Grounded retrieval

Syllabus study companion

A model held to a defined corpus, so it answers from the syllabus rather than inventing, and says so when the answer is not there.

Sheet
W-05
Title
Case study
Issued for
Information
Drawn by
Snilld

Sheet W-05 · Grounded retrieval · Education

Live
Goes in
A student’s question, and one ingested, versioned syllabus corpus
What we built
Retrieval restricted to that corpus, so the answer comes from the chapter
Comes out
An answer out of the syllabus, on an installable web app
The line it does not cross
A question outside the corpus gets a refusal rather than an improvisation

Note 1 · The constraint

Students already use general chatbots for homework. Those tools do not know which syllabus a child is on, wander outside the chapter, and answer a Class 8 question at undergraduate level. On a shared family phone with a patchy connection they are worse than nothing.

Note 2 · What we built

A retrieval pipeline that restricts answers to an ingested syllabus corpus. When a question falls outside the ingested material, the answer is a refusal rather than an improvisation. That refusal is the feature. Grounding of this kind reduces invention substantially; it does not eliminate it, and the product is built on the assumption that a wrong answer is still possible.

Note 3 · What transfers

The control you need before a model is allowed to speak about your products, your policies or your prices. The mechanism that makes it refuse is the same one that keeps it inside your documentation.

Status
Live and public
Corpus
Defined and versioned
Out-of-corpus behaviour
Refuses
Delivery
Installable web app

Start with a prototype

Two to four weeks working from your own files, and at the end you hold something real. It is the cheapest way to find out whether any of this applies to you.

Start with a prototype