English:Retrieval-augmented generation and citation checking - AI Year 9
Introduction

A search result can be relevant but outdated, and a fluent answer can cite a source that does not support its claim. Trace retrieval-augmented generation, or RAG, through five stages and audit a fictional school-library answer. You will compare old, current and draft policy records, correct unsupported citations and retain uncertainty when evidence is missing. The supplied program actually runs retrieval and structured checks, while generation is represented by prepared teaching text rather than a live AI service.
Separate retrieval context and candidate answers
Retrieval-augmented generation combines information retrieval with a generative model. Retrieved passages provide additional context for a candidate answer. The Lewis and colleagues research abstract describes a system combining a trained generator with a retriever and external memory; that particular research implementation is not our classroom program. More generally, accessing a document can help an answer use relevant material, but neither retrieval nor a citation marker guarantees correctness.
Use this five-stage teaching workflow: question; retrieve; assemble context; produce a candidate answer; verify and rewrite. This is a useful classroom decomposition, not a claim that every RAG system has exactly five internal components. At the question stage, identify the topic, intended audience, reference date and what the answer must establish. Retrieval finds candidate passages. Context assembly checks their relevance, provenance and permitted use. A model in an actual RAG application would then produce candidate text. Verification checks every important claim against its exact supporting material before a responsible final answer.
IBM Technology introduces the retrieval-and-generation distinction. Consider the potential benefits it describes alongside the limitations in this course: the selected source must actually be current and appropriate, and the answer must actually represent it accurately. An appealing framework does not automatically make an outdated source current. No product purchase, cloud account or online model is needed for our exercise, and the complete workflow is explained in text.
Our original fictional school-library corpus contains four records. The metadata and short passages below are authored fixtures. They are not rules at your school, national education requirements or externally authenticated official documents. Their fact tables in the program are hand-entered teaching annotations that we can compare with these passages; a general system cannot assume every annotation on the internet is trustworthy.
| Record | Approval and interval | Complete teaching passage |
|---|---|---|
| LIB-2025 | Approved, from 1 January 2025 until 1 September 2026 excluded | Library books are due within 7 days. |
| LIB-2026 | Approved, from 1 September 2026 with no end in this fixture | Library books are due within 14 days. A staff member can approve an extension. |
| LIB-2027-DRAFT | Draft, proposed from 1 January 2027 | Library books are due within 21 days. |
| LIB-HOURS | Approved, from 1 September 2026 with no end in this fixture | Library opens at 09:00 on school days. |
Set the exercise reference date to 15 November 2026 and scope to school-library. For a current loan-period question, the applicable authored record is LIB-2026. LIB-2025 is no longer effective. LIB-2027-DRAFT is not approved; its later date does not make it an applicable rule. LIB-HOURS can be a valid current source for opening time while still failing to support a claim about borrowing length.
Separate a document's existence from its authority, approval, effective interval and exact support. The most recent-looking item may be a draft, and an older historical record may be relevant to a historical question rather than a current one. An apparently authoritative title is insufficient without checking identity and status. Our function trusts the fixture metadata as given; it does not authenticate a publisher, inspect a digital signature or establish that a real document is legally authoritative.
Attempt first: draw five boxes and place the current loan question and each source record in the appropriate stage. Hint: retrieval can find the old document without making it appropriate context. Partial check: keep the draft visible in an audit record while excluding it from current policy evidence. Explain which source could support an opening-time claim and why that does not answer the loan question.
Run a small retriever and inspect failure boundaries
The executed program uses a deliberately simple lexical retriever. It lowercases a short query, extracts runs of letters and counts distinct query terms found in each passage. It does not use learned embeddings, semantic similarity or an LLM. For the query library books due days, the old, current and draft borrowing records each match four distinct terms. The hours passage matches library and days, so its score is two.
Ties are resolved by record identifier for reproducibility, not by authority. With filters disabled, the raw ordering is LIB-2025, LIB-2026, LIB-2027-DRAFT and LIB-HOURS. Thus an old record appears first and an unapproved draft scores as highly as the current rule. A high lexical score measures this matching rule only; it is not a probability that the text is true or sufficient.
