A common complaint when a teacher tests a conversational AI in class for the first time: where does what the character says come from? Did it invent this fact, or did it draw it from a real document? On Parlova, the answer involves a precise technical mechanism — RAG (retrieval-augmented generation) — which the teacher fully controls: which documents to provide, with what reliability, and exactly what the student sees from this source. Here's how it works, in detail.
RAG, in one sentence: the AI responds from YOUR documents
RAG stands for retrieval-augmented generation. The principle: before responding, the model searches ('retrieval') for relevant passages in the documents you have provided, then drafts its response based on them rather than solely on its general knowledge. On Parlova, this mechanism applies to the character the student questions: when a teacher adds a textbook excerpt on the Crusades to a mission, the historical character discussing it with the student can cite this precise passage — and the student sees exactly which one.
Add a document to a mission
In the educational brief editing — when creating the mission, or later from 'Edit' — the '📚 Documentary corpus' section allows you to add sources one by one. For each: a title (the only mandatory field), a type (primary source, secondary source, testimony, official document), an author and a date, a bibliographic reference and a URL if needed, and especially content — imported (PDF, Word, ODT, plain text, or Markdown, up to 15 MB) or pasted directly. Without content, the source remains a reference: characters can say it exists, but never draw a specific fact from it.
A checkbox specifies if this text is exact — the AI can then quote it word for word in quotation marks — or a rephrasing, usable as a factual basis but never cited as is. An automatically imported document remains 'awaiting validation' until the teacher has reviewed and confirmed the extracted text: a safeguard against a failed extraction that would become, without control, a 'fact' presented to the student.
An optional field, 'Known from', specifies since when a document exists in the character's world. Example: if a teacher adds the text of the Declaration of the Rights of Man (August 26, 1789) to a mission where the character operates in July 1789, they are not allowed to rely on it — in their world, at that date, it does not yet exist. The same principle as the temporal constraint preventing Napoleon from knowing Waterloo, applied this time to documents rather than events.
The 3 levels of reliability: Limited, Semi-limited, Free
Once at least one document is validated, a selector appears to choose how far the character can go beyond what is provided:
- →Limited — the character draws ONLY from the provided documents and the mission's own content (their biography, concepts, events from the brief) — never their general knowledge, even for a fact they would know otherwise. The strictest, most reliable — but they often answer 'I don't know', even for simple things not covered by the documents.
- →Semi-limited — the character prioritizes the provided documents for any question concerning them, but retains their general knowledge for the rest. This is the mode automatically chosen as soon as a first document is added; the teacher remains free to change it.
- →Free — the character can use the provided documents as well as their general knowledge, without restriction. This is the default mode as long as no document is provided.
This choice never affects the issue, concepts, benchmarks, or final question of the mission — only the origin of the facts the character can use to respond.
What the student sees: the cited source, passage by passage
When a response relies on a provided document, a discreet link appears under the message: '📄 Documents made available for this response'. By clicking, the student sees the document's title, the relevant page if known, and the exact excerpt used — in quotation marks, truncated with a 'Show more' button for long passages.
The choice of words is deliberate. This mention attests that these passages were provided to the model at the time of responding — not that every word of the response directly stems from them. An important nuance to avoid implying stronger traceability than what it actually is.
The printable student sheet, without the answers
This sourcing work extends to paper. From '📚 My missions', two sheets are printable for each mission: the classic PDF sheet, with summary and complete answer key, reserved for the teacher — and a distinct '📝 Student sheet', to be distributed before the session. This one never reveals the answers: neither the concepts to discover, nor what the characters know, nor the answer key to the final question — only blank spaces to fill in during or after the investigation, in response to an explicit request from a teacher who did not want a student to go through a session without a written record.
These three components — documentary corpus, reliability mode, visible citation — address the same concern: that a teacher relying on AI in class retains real control over what their students read, without having to blindly trust a language model.
All you need is a free teacher account. Your students join the mission with a link and a password.
Create my account →