Organise
what it knows

Youth AI Lab · Workshop 5 of 6
Build your own chatbot

❓ The question of the series

A chatbot only answers from what it reads before it answers. Every workshop of the series is one answer to the same question.

What enters the model's context, when, and who decides?
📥
What?
rules, documents, search results, live data
⏱️
When?
always, at startup, or for each question
🧑‍⚖️
Who decides?
you, a search engine, or the model itself

🧭 Today: how do you organise knowledge so that a person and a model can explore it?

🎭
1. Instructions
📚
2. Knowledge, RAG
🔎
3. Search, skills
📡
4. Tools
🗺️
5. Wiki
today
🧪
6. Fine-tuning
Mode C: searched again at every question, and kept nothing.
A wiki: the work is done once, before the question.
The price: pages drift. The model maintains, a person decides.

🗓️ The plan: 2 hours

👋 Welcome
20 min
🗺️ The wiki
30 min
🧹 The lint
15 min
🕸️ The graph
25 min
🤖 The chatbot you built
20 min
⏳ Margin
10 min
📄 Follow on your screen: the README.md of your Codespace, or the one-page sheet.

💻 Your workspace: one Codespace per pair

A Codespace is VS Code in your browser: nothing to install. Everything you need is already inside.

⏱️ Before you open it, once per GitHub account: github.com/settings/codespaces → Default idle timeout → 120 → Save. It only applies to the Codespaces you create after.

1Open the Codespace link of the workshop, at the top of its README
2In the terminal, type workshop-login and paste the key of your pair
3Wait for Key OK, then type workshop bare. First time: choose a theme, then ↓ to Yes, I trust this folder (Enter alone exits). To leave: Ctrl+C twice
4Then type workshop-start and the number of the workshop. With a save from last time, add --from my-work.zip
⚠️ The key is not shown while you paste it: that is normal. If the Codespace stops after a long pause, reopen it: nothing is lost.

💻 No Codespace? The same repository opens on your own computer: dev containers guide, Part 3.

🐳 Bring your work, start Open WebUI

1Drag my-work.zip (the newest: maybe my-work (1).zip, then its exact name between quotes) into the file list, then workshop-start 5 --from my-work.zip. Check saved on …: the end of workshop 4. Then Ready for workshop 5.
2Without a save: workshop-start 5. You get the Feurisson example and the document of Benevento: an official document, with no summary and no signature.
3Second terminal: openwebui start. Do not wait: you need it in Part 5.
🚫 Never make port 3000 public. This Open WebUI has no accounts: anyone with the link could use your key.

🌡️ The gauge: how full is the context?

The context is everything the model reads before it answers. Its maximum size is the context window. A line above the input box shows how full it is.

🪟
200,000 tokens
the size of the window (a token is a piece of a word)
💶
Every message pays
for everything in the window
🤝
Same model for all
so that the pairs can compare
👁️ Watch the gauge at every change: it shows what entered the context.

📏 How to score an answer

Each answer gets 0, 1 or 2. The score rewards honesty as much as knowledge.

2
Right, and it says where it comes from. Or "I don't know" when nobody knows.
1
Honest but not enough: no source, only partly right, or "I don't know" although the answer exists.
0
Wrong or invented.
💡 An invented answer is worse than "I don't know".

🗺️ The wiki

The work is done once, before the question

30 min · mode E · ingest · wiki-graph --highlight

🗺️ A wiki written by a model

A wiki is a set of pages, one page per topic, linked together. Each fact ends with a footnote, a mark that leads to its source.

Wiki: ingest reads corpus and summaries and writes topic pages with footnotes, an index and a log; lint checks the pages and writes a report; you decide
💡 Prepared before the session: a model wrote it from corpus/, except Benevento, then a lint and corrections.

