ChatGPT hallucinates about subject-specific pedagogy.
A generic language model sounds convincing, even when its advice on testing or assessment isn't based on anything.
Case · Our own product · Education
DocentRAG
An AI assistant (the engine behind TeachFoundry) that grounds teachers in formative assessment practice, with a source citation on every answer.
Four fields, one standard.
Education and learning materialsConsulting and strategyPublishers and editorial teamsResearch and think tanksTeachFoundryBookaflowValoraEducation and learning materialsConsulting and strategyPublishers and editorial teamsResearch and think tanksTeachFoundryBookaflowValora
Who it's forTeachers & schools
WhatRAG knowledge assistant + material generator
BasisRecognised literature · formative assessment practice
The challenge
Generic AI hallucinates about subject-specific pedagogy.
ChatGPT and similar chatbots give convincing-sounding advice on testing and assessment. Even when that advice isn't based on anything. For subject-specific pedagogy that's a risk: wrong advice hits the student directly.
Teachers who take formative assessment practice seriously don't want loose tips. They want to know what an approach is based on. A recognisable framework, not guesswork from a language model.
There was no knowledge assistant curated specifically for formative assessment practice, with source citation on every answer. That's what DocentRAG adds.
What we built
A knowledge assistant that shows its source.
This is our own product, known publicly as TeachFoundry. DocentRAG is the internal working title for the engine underneath. Together with a subject expert in formative assessment practice, we built a RAG knowledge assistant: teachers ask a question in plain language and get an answer from a curated knowledge base, with its source attached.
The system also generates ready-to-use lesson materials (lesson plans, exit tickets, rubrics, implementation plans) that export straight to PDF or Google Docs.
Technically this is called retrieval-augmented generation: the system first searches its own knowledge base for relevant fragments and only then lets the language model formulate an answer, based on what it found. Not on whatever it happens to “know”. That way, every answer stays traceable to a concrete document from the formative assessment practice framework.
What is this based on?Source citation with every answer.
Formative practice starts with a learning goal the student can explain in their own words.
The sheet the answer rests on, in the knowledge base.Source · Core standards 2024, §4.2
I don't know.
No sourceLow confidence · no source in the knowledge base
How it works
From question to ready-to-use material.
+ 01 · Step
The teacher asks a question.
In plain language, the way you'd ask a colleague. About formative assessment practice, testing or grading.
+ 02 · Step
Searching the curated knowledge base.
Not the open internet and not the language model's memory: only the walled-off knowledge base, based on recognised literature.
+ 03 · Step
Answer with source citation and confidence indicator.
Summary, elaboration, sources and follow-up questions. Plus an indicator: high, medium or low confidence.
+ 04 · Step
Generating ready-to-use lesson material.
From lesson plan to exit ticket, rubric or implementation plan, ready to export straight to PDF or Google Docs.
+ 05 · Step
The teacher judges and uses it.
The system supplies the grounding and the material, the teacher knows the class and decides what goes into the lesson.
Who it's for
Built for teachers and schools.
Teachers and schools use DocentRAG (publicly: TeachFoundry) to ask questions about formative assessment practice and quickly get ready-to-use lesson material, grounded in recognised literature.
Why
What an ordinary chatbot doesn't offer.
Loose tips, no framework.
Advice without reference to a pedagogical framework. You can't check where it comes from, or whether it matches what's already known about formative assessment practice.
An answer isn't lesson material.
A chatbot hands back a block of text, not a lesson plan, rubric or exit ticket you can actually use in class tomorrow.
Every teacher works it out again from scratch.
The same question about formative assessment practice gets put to a chatbot separately in a hundred places, with no shared, curated knowledge base behind it.
The difference
Why this is different from an ordinary chatbot.
A general-purpose chatbot often sounds convincing on educational questions, even when the answer is a hallucination. The model fills a gap in its knowledge with something that sounds plausible about subject-specific pedagogy, without showing it.
DocentRAG works the other way round: the answer only comes after the system has found something in the curated knowledge base. No match (or a low confidence score) and you see that reflected in the answer straight away.
An ordinary chatbot sounds convincing. DocentRAG shows where the answer comes from, and how sure it is.
How we approached it
Source citation and confidence as a design principle.
What the engine did, per questionadmin environment
Raw matches33
After filtering11
In the final context4
The recognised intentAbout formative assessment practice, testing or grading.
Curated knowledge base.
Around 196 documents per language instance, put together with a subject expert in formative assessment practice. Based on recognised literature, including the Toetsrevolutie framework (René Kneyber) and the HOP model.
Source citation with every answer.
No answer without a locatable source. Structure: summary → elaboration → sources → follow-up questions.
Confidence indicator + refusal when off-topic.
Every answer shows high, medium or low confidence. Questions outside the knowledge domain are refused rather than have the system make something up.
Multilingual, multiple instances.
Separate, live instances in Dutch, English and for Hong Kong. Each with its own language and its own knowledge base. The English one goes to market under its own brand, on the same engine.
The teacher keeps the final say.
The system supplies the grounding and the material. The teacher judges and decides what goes into the classroom.
What it means for teachers
What this delivers, beyond the answer itself.
A grounded answer, fast.
No hunting through books or loose articles: the question in plain language, the answer with its source attached.
Ready-to-use lesson material.
Lesson plans, exit tickets, rubrics and implementation plans. Ready to export, instead of drafting them yourself.
Grounded in recognised pedagogy.
Every answer is rooted in a coherent framework for formative assessment practice, not in loose tips from a generic chatbot.
The judgement stays with the teacher.
The system supplies, the teacher weighs up what fits the class. That stays human work.
In numbers
What's live now.
~196
per language instance
Documents in the curated knowledge base
3
live
Language instances: Dutch, English, Hong Kong
28
types
Material types the system generates ready-to-use
It runs live in three languages, on a curated knowledge base built with a subject expert. Every answer comes with its source and a confidence indicator. That's the point, not a dashboard full of numbers.
Start small, go far.
Want to look for yourself first? The supplier scan and the Chain check are free and need no conversation.
The supplier scanWant something like this for your knowledge domain?
In a half-hour conversation we'll see whether your knowledge domain lends itself to the same thing.
Book an intro callHere's what that looks like in practice.
Three different worlds, one principle: the AI does the work, the human checks and approves.
Read the case