every call is placed at the start of the prompt. Why does that work against a prompt cache?

Accepted Answer

The cache stores a repeated prefix, and a changing start is not a repeated prefix

Question 24

Which claim matches text that was never placed in the context?

Accepted Answer

What the model cannot see was never in the window

Question 25

A metadata filter requires the topic to be billing. A shipping note is in the index. Where is that note during the call?

Accepted Answer

Outside the set the model sees, because the filter ran first

Question 26

Policy text in the window is out of date, and the KV cache for those tokens is still in memory. What actually changes the policy the model read?

Accepted Answer

New text processed in place of the old instruction, not the KV cache by itself

Question 27

More chunks are inserted for one question. What increases for certain?

Accepted Answer

The amount of inserted text, not the share of chunks that are true

Question 28

An exception sits in the tail of a passage, and the caller truncates that tail. What remains of the exception in the window?

Accepted Answer

Nothing, because those tokens were dropped from the chosen end

Question 29

An index finds three nearby chunks and those chunks are inserted. What did the embedding step contribute?

Accepted Answer

Neighbor text chosen by similarity, which may still be wrong

Question 30

What does an embedding search supply for a query when the results are used as context?

Accepted Answer

Neighbor texts whose vectors sit near the query

Question 31

Why reuse a KV cache while the next token is produced?

Accepted Answer

Attention state for tokens already processed can be kept instead of recomputed

Question 32

Which description fits retrieval into a context window?

Accepted Answer

It pastes chosen text into the window and does not certify that text

Question 33

A sentence is left out of the request. What is true of that sentence on this call?

Accepted Answer

The model cannot see it, because it was never in the window

Question 34

Which order matches a retrieval call that uses a filter?

Accepted Answer

Filter the candidates, then insert what remains for the model to see

Question 35

The source says "send the report, but never include account numbers." The summary says "send the report." What happened?

Accepted Answer

The summary replaced the source and dropped the constraint about account numbers

Question 36

Two chunks disagree, and both are inserted. What has retrieval established?

Accepted Answer

That both texts were inserted, not which one is true

Question 37

Where can a length-based splitter place a boundary?

Accepted Answer

At a length that falls inside a rule, not only after a finished rule

Question 38

Why does rewriting an instruction change what the model receives?

Accepted Answer

The instruction is text the model sees, so new wording is new text

Question 39

A writer leaves 200 words of room under a token limit. Why is that room uncertain?

Accepted Answer

Those words may be more tokens than the room that is left

Question 40

A constraint lives only in a document this call does not include. Why is that constraint absent from what the window shows?

Accepted Answer

What the model cannot see was never in the window, and the document was not included

Question 41

A permission filter drops three private notes. What does the model see of those notes?

Accepted Answer

Nothing from them, because the filter ran before the model sees the set

Question 42

In a context window, what is a prompt?

Accepted Answer

Text that is placed where the model reads it

Question 43

The KV cache is cleared, and the same stale instruction is sent again. What does the new cache contain once that text is processed?

Accepted Answer

Attention state for that same stale text

Question 44

Why describe ordinary truncation as dropping tokens from a chosen end?

Accepted Answer

The caller selects an end, such as the start or the finish, and those tokens are dropped

Question 45

Three solved examples appear before a new question. How do those examples enter the call?

Accepted Answer

As text the model sees, in the same way the question does

Question 46

A rule says "approve refunds under $20, except for opened goods." The cut between chunks falls between those clauses. What can go wrong?

Accepted Answer

The boundary can split the rule, so one chunk lacks the exception

Question 47

A safety filter rejects a chunk, and a later step pastes the chunks that remain. What happens to the rejected chunk?

Accepted Answer

It is out before the model sees the set, so it is not pasted

Question 48

An answer ends with a source title. The page body was not inserted. How far does that citation go?

Accepted Answer

Only as far as the inserted text, which here is the title rather than the page

Question 49

Inside the window, what is a citation string by itself?

Accepted Answer

More text, only as useful as the source text inserted with it

Question 50

Why is "nearest neighbor" a weak stand-in for "trusted source"?

Accepted Answer

The nearest chunk is the most similar one, and that neighbor can be wrong

Question 51

The first line of a long instruction is edited, and the rest is unchanged. What happens to a cache of that prefix?

Accepted Answer

The repeated prefix no longer matches, because the stored prefix included the old first line

Question 52

What does a prompt cache store?

Accepted Answer

A repeated prefix of the prompt, so that same leading text can be reused

Question 53

A stale instruction has already been processed into the KV cache. What does the cache do to that instruction on its own?

Accepted Answer

It keeps attention state for the old tokens and does not refresh the instruction

Question 54

The inserted chunk says "always share the file." The next chunk, which was not inserted, says "never share it outside the team." What did the boundary do?

Accepted Answer

It split the rule, so the inserted half leaves out the limit

Question 55

Which unit decides whether more text still fits in the window?

Accepted Answer

The token budget, not a word count

Question 56

A language filter keeps only English chunks. A relevant note is in another language and does not appear. Why is it absent?

Accepted Answer

The filter removed it before the model sees the set

Question 57

The oldest instructions are cut so a new question fits. Can the model use those cut instructions on this call?

Accepted Answer

No. Truncation dropped those tokens from the chosen end

Question 58

Two notes use the same product name, and one note is a rumor. Why might the rumor be retrieved?

Accepted Answer

Its embedding can be a near neighbor even though the rumor is wrong

Question 59

A bracket names a manual, and none of the manual's sentences are in the window. What can that citation supply?

Accepted Answer

Only the bracket's own text, because the manual was not inserted

Question 60

A limit is written in the first sentence. Truncation keeps the tail and drops the head. What happens to that limit?

Accepted Answer

Its tokens are dropped with the chosen end, so the limit is not in the window

Question 61

Which text is suited to a prompt cache?

Accepted Answer

A long instruction block that is the same prefix on many calls

Question 62

A leading line says "answer briefly," and the user then asks a detailed question. How should both parts be described?

Accepted Answer

Both are text the model sees in the prompt

Question 63

Why can an answer quote a retrieved chunk and still be wrong?

Accepted Answer

The quote is the inserted text, and that text may itself be wrong

Question 64

Someone treats a reference marker as proof the claim was checked. What is wrong with that treatment?

Accepted Answer

The marker is only as good as the text inserted with it, and the marker is not a check