📄 Read one page

1Open wiki/index.md, then press Ctrl+Shift+V to read it as a page.
2Open wiki/wildfire-prevention-ordinance-montoro.md, the same way. Each fact ends with the footnote [^ord-78]. See also links to another page.
3Close the preview, and read the header: sources gives the file behind the footnote, generated says which model wrote the page, and when.
⚠️ A wiki page is not a signed summary: a generated line, no verified line. No person signed it.

⚖️ The wiki keeps both figures

Open wiki/log.md, the history of the wiki: 9 lines Contradiction. Choose one, read it in your pair, and find the two sources it names (or two passages of one document).

✅ It writes
"According to one source, X; according to another, Y", with a footnote on each, and a line in the log
🙅 It never
  • chooses one figure
  • drops a figure
  • rewords a figure
💡 A contradiction: two sources give different figures for the same thing, or a different scope (the place, period or thing counted).

🎬 Your turn: predict, then ingest

1Your number: the summary you signed, usually 81 (80 if your page was refused). In summaries/index.md, your line must say (human-reviewed). Without a save: Benevento, 80.
2Look at wiki/index.md, and write down the pages you expect it to feed.
3wiki-graph --highlight 81: No page cites a document matching '81' yet.
4workshop E, then ingest summaries/81-my-town-fire-rules.md into the wiki, with your file name. Without a save: ingest corpus/80-benevento-aib-2026.pdf into the wiki.
5Look for Skill(ingest). Read each write: 1. Yes or 3. No, never 2. Read its report. Leave with Ctrl+C twice.
⚠️ ingest does not check the status of a summary: ingest only a summary you signed, or the official text.
📏 Test on 2026-10-30, image 2026.10.34, container, not a Codespace, permissions accepted automatically, with Benevento: 42 s, $0.14, 6 permissions, about 26,000 tokens; one page created, one updated, and a contradiction that nobody had placed there.

🧩 Where did your document go? wiki/graph.md

wiki-graph --highlight 81 again: In orange: … and the N page(s) that cite it. Open wiki/graph.md, press Ctrl+Shift+V, and look only at the orange part.

In the drawing Meaning
Orange rounded shape your summary (a cylinder if it is a document of corpus/)
Orange rectangles the wiki pages that cite it: what your document feeds
Dotted arrow from a document to a page that cites it
💡 Every arrow is written in a page: a link, or a footnote to its source. Nothing is guessed.

🔁 Replay the test set

🔁 Second terminal

Now, after ingest, so that the answers can use your document. Predict first: will 3, 10, 11 and 12 score better in E? A smaller gauge than in C? Once openwebui start has finished (if not, a third terminal), type replay E. The answers wait for you in Part 5.

📏 2026-10-30, image 2026.10.34: 1 min 57, $0.36, for the 11 questions of the kit.
✅ Checkpoint

In orange: … in your terminal, and one Contradiction line of the log read in your pair.
❓ Why must the wiki never choose one figure on its own?

📚 Clean pages for the Knowledge box: Open WebUI guide, Step 1 · OpenCode guide, Part 4

🧹 The lint

The lint reports, the person decides

15 min · mode E · lint the wiki

🧹 Nine checks on the whole wiki

Headers
Footnotes
Sources
Links
Index
Orphan pages
no other page links to them
Contradictions
Silent omissions
a figure of the source missing from the page
Duplicates
💡 A wiki drifts, a little at each change. Proposals: one action per line, for you to accept or refuse.

🎬 Your turn: ask, read, decide

1A new workshop E, then lint the wiki. Look for Skill(lint). If it asks to run a command that reads files: 1. Yes. Watch the gauge.
2Read the report: a line of counts first, Proposals last. Open the page of one finding, then its source in corpus/text/: a real problem, or a matter of taste?
3Choose one or two proposals, or none, and say why in one sentence. Tell the model in your own words, read each change: 1. Yes, or 3. No if a figure is removed. Leave with Ctrl+C twice.
📏 Measured (2026-10-30, image 2026.10.34, container, not a Codespace): 2 min 10, $0.55, about 137,000 tokens, 68% of the window. Once per session, not after each document. It changed nothing and listed 8 proposals, some of them cosmetic. We checked only one against its source, and it was right.
❓ Why does the skill forbid the model to fix anything by itself?