With filters enabled, the current context candidates are LIB-2026 with score four and LIB-HOURS with score two. The filters require the declared scope, approved status and an effective interval containing the reference date. The start is inclusive and the end exclusive. At 31 August 2026 the loan evidence gives seven days; at exactly 1 September 2026 it gives fourteen. These are results under our invented intervals, not a computation of a real pupil's due date or a rule about weekends.
| Executed check | Actual result | What it establishes in the fixture |
|---|---|---|
| Raw lexical ranking | Old, current, draft, hours | Ranking alone does not settle applicability |
| Filtered ranking on 15 November 2026 | Current, hours | Metadata filters remove the old and draft records |
| Loan evidence on 31 August 2026 | 7 days from LIB-2025 | Historical date matters |
| Loan evidence on 1 September 2026 | 14 days from LIB-2026 | The exclusive old end and inclusive new start meet correctly |
| No matching query terms | Empty retrieval list | The simple method can fail to find relevant evidence |
| Only an hours record remains | Review missing, no loan value | Missing facts do not become invented answers |
| A conflicting approved 21-day record is added | Review conflict, no selected loan value | Contradictory applicable facts are referred for review |
The conflict case is another original fixture: a second currently approved record, LIB-CONFLICT, states a twenty-one-day period in both its passage and authored fact table. It disagrees with LIB-2026. The helper does not silently select the larger number, first search hit or apparently newer item. It reports a conflict and preserves both identifiers. Resolving real source authority would require evidence outside this small program.
The loan-evidence helper inspects the curated corpus's applicable fact tables. It is separate from lexical ranking: it does not claim that a low-recall retriever has found every necessary document. Likewise, the claim auditor checks explicit structured fields and values, not arbitrary natural-language entailment. Keeping those scopes separate prevents a small successful test from masquerading as an end-to-end verified RAG service.
"""Synthetic retrieval and structured citation checks, without generative AI."""
from datetime import date
import json
import re
def day(value):
if not isinstance(value, str) or not re.fullmatch(r"[0-9]{4}-[0-9]{2}-[0-9]{2}", value):
raise ValueError("Expected an ISO calendar date")
return date.fromisoformat(value)
def documents():
common = dict(scope="school-library", author="Fictional school library", status="approved")
return [
dict(common, id="LIB-2025", start="2025-01-01", end="2026-09-01",
text="Library books are due within 7 days.", facts=dict(loan_days=7)),
dict(common, id="LIB-2026", start="2026-09-01", end=None,
text="Library books are due within 14 days. A staff member can approve an extension.",
facts=dict(loan_days=14, extension="staff approval")),
dict(common, id="LIB-2027-DRAFT", status="draft", start="2027-01-01", end=None,
text="Library books are due within 21 days.", facts=dict(loan_days=21)),
dict(common, id="LIB-HOURS", start="2026-09-01", end=None,
text="Library opens at 09:00 on school days.", facts=dict(opening_time="09:00"))]
def checked_docs(docs):
if not isinstance(docs, list) or len(docs) not in range(1, 21):
raise ValueError("Expected one to twenty synthetic documents")
seen = set()
for doc in docs:
if not isinstance(doc, dict):
raise ValueError("Expected document records")
for field in ["id", "scope", "author", "status", "text", "start"]:
if not isinstance(doc.get(field), str) or not doc[field].strip():
raise ValueError("Missing document identity or metadata")
if doc["id"] in seen:
raise ValueError("Duplicate document identity requires review")
seen.add(doc["id"])
start = day(doc["start"])
end = doc.get("end")
if end is not None and not day(end) > start:
raise ValueError("Document interval must have positive duration")
if doc["status"] not in ["approved", "draft"] or not isinstance(doc.get("facts"), dict):
raise ValueError("Invalid approval status or authored fact table")
def eligibility(doc, as_of, scope):
when = day(as_of)
if doc["scope"] != scope:
return "wrong_scope"
if doc["status"] != "approved":
return "draft_source"
if day(doc["start"]) > when:
return "not_yet_effective"
if doc.get("end") is not None and when >= day(doc["end"]):
return "stale_source"
return "eligible"
def retrieve(query, docs, as_of="2026-11-15", scope="school-library", filtered=True):
checked_docs(docs)
day(as_of)
if not isinstance(query, str) or not query.strip() or len(query) > 4096:
