THE WORKING SET / AN EXPLAINER
01 / THE WORKING SET

A smaller context.
A deliberate boundary.

Retrieve from your documents.
Keep only candidates that clear the floor.

01

Question

Tokenize the words

refund
02

Rank

BM25 keyword matching

03

Gate

Apply a relevance floor

A lexical threshold
04

Working set

A working set can be empty. Returning something is not the goal.

02 / A SOURCE TO USE

A question.
A source that fits.

THE QUESTION

Tokenizedrefund
Returns & Refunds
BM25 score

One source clears the gate

SOURCE EXTRACT01 DOCUMENT

Returns & Refunds

Retrieved knowledge → your model's prompt

Bring your own model. groundedwork assembles evidence; it never calls an LLM.

03 / THE EMPTY SET

No evidence?
Don't fill the silence.

THE QUESTION

capitalfrance

No matching documents in this corpus.

EMPTY WORKING SET

No knowledge injected

grounded = false
ABSTENTION
“I don't have that in my knowledge base, so I won't guess.”

No LLM call. An explicit no-evidence response.

A boundary around these documents — not a claim about everything the world knows.

04 / KNOW THE BOUNDARY

A lexical gate.
Not a truth detector.

A RECORDED WRONG MATCH

SOURCE RETURNED

Pausing Your Subscription

Score

grounded = true means a source passed the floor.
It does not mean the source answers the question.

THE TAKEAWAY

Retrieve evidence.
Keep the right to say no.

Evaluate the retrieval. Bring your own model. Keep the boundary visible.

EXPLORE THE LIVE PLAYGROUND

Threshold ≠ semantic correctness verification. Evaluate your sources and your model separately.