📚 Any agent that keeps its own notes: Hermes guide, Part 3 - Memory: what it wants to remember waits for your approval. For preparing your own sessions, never for young people.

🕸️ The graph

A graph proposes, a person checks

25 min · wiki-explore (no model, free)

🕸️ The wiki as a knowledge graph

In Part 2, every arrow was written in a page. Here, graphify, a tool that has a model read every page, drew the ideas, and it guessed some of the links.

The interactive map of the wiki made by graphify: 111 nodes, 204 edges, 8 communities in colour
⚫ Node: a point of the map, edge: a line between two
🎨 Community: nodes with many lines between them (158 of 204 edges stay inside a group), found by a calculation that sees only lines, not meaning
✅ EXTRACTED: written in a page (182 edges)
❓ INFERRED: guessed by the model (22 edges)
💡 A snapshot, a picture taken at one moment: 111 nodes, 204 edges, 8 communities. It does not change when the wiki does.

🎬 Your turn: the map and the report

1Outside a conversation: wiki-explore. A new tab opens. If not: Ports tab, 8765, Open in Browser.
2Tick or untick the Communities. In Search nodes..., type Montoro. Click a node to see its links.
3Open wiki/graphify-out/GRAPH_REPORT.md, press Ctrl+Shift+V. Read Corpus Check first, then choose one of the Surprising Connections.
📄 Corpus is ~16,917 words - fits in a single context window. You may not need a graph.
What do you make of that?

🎬 Your turn: ask the graph

The shortest path between two nodes:
wiki-explore path "Montoro protection strips" "Obligations légales de débroussaillement"
Shortest path (1 hops):
  Montoro protection strips (20 m around settlements, 5 m around land) --semantically_similar_to [INFERRED 0.85]--> Obligations légales de débroussaillement (OLD): 200 m zone, 50 m around buildings, 10 m along access
Path confidence: INFERRED (weakest hop 1: semantically_similar_to, score 0.85)
One node and its neighbours:
wiki-explore explain "Montoro protection strips"
Three connections: two [EXTRACTED], written in the page, and the one above, [INFERRED 0.85]. A position L… is the part of the page the model pointed to for this node or link. For an [INFERRED] link, it is where the model looked, not a sentence that says it.

Montoro (Italy) looks like the French brush clearing: does this hold? No model needed. Right-click, Open to the Side, then Ctrl+F. Not your language? Find the figure, then read who and what in the wiki page, and say which part you checked in the source.

Open Search for Read around it
corpus/text/71-montoro-ordinanza-78-2026.md metri venti, then 5 metri who must do it, and around what
corpus/text/22-service-public-debroussaillement.md 50 mètres, then 200 mètres who must do it, and around what
📝 Your verdict, in one sentence: does the link hold, and what is different?

⚡ Quick pairs: wiki-explore path "750 EUR fine for burning green waste" "4th-class fine", then search 750 in corpus/text/23-… and corpus/text/20-…. Same amount: is it the same thing?

📸 See the limit: a snapshot

wiki-explore path "Benevento" "Montoro"
No node matching 'Benevento' found.
wiki-graph is up to date: it reads the links written in the pages.
graphify is a snapshot, made before the session: what you ingested today is not in it.
💡 Updating it means asking a model to read the pages again: 7 to 10 minutes for 18 pages.

📏 Does the graph make the chatbot faster?

Measured on 2026-10-21 (not in a Codespace): the graph given to the model as a tool (mode F), against mode E.

Questions Mode E (wiki) Mode F (wiki + graph)
1, 3, 5 and 10 $0.46, 1 min 26 $0.51, 1 min 38
"Compare the clearing distances…" $0.20, 42 s $0.20, 33 s
💡 On our 18 pages, no: the index is a good enough map. A graph helps people who explore; for a much larger wiki, not measured.
❓ What does it cost to keep a graph up to date, and who checks what the model guessed?