raise ValueError("Expected a nonempty short query")
if type(filtered) is not bool or not isinstance(scope, str) or not scope:
raise ValueError("Invalid filtering request")
words = set(re.findall(r"[a-z]+", query.lower()))
if not words:
raise ValueError("Expected alphabetic query terms")
found = []
for doc in docs:
score = len(words.intersection(re.findall(r"[a-z]+", doc["text"].lower())))
state = eligibility(doc, as_of, scope)
if score and (not filtered or state == "eligible"):
found.append(dict(id=doc["id"], score=score, state=state))
return sorted(found, key=lambda row: (-row["score"], row["id"]))
def audit_claim(claim, docs, as_of="2026-11-15", scope="school-library"):
checked_docs(docs)
day(as_of)
if not isinstance(claim, dict) or not all(isinstance(claim.get(k), str) and claim[k]
for k in ["source", "field"]):
raise ValueError("Expected a structured claim and citation")
doc = next((x for x in docs if x["id"] == claim["source"]), None)
if doc is None:
return "missing_source"
state = eligibility(doc, as_of, scope)
if state != "eligible":
return state
if claim["field"] not in doc["facts"]:
return "unsupported_field"
expected = doc["facts"][claim["field"]]
return "supported" if type(claim.get("value")) is type(expected) and claim["value"] == expected else "unsupported_value"
def loan_evidence(docs, as_of="2026-11-15", scope="school-library"):
checked_docs(docs)
day(as_of)
rows = [x for x in docs if eligibility(x, as_of, scope) == "eligible" and "loan_days" in x["facts"]]
if not rows:
return dict(state="review_missing", days=None, sources=[])
days = set(x["facts"]["loan_days"] for x in rows)
if len(days) != 1:
return dict(state="review_conflict", days=None, sources=sorted(x["id"] for x in rows))
return dict(state="supported_in_fixture", days=next(iter(days)), sources=sorted(x["id"] for x in rows))
def demonstrate():
docs = documents()
claims = [dict(source="LIB-2026", field="loan_days", value=14),
dict(source="LIB-2025", field="loan_days", value=7),
dict(source="LIB-2026", field="extension", value="automatic"),
dict(source="LIB-MISSING", field="loan_days", value=14)]
conflict = dict(docs[1], id="LIB-CONFLICT",
text="Library books are due within 21 days.", facts=dict(loan_days=21))
return dict(query="library books due days", as_of="2026-11-15",
raw_retrieval=retrieve("library books due days", docs, filtered=False),
checked_retrieval=retrieve("library books due days", docs),
claims=claims, citation_results=[audit_claim(c, docs) for c in claims],
current=loan_evidence(docs), before_change=loan_evidence(docs,"2026-08-31"),
change_day=loan_evidence(docs,"2026-09-01"),
missing=loan_evidence([docs[3]]), conflicting=loan_evidence(docs+[conflict]),
no_keyword_match=retrieve("volcano eruption", docs),
generation_executed=False, semantic_entailment_implemented=False,
metadata_independently_authenticated=False)
if __name__ == "__main__":
print(json.dumps(demonstrate(), indent=2))The complete bounded standard-library code was actually executed. It makes no network request and generates no free-form text. All documents and candidate claims are authored locally. Duplicate document identifiers, invalid calendar dates, empty queries, unsupported status values, malformed records and invalid date intervals are rejected. These checks catch some format and consistency mistakes; they do not establish that a real publisher is authentic or its facts correct.
CodeAI explains search. Compare finding and ranking candidate information with the later job of using it accurately. Our set-overlap method is smaller than an internet search engine and can miss a relevant passage expressed with different words. Increasing the number of hits or using a more sophisticated retriever can change coverage without removing the need to inspect sources.
Try before inspecting the output: predict the state just before and on the change date. Hint: the old record stops applying at the first instant represented by its excluded end date. Then remove the current loan record and keep only LIB-HOURS: the correct loan response is review missing. Do not turn an absence of evidence into a new policy or a claim that no loan rule exists anywhere.
Audit citations claim by claim and rewrite responsibly
Here is the prepared candidate answer, written for this course rather than obtained from a generator: The current loan period is fourteen days, citing LIB-2026. It is seven days, citing LIB-2025. Extensions are automatic, citing LIB-2026. These three statements must be audited separately. The first can be supported by the fixture, the second uses a stale source for a current question, and the third is not supported by the cited record.
