The School Knowledge Base (RAG): AI That Looks Up Your Documents Before Answering

In the first week of term, a Chinese panel head gets a WhatsApp from a supply teacher: "Which P5 unit are we on this term, and what's the homework cap?" The supply teacher does not want to ping the whole panel, but generic AI blends "suggestions" with "school rules". The panel head later asks the same question in Edor.ai; the system retrieves the school pacing chart and Homework Policy 2026 first, and the reply ends with "Source: Chinese pacing chart 2025–26, section 3; Homework Policy, chapter 2" — two minutes to verify, no guessing.

A school knowledge base (RAG) turns AI from a stranger with impressions of the whole world into someone who can look up your files: retrieve before answering, generate with citations, documents owned by the school, exportable in full, never used for training.

Why generic AI fails on school questions

Ask "which P4 Chinese unit are we on", "what is the homework policy" or "write a notice in our format" and a generic model can only guess from public corpus — often stating another school's practice or outdated advice as fact. Retrieval-Augmented Generation (RAG) turns the question into a search, finds the most relevant passages in the school library, asks the model to answer from those passages only, and attaches source tags. Accuracy and traceability beat relying on model memory alone.

Four steps

Upload — PDFs, Word files, slides and web pages enter the knowledge base; the system splits them into passages and indexes them. Tag every file with subject, grade, school year and owner.

Retrieve — When a teacher or student asks, the system finds the most relevant passages (not the whole book stuffed into the prompt).

Generate — The AI answers from retrieval results, with sources at the end or beside sentences, e.g. "Source: Homework Policy 2026, section 3".

Trace — Users click through to verify; retrieval logs show which documents were cited and which questions returned nothing (= missing documents).

Key point: documents are never used in model training; they are consulted only at answer time.

The value of a knowledge base is not the upload moment but the habit of "no source, no confident claim".

What to upload first

Start small and precise. Priority three: pacing and unit objectives (panel charts, school curriculum outlines), policies (homework, assessment, SEN, IT acceptable use), admin templates (notices, circulars, parent letters, minutes formats). Phase two adds school notes and worksheets (opened by class), rubrics (for AI marking) and school profile plus event lists (for eAdmin).

Six practical scenarios

Planning — "Generate a lesson outline for our P5 Maths unit 6 objectives": the AI cites the pacing chart; the teacher does not retype the year's plan.

Scaffolded tutor — A student asks "what are the key points of this lesson?": the tutor cites teacher-opened school notes, not random web articles; paired with the scaffolded tutor that does not ghostwrite.

AI marking — First-pass feedback cites panel rubric clauses with criterion numbers; teachers see the basis when reviewing.

eAdmin — Notice drafts retrieve past templates for consistent salutations and reply-slip layout; see AI admin agents.

Staff Q&A — "How do we apply for exam accommodations for SEN students?": cites policy chapters so new colleagues need not search shared drives.

New-teacher onboarding — One chat interface for rules, systems and procedures, each answer with a verifiable source, reducing oral-handoff drift.

Permission tiers

School — Policies and admin templates; all staff can retrieve.

Panel — Pacing, materials, rubrics; that subject's teachers only.

Class — Materials a teacher opens to a specific class; student mode touches only opened items here.

Personal — A teacher's private notes; default not in student retrieval.

Student mode passes the safety guard and must not retrieve unopened documents. HR files and case records should not sit in broadly searchable tiers.

Quality and versioning

When a policy updates, keep the old file but mark it "no longer retrieved"; panel heads review outdated documents each term. Retrieval logs show high-frequency citations and zero-result questions — the latter are knowledge-base gaps more useful than guessing.

Data belongs to the school

The knowledge base is accumulated school property. The platform must allow: export everything anytime (original files plus index metadata, free and unlimited); delete means stop citing (AI stops referencing immediately); migration help and deletion certificate at contract end. A knowledge base grows more valuable with use — which is exactly why the school must hold it entirely.

Common misreadings

• The claim: The more documents uploaded, the smarter the AI. What actually happens: Unlabelled, unversioned clutter introduces conflicting clauses. Fewer, accurate files with tags, owners and "current version" flags beat volume.

• The claim: RAG means school data trains the model. What actually happens: RAG retrieves at answer time; training and retrieval are different flows. Contract and architecture should forbid training on school documents and allow export and deletion.

• The claim: With a knowledge base, teachers need not review AI output. What actually happens: Retrieved passages may be outdated, truncated, or the question may exceed document scope and the model still fills gaps. Source tags help verification; they are not a waiver.

Three things to try on Monday

1. Upload three current documents: homework policy, one notice template and one panel pacing chart — as knowledge-base seeds.
2. Ask the same question of generic AI and the school knowledge base (e.g. "what are our P5 unit 6 objectives") and compare whether sources appear and can be clicked to verify.
3. Confirm with ICT: student-mode retrieval scope, export workflow and "old version no longer retrieved" settings are active.