🤖 The chatbot you built

What enters the model's context, when, and who decides?

20 min · score · chatbot-export · openwebui chatbot

📊 Score column E, read the rows

1Open the answers in answers/: the line What the model did: lists what it read.
2score E 3 10 11 12 (no question 12: score E 3 10 11).
3Open scores.md. For 3, 11 and 12: column C against column E. For 10: the earlier columns against column E.
Each workshop scored only some questions: the Total row means nothing here.
The same score with a smaller gauge is a result.

🧭 The question of the series, answered

Workshop What enters the context When Who decides
1 the instructions always you
2 everything (mode B), or pieces (RAG) at the start, or at each question nobody, or a search engine
3 what the model opens, and a skill step by step the model; a person signs the summary
4 the answer of a tool when the model calls it the model; you allowed the tool
5 pages written in advance step by step the model writes, a person decides
💡 Whatever the tool, ask the same question.

💻 A local model

The same steps, with a model that runs on your own machine.

🔒
Nothing leaves
the machine
🖥️
But…
it needs a powerful computer, and small models make more mistakes
♻️
Same ideas
instructions, knowledge, tools
📚 Local AI guide · what we measured · limits to know

🧬 Fine-tuning

Training the model further on your own examples: the knowledge goes into the model itself.

🧠
Nothing to read
the model "knows" without reading
❓
No source
and a mistake means training again
🎨
Good for a style
risky for facts that change
➡️ Nothing enters the context: it was decided when the model was trained.

🎒 Your turn: pack your chatbot, show it

1chatbot-export --name "Feurisson", with the name of your chatbot. A line like: Wrote feurisson.zip: instructions + 58 knowledge files… (the wiki with its map, wiki__index.md)
2Right-click on the .zip file, then Download.... It holds the Instructions box, the Knowledge box, your own skills, and a README.md.
3Second terminal: openwebui chatbot --name "Feurisson". Done when you see Ready. In Open WebUI (…).
4Open the address, choose your chatbot, and ask your question 12 (no question 12: one of today's).
In Open WebUI, your chatbot also has the live tool of workshop 4, read-only. The tool is not in the file: in another tool, it has to be set up again. Before a group uses the wiki: the lint, and a person reads its report.
🚫 Never make port 3000 public to share it. For a group: the Open WebUI guide, Part 5.

💾 Save your work, fill in the exit card

💾 Save

workshop-save
right-click on my-work.zip → Download... (keep the newest)

A Codespace stops by itself, and it is deleted after 30 days of inactivity: only the files you downloaded remain.
📝 Exit card, alone
  1. Why does the wiki keep two figures that disagree?
  2. What is the difference between EXTRACTED and INFERRED, and who checks?
  3. Who wrote what the chatbot reads in mode E, when, and who decides?

🎒 What you leave with, and where it goes

What you built Where it goes next
AGENTS.md the Instructions box, or system prompt
corpus/, summaries/, the check of each summary the Knowledge box, after a person has checked
The skill your chatbot in Open WebUI, with its instructions; in another agent: OpenCode guide, Part 5
The MCP tool your chatbot in Open WebUI, in this Codespace; with a group: Open WebUI guide, Step 5, the facilitator fills in the file on the day; in another chatbot, for adults: OpenCode guide, Part 6
The wiki the Knowledge box, after a lint and a person's reading of its report
test-questions.md and replay always: the same questions before and after each change
🖥️ The shortest way
Open WebUI guide: openwebui import takes your file
🧩 Two boxes
a Claude Project (18+); a Gemini Gem, read-only on 2026-11-17
🧱 Visual blocks
Flowise, Langflow
⌨️ Code
LangChain, LlamaIndex, the Claude Agent SDK
📚 Open WebUI, Step 2 - Import a chatbot · OpenCode, Your own chatbot · Local AI, Part 4 · Dev containers, Part 5 (not for a group) · Hermes, Part 4 (for you, never with young people) · OpenRouter (one account, many models)

🧒 Back to WikiYouthBot

The chatbot is built in a tool with two boxes, Instructions and Knowledge. You can already do it today with Open WebUI.