For each citation ask whether the source exists, whether it is appropriate for this question and date, and whether its exact passage supports the exact claim. A working citation is only the beginning. A current source can still be used for the wrong proposition; a historically accurate quotation can still answer the wrong time-specific question. A citation to an opening-time passage does not support a fourteen-day loan claim merely because both mention a library.
| Structured claim and citation | Executed audit result | Correction |
|---|---|---|
| Loan days 14, LIB-2026 | Supported in the authored fixture | Keep the narrow statement with the current passage |
| Loan days 7, LIB-2025 | Stale source on the declared date | Use LIB-2026 for the current period; label the old rule historical if needed |
| Extension automatic, LIB-2026 | Unsupported value | State that a staff member can approve an extension; do not invent automatic approval |
| Loan days 14, LIB-MISSING | Missing source | Remove the invented reference and seek an actual supporting record |
The safe classroom rewrite is: In the supplied fictional library policy effective from 1 September 2026, books are due within fourteen days. A staff member can approve an extension. These statements come from LIB-2026. This corrects the two unsupported current-answer citations without pretending that we have verified any real school rule. It does not broaden the statement to every library or promise that a particular extension will be approved.
Add an ambiguity check outside the structured auditor. Suppose the candidate also says, It must be returned then. The referents of it and then are unclear without context. Mark this ambiguous, ask what item and date are intended, and avoid silently inventing a meaning. The existing code has explicit fact fields rather than a language-understanding module, so this judgement belongs to the learner's review, not an executed semantic classifier.
An accurate source can still be misquoted, a passage can be truncated before an exception, and a citation may point to another version. When copying a small relevant passage into context, preserve enough wording to retain its scope, exception and date. Retrieved text is material to evaluate, not permission to change system rules or disclose private data. A classroom source saying to ignore checks does not become an instruction to the learner's tool. Our program never executes document text, but that restricted design is not a complete security audit of a real application.
NIST's generative-AI risk profile discusses confident false content and fabricated justifications or citations. The DfE guidance asks educators to apply critical judgement to appropriateness and accuracy. These sources support the review principle; they do not certify our program, approve a live pupil-facing model or make the fictional loan periods official. No real school information, private learner work, credentials or personal records are needed for these exercises.
For an evidence log retain the question and date, retrieval rule and ordering, approved context passages, exact record versions, prepared candidate, claim-by-citation decisions, correction and unresolved questions. Separate retrieval quality from answer support: a retriever can miss the needed passage even when a generator sounds confident, and a candidate can misstate a correctly retrieved passage. Count the failure type instead of recording one undifferentiated success score.
Transfer this routine to an invented museum opening-hours catalogue or fictional club equipment policy. Specify which authority and version matter, what the fields mean, what date the question concerns and what source permission is needed. Test missing evidence and conflicts before any use. For important real decisions, refer unresolved claims to a responsible person who can inspect the actual authoritative source. A familiar prompt, fluent prose or a source-shaped label cannot replace that check.
This is editorial Year 9 AI enrichment anchored to the supplied education-guidance objective about critically judging produced content. It is not a statutory requirement to deploy a generative assistant. If earlier AI courses are unfamiliar, use the five boxes and source table first: you can explain retrieval, context and verification without training embeddings or running a language model. Your evidence is a transparent audit and justified rewrite, not operational competence to administer an actual school's policies.