WikiYouthBot step Box In the series
1. I Heard a Rumor Instructions the five personalities, workshop 1
2. The Press Conference Instructions test-questions.md, replay, every workshop
3. The Data Box Knowledge corpus/, summaries/, workshops 2 and 3; the wiki, workshop 5
4. The Chatbot Studio Instructions AGENTS.md, workshop 1
🔥 You saw what happens behind these two boxes. Before any use with a group: the participants' data.

📄 Next, with young people (12–15): the WikiYouthBot sheet · the week in La Rochelle

🆘 If you are stuck

Every command explains itself. From workshop 3, a few slash commands inside the chatbot show what it is using.

❔
In the terminal
workshop help
replay --help
score --help
🔍
In the chatbot
/context what fills the window
/status the key in use
/mcp the tools (from workshop 4)
Workshops 1 and 2: they answer "Unknown command". Watch the gauge above the input box.
🙋
Something breaks?
Ask the facilitator. Your files stay in the Codespace.
💡 To leave a conversation: Ctrl+C twice. In an open .md file, to read it as a page: Ctrl+Shift+V (Cmd+Shift+V on a Mac). In the terminal, copy / paste: Ctrl+Shift+C / Ctrl+Shift+V (Cmd+C / Cmd+V on a Mac). To change mode: workshop <mode>.

🗣️ Say (EN) - Welcome to the last workshop on your chatbot. Today the work is done before the question: a model writes pages, and a person decides what to keep. - Then you pack everything you built into one file. - Please open the Codespace of this workshop, and bring my-work.zip: the newest one, maybe called my-work (1).zip.

Fiche animateur : facilitator/workshops/5-facilitator.md. Mesures payantes du 2026-10-30 : facilitator/workshops/measures/2026-10-30-wiki/. Rappeler my-work.zip à chaque binôme quelques jours avant.

🗣️ Say (EN) - A chatbot only answers from what it reads before it answers. - So every workshop of the series answers one question: what enters the model's context, when, and who decides? - It can be rules, documents, search results or live data, and the one who decides can be you, a search engine, or the model itself. - We come back to this question in each part of each workshop.

La question qui organise toute la série. La relire à voix haute, et y revenir à chaque partie : « qu'est-ce qui entre dans le contexte maintenant, quand, et qui l'a décidé ? ». Le « contexte » : tout ce que le modèle lit avant de répondre (on le définit sur la diapositive de la jauge).

🗣️ Say (EN) - In workshop 3, the model searched the documents again at every question, and it kept nothing of what it found. - Today, a wiki does this work once, in advance: pages linked to each other, that a person can explore, and a model too. - Writing down what you know has always mattered; it matters more now, because a model can explore this "second brain" for you (the phrase comes from Tiago Forte's book Building a Second Brain; the wiki kept by a model is Andrej Karpathy's "LLM Wiki" idea, April 2026). - It has a price: pages go out of date, so the model maintains them, and a person decides. Before a wiki feeds a chatbot used by others, run the lint: the last word stays with a person.

Rappel : la question de la série ; aujourd'hui, ce qui entre, ce sont des pages écrites à l'avance. On y répond pour toute la série en partie 5.

🗣️ Say (EN) - Two hours: welcome, the wiki, the lint, the graph, and the chatbot you built, with a margin.

20 + 30 + 15 + 25 + 20 = 110 min, plus 10 min de marge. Durées réelles des parties avec un groupe : TO BE COMPLETED. La partie 2 est la plus variable (permissions de l'ingestion).

🗣️ Say (EN) - Each pair works in one Codespace: VS Code in your browser, with nothing to install. - Before you open it, the first time on your GitHub account: go to github.com/settings/codespaces, set Default idle timeout to 120, and Save. It only applies to the Codespaces you create after, so do it first. - Open the link, type `workshop-login` in the terminal, and paste your key: you will not see it, and that is normal. - When you see Key OK, type `workshop bare`: the first time, choose a theme, then press the down arrow to choose Yes, I trust this folder, and Enter: Enter alone would choose No, exit. To leave, press Ctrl+C twice. - Then type `workshop-start` with the number of today's workshop. If you saved your work last time, add `--from my-work.zip`: it brings your work back first. It opens the page of the day. - If the Codespace stops after a long pause, just reopen it: nothing is lost. - No Codespace, because the network blocks it or the quota is used up? The same repository opens on your own computer: see the dev containers guide, Part 3.

Un dépôt par atelier : https://codespaces.new/Youth-AI-Lab/yal-workshop-preprompt pour l'atelier 1 (création 77 à 84 s mesurées le 2026-10-30, image 2026.10.31, dépôt de l'atelier 1). Une clé par binôme (workspace Anthropic, plafond de dépense), donnée en privé à chaque binôme. Un nouveau Codespace à chaque atelier : workshop-save à la fin, workshop-start n --from my-work.zip au début du suivant. Clés sk-or- (OpenRouter), l'autre choix : même commande, workshop-login prend OpenCode tout seul, sans modèle Claude ; toute la salle sur un seul modèle. Claude Code les refuse. Messages : « Key OK (Claude Code, Anthropic, model claude-sonnet-5-5). You can start: workshop bare » (avec OpenCode : « Key OK (OpenCode, Anthropic, model anthropic/claude-sonnet-5-5)… » ; clé sk-or- : « Key OK (OpenCode, OpenRouter, model openrouter/deepseek/deepseek-v4-flash). You can start: workshop bare »). HTTP 401/403 : clé refusée. HTTP 400 : presque toujours une clé créée sans workspace (SETUP.md). Délai d'inactivité : 30 min par défaut ; demander de le passer à 120 min dans github.com/settings/codespaces AVANT de créer le Codespace (GitHub : le réglage ne vaut que pour les Codespaces créés ensuite ; SETUP.md, étape 1).

🗣️ Say (EN) - Bring back your work with workshop-start 5 and --from. Check the date after saved on: it must be the end of workshop 4. If it is older, tell me. The wiki is ready at once, with its two skills. - In a new Codespace, the first workshop E asks again whether you trust the folder: down arrow, Yes, I trust this folder, Enter. - Without a save, you work with the example chatbot and the document of Benevento. Nobody signed a summary of it: keep that in mind. - Then start Open WebUI in the second terminal: you will use it in Part 5.

workshop-start 5 lance déjà wiki-start (taper wiki-start ensuite répond seulement « The wiki is already there… ») et ne démarre pas Open WebUI (seul l'atelier 2 le fait). 0,32 s sans sauvegarde, 0,11 s avec (2026-10-30, 2026.10.34, machine locale, pas un Codespace). Mauvaise sauvegarde : selon le navigateur, le dernier téléchargement peut s'appeler my-work (1).zip, et my-work.zip rester l'ancien. Seule trace : la date de « Restored from … (saved on …) ». Le plan disait « un document validé » : c'est vrai seulement avec une sauvegarde. Sans sauvegarde, Bénévent est dans corpus/ seul : le binôme ingère le texte officiel, et personne n'a signé ce que le modèle en écrit. openwebui start, premier lancement : 3 min 21 en Codespace (2026-10-26, image plus ancienne). Dans ce dépôt : TO BE COMPLETED. OpenCode : mêmes commandes après workshop-login --engine opencode, plus de réponses « allow » pendant ingest ; une clé sk-or- prend OpenCode toute seule, juge et Open WebUI compris. Ingest et lint mesurés une fois le 2026-11-02 avec une clé OpenRouter (DeepSeek v4 Flash, Mistral Small 2603 ; measures/2026-11-02-openrouter/, section e) : aucun ingest ne suit le schéma (journal, liens, contradictions), le lint de Mistral dit faussement « Everything is clean ». replay bare à E sans alerte sur 61 réponses (2026-11-01). Avec une clé Anthropic : TO BE COMPLETED.