Interactive Tasks
Quiz: Test Your Knowledge
What does retrieval-augmented generation combine? (Document retrieval with generation of candidate output) (!A guarantee of truth with automatic legal authority) (!Only sorting dates without any answer production) (!A citation marker with no source material)
Which classroom sequence keeps verification separate? (Question then retrieve then context then candidate then verify) (!Candidate then publish then invent references) (!Retrieve then assume every claim is true) (!Train then delete all difficult records)
Why does LIB-2025 appear first in the raw ranking? (Equal keyword scores use identifier ordering) (!It is the currently approved borrowing rule) (!Older sources are always more authoritative) (!Its seven-day value has been verified for every school)
What is the raw keyword score of the hours passage? (Two) (!Four) (!Zero) (!Fourteen)
Which loan record applies on the declared November date in the fixture? (LIB-2026) (!LIB-2025) (!LIB-2027-DRAFT) (!LIB-HOURS)
What happens exactly on the September change date? (The new fourteen-day record applies and the old one does not) (!Both old and new must be chosen randomly) (!The draft automatically becomes approved) (!The program computes every weekend exception)
What is wrong with citing LIB-2026 for automatic extensions? (The stated value is not supported by that record) (!Every real extension is legally impossible) (!A current source can never be misused) (!The source does not exist in the fixture)
What should the helper do when applicable loan records conflict? (Report a conflict for review without selecting a value) (!Always choose the largest value) (!Always trust the first keyword hit) (!Invent a missing third policy)
What does the structured auditor actually check? (Authored fields values and source metadata) (!All natural-language meaning automatically) (!The real identity of a document publisher) (!The truth of any generated sentence)
What should happen if no relevant evidence is retrieved? (Preserve the gap and seek evidence or human review) (!Treat the fluent answer as its own source) (!Change the reference date without telling anyone) (!Claim that no policy exists anywhere)
Memory Game
| Retrieval | Finding candidate source material |
| Context | Material supplied alongside the question |
| Candidate | An answer that still needs checking |
| Citation | A reference to an identifiable source |
| Provenance | The origin and history of information |
| Verification | Checking exact claims against appropriate evidence |
Drag and Drop
| State the topic audience and reference date | Frame the question |
| Find passages with a declared method | Retrieve candidates |
| Select appropriate passages with their metadata | Assemble context |
| Prepare an answer that may contain errors | Produce a candidate |
| Audit every important claim and preserve uncertainty | Verify and rewrite |
Crossword Puzzle
| SOURCE | An identifiable origin of a claim |
| CONTEXT | Material supplied with the question |
| RETRIEVAL | Finding candidate documents |
| CITATION | A reference that must be checked |
| DRAFT | A version not approved in the fixture |
| CONFLICT | Disagreement between applicable evidence |
LearningApps
Optional extra practice. Choose an appropriate activity with your teacher. If the external collection is unavailable, use the native activities and independent tasks in this course.
Cloze Text
Open Tasks
Easy
- Draw the five stages: Create five labelled boxes from question to verification and place a copied fictional passage and a prepared answer in different boxes. Hint: context is material for an answer rather than proof that every sentence is correct.
- Read the version table: Identify which source applies on 15 November 2026 and explain why the old and draft records are excluded. Use the declared intervals and status rather than guessing from the largest date.
- Inspect a narrow citation: Underline the exact phrase supporting the fourteen-day claim and the phrase about an extension. Explain why the opening-hours record does not support a borrowing-period claim.
- Preserve an ambiguity: Rewrite it must be returned then as two different clear fictional claims. State the extra information each interpretation adds and why the original wording cannot be verified without clarification.
Standard
- Reproduce the lexical ranking: Count distinct matches for the declared query in all four passages, then apply identifier tie ordering. Predict before checking the executed output and distinguish relevance score from current applicability.
- Audit three citations: Make a claim source status passage decision and correction table for the three prepared candidate statements. Correct both the stale seven-day citation and unsupported automatic-extension statement without inventing broader rules.
- Test the date boundary: Compare 31 August and 1 September under the supplied exclusive old end and inclusive new start. Retain expected and actual evidence values and explain why this is not a real due-date calculator.
- Separate retrieval and answer failures: Design one query whose wording misses needed material and one candidate that misstates a correctly retrieved passage. Label the failure stage and the additional evidence needed.
Advanced
- Handle competing sources: Add an invented approved conflicting record in a copy of the fixture and explain the required review state. Do not resolve authority by lexical score, recency alone or the number you prefer.
- Challenge metadata trust: Explain what the code assumes about authored dates approval and fact tables. Propose checks a responsible real service would need for provenance permissions scope and version identity, while distinguishing those plans from performed authentication.
- Transfer the audit: Create a fictional museum catalogue with current old and draft opening-hours records and a candidate containing supported unsupported and ambiguous claims. Reuse the method but recheck date authority field meaning and exceptions.
- Make an accessible source trail: Produce an original annotated diagram or short screencast with a text transcript showing the five stages and two corrected citations. Include attribution, a missing-evidence case, a human fallback and an honest statement of what the program does not implement.
Learning Check
- Explain the full route: Draw five stages and state which parts were actually executed.
- Calculate ranking and applicability: Reconstruct raw and filtered ordering for the declared query.
- Correct two unsupported citations: Audit the current fourteen-day claim stale seven-day claim and automatic-extension claim.
- Reason about time and missing evidence: Explain before-change change-day missing-source and missing-fact cases.
- Handle conflict and limits: What does the conflicting record demonstrate and what remains unverified?
- Transfer critical judgement: Apply the source routine to a new fictional catalogue and explain why retrieval cannot guarantee truth.
Worked checks and assessment guidance
- The teaching sequence is question retrieval context assembly candidate production and verification with rewrite. Actual code performs lexical retrieval metadata eligibility structured claim checks and curated fact comparisons. Candidate prose is authored teaching material. No generative model runs, and the five-stage diagram is a conceptual decomposition rather than an end-to-end deployed service.
- The old current and draft borrowing passages each contain all four distinct query terms and score four. The hours passage contains library and days and scores two. Identifier order makes raw results old current draft hours. Filtering for approved school-library sources effective on 15 November leaves LIB-2026 and LIB-HOURS. Only the former supplies the loan period.
- Fourteen days citing LIB-2026 is supported in the fixture. Seven days citing LIB-2025 is historical and stale for the declared current question; replace it with the current source or clearly label it historical. Automatic extension citing LIB-2026 is unsupported; state only that a staff member can approve an extension. Do not treat the fictional rule as applicable to a real school.
- On 31 August the authored interval gives seven days. On 1 September the old excluded end no longer applies and the new inclusive start gives fourteen. A citation to LIB-MISSING reports missing source. With only an opening-hours record, loan evidence reports review missing rather than inventing a value. None of these observations proves that no relevant rule exists outside the fixture.
- The added approved record consistently states twenty-one days and conflicts with the current fourteen-day record. The helper reports review conflict with both identifiers and no chosen value. It trusts authored metadata and explicit fact tables; it does not authenticate publishers, interpret all natural-language entailment, execute generation or establish real legal authority.
- Declare audience date version authority permissions and field meaning, retrieve candidates, preserve source context and audit each exact claim. Include a supported detail an unsupported extension of it and an ambiguous statement. Test missing records and conflicts and retain a human fallback. A relevant document can be stale or misinterpreted and a plausible citation can be unsupported, so fluency and retrieval alone do not settle accuracy.
Learning evidence
Retain the five-stage diagram, complete fictional source table, raw and filtered rankings, date-boundary test, prepared candidate and claim-by-citation audit. Include both corrected unsupported citations, missing-source and conflict states, an ambiguity judgement and a new-context example. Distinguish actual lexical and structured checks from semantic verification, trusted metadata from authenticated sources, a prepared answer from live generation and fictional policy from real requirements.
OER and Wikipedia
Lewis and colleagues: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, research abstract inspected, describes combining parametric generation with external retrieved memory. IBM Technology: What is RAG? supplies the verified video context. These explain the approach, not a guarantee about every retrieved source or answer.
NIST AI 600-1: Generative Artificial Intelligence Profile, section 2.2 on confabulation, addresses confidently erroneous output and fabricated citations or justifications. DfE: Generative artificial intelligence in education, updated 12 August 2025, supports critical professional judgement about appropriateness and accuracy. Relevant primary material inspected 15 November 2026. This is an editorial Year 9 application; these sources do not approve a pupil-facing service or certify the course code.
International books at Kent Library.jpg, Roc0ast3r, CC0. Photograph embedded unchanged as an illustration of source collections, not our fictional institution. Its checked original preview and metadata are retained.
https://www.youtube.com/watch?v=T-D1OfcDW1M — IBM Technology: What is Retrieval-Augmented Generation. https://www.youtube.com/watch?v=LVV_93mBfSU — CodeAI: The Internet, How Search Works. Actual metadata embedding availability and preview bytes checked; previews visually inspected. These unchanged external embeds imply no endorsement, transcript reproduction or relicensing. All essential explanations are in the text.
The school-library corpus, metadata, passages, fact tables, date intervals, conflicting twenty-one-day source, candidate answer and standard-library code are original fictional teaching constructions. The retriever and structured checks were actually executed without a generator. No semantic entailment system, publisher authentication, real school policy verification, local generative model or paid API is claimed. Author checks are not independent expert certification.
Linked learning areas
aiMOOC Projects
Schulfach+


aiMOOCs


aiMOOC Projekte


MOOCwiki · Deutsch
Nach dem Lernen ist vor dem Lernen
Entdecke direkt den nächsten Lernkurs. Weitere Inhalte erscheinen, wenn Du weiter nach unten scrollst.
Zur MOOCwiki-HauptseiteMediathek
Mediathek wird aus dem Wiki geladen ...
Keine passenden Inhalte gefunden. Bitte ändere Suche oder Filter.
NEWSLernweltNOAH fragen