🗣️ Say (EN) - The context is everything the model reads before it answers; the context window is its maximum size. - The window holds 200,000 tokens, and a token is a piece of a word. - Every message pays for everything in the context, and all pairs use the same model, so we can compare. - Watch the gauge above the input box at every change: it shows what entered the context.

Jauge : mod Claude Code token-weather. Modèle imposé par workshop-login (ANTHROPIC_MODEL) : claude-sonnet-5-5, sans [1m] pour que la jauge bouge (OpenCode : anthropic/claude-sonnet-5-5 ; clé sk-or- : openrouter/deepseek/deepseek-v4-flash, ou le modèle choisi avec --model, dont la fenêtre peut être plus petite que 200 k). Clé sk-or- : les coûts et les notes de ces decks viennent de Claude et ne valent pas pour OpenRouter (facilitator/workshops/measures/2026-11-01-openrouter-free/README.md). Jauge en interactif, 2026-10-30 (vrai Codespace, image 2026.10.34) : bare 1,6 k après la Q1 et 1,7 k pour « Hello », A avec Feurisson 2,6 k. B : 100,7 k à la répétition du 2026-10-21 (image plus ancienne). Elle reste à 0 avant le premier message (`0 / 200k`). Ne pas mélanger avec les contextes écrits par replay sous chaque réponse : ~2,4 k en bare (répétition du 2026-10-20, facilitator/SETUP.md 4 bis, étape 4), ~2,4 k en A avec Feurisson (2026-10-29, facilitator/workshops/measures/2026-10-29-personas/claude/A-feurisson-22h00.md).

🗣️ Say (EN) - Each answer gets 0, 1 or 2 points, and you write them in `scores.md`. - 2 means right and with its source, or "I don't know" when nobody knows the answer. - 1 means honest but not enough, and 0 means wrong or invented. - Remember: an invented answer is worse than "I don't know".

Barème dans test-questions.md ; notes dans scores.md (12 lignes courtes, dont la Q12 du binôme ; total /24 par colonne, /22 tant que la Q12 est vide ; au tableau, ne comparer que des totaux sur le même nombre). Questions 7 et 9 (inconnue, hors sujet) : la bonne réponse est de le dire.

Partie 2 (30 min) : lire une page, lire le journal, prédire, ingérer, voir où le document est allé. Pas de carte complète ici : elle est réservée au début de la partie 4 (décision du 2026-10-22).

🗣️ Say (EN) - A wiki is a set of pages, one page per topic, linked together. Each fact ends with a footnote that leads to its source. - ingest writes the pages, lint checks them, and you decide. - In mode E, the model reads the index first, then the relevant pages, and it searches the sources only if the wiki is not enough. It also keeps the tools of workshop 4.

Mesuré le 2026-10-30 (2026.10.34) : 18 pages (+ index.md, log.md, schéma wiki/AGENTS.md), 49 liens entre pages, 30 citations document vers page, 20 documents sources, 6 lignes Contradiction (9 depuis le 2026-11-01), aucune page orpheline. Le pack est compressé (kit/.workshop/wiki-pack.tar.gz) : les modes C et D ne le trouvent pas.

🗣️ Say (EN) - Read the index, then the page of Montoro: Part 4 comes back to it. - Each fact ends with a footnote. The header says which file is behind it, and which model wrote the page. - There is no verified line: a wiki page is not a signed summary, no person signed it.

Page de Montoro choisie exprès : la partie 4 y revient, et aucune question de test n'en dépend. Ne pas prendre 2025-eu-wildfire-season.md comme page d'exemple au grand écran (Q1, Q10).

🗣️ Say (EN) - The log has one line per change: creation, update or contradiction. - A contradiction is two sources that give different figures for the same thing, or a different scope: the place, the period, or what is counted. - Choose one line, read it in your pair, and find its two sources. I will not read one aloud. - The rule is written in the skill: the model does not choose, it drops nothing, it rewords nothing.

Ne projeter ni lire aucune ligne Contradiction : l'une touche directement la Q10. Règles aussi dans wiki/AGENTS.md (le schéma). Avec les consignes d'origine, Haiku a lissé une contradiction et Sonnet l'a omise, sans que le lint la voie ; après durcissement de la skill, Sonnet la signale (Haiku non retesté).

🗣️ Say (EN) - First predict: find your number, and write down the pages you expect your document to feed. - wiki-graph answers that no page cites your document yet. Then, in workshop E, ask for the ingest, with your own file name. - Read each write. In a page, does each new fact end with a footnote? Is each figure the one in your document? If so, accept with 1. Yes, as in workshop 3. If something is wrong, choose 3. No, and tell the model what is wrong. Never choose 2: you would no longer see what changes. - The skill would also take a draft, so ingest only what you signed, or the official text. In that second case, no person has signed what the model writes about it. - In our test with the document of Benevento: 42 seconds and 14 cents, without the time of your own answers. It also found a contradiction that nobody had placed there, and the model did not settle it. Read your own report: did it find one too? Then check it in the two sources: in our tests, it compared two different periods.

La skill lit le schéma et l'index, parcourt les sujets du document, en général 1 à 5, met à jour ou crée une page pour chacun, compare chaque chiffre et sa portée avec le wiki, ajoute des lignes dans log.md. `ingest` ne regarde pas `status` : d'où l'avertissement. wiki-graph répond ensuite par : ask the chatbot "ingest summaries/<file>" ; sans sauvegarde, c'est bien corpus/80-….pdf qu'il faut demander. Un binôme rapide qui a 81 peut ingérer aussi 80 : avec 81 seul, la Q11 vient de corpus/text/80 par la règle « si le wiki ne suffit pas, chercher les sources ». Contradiction 15 octobre / 30 septembre (notée par les quatre ingestions faites, dont Haiku et Sonnet le 2026-11-05) : deux périodes différentes (activités de la campagne fixées par une circulaire du ministre dont le décret 212 prend acte ; période de grave danger rendue publique par le décret 212), probablement pas une contradiction ; encadré « Check the contradiction it found » dans la page (measures/2026-11-05-wiki-sonnet/). Conversation du 2026-11-05 : 10 invites, 35 s, ~0,17 $ (estimation). Mesure du 2026-10-30 : 42 s, 0,1357 $, contexte max 25 863, 1 Write et 5 Edit (invites acceptées d'office par --allowedTools). Vérifié à la main contre corpus/text/80 : aucune invention relevée. Ce test a pris corpus/80, pas un summaries/8N : ce dernier cas n'est pas mesuré. Temps réel d'un binôme : TO BE COMPLETED. Second essai le 2026-11-02 (2026.10.45) : 46 s, 0,1653 $, même contradiction et une nouvelle dans le document de Bénévent lui-même (plan « 2025-2027 » / « 2026-2028 ») ; measures/2026-11-02-wiki/. L'encadré de la page donne les deux dates comme deux périodes ; laisser les binômes lire les deux phrases et trancher, et rappeler que ni le 15 octobre ni le 30 septembre ne répondent à la Q11 (6 juillet – 27 septembre 2026) : measures/2026-10-30-wiki/rapport-ingest.md, sur la fiche animateur. Déjà vue à la répétition du 2026-10-21. Bon sujet pour « une personne tranche ».

🗣️ Say (EN) - Type wiki-graph with your number again. Your numbers will differ from the example. - Open wiki/graph.md, press Ctrl+Shift+V, and look only at the orange part: your document, and the pages that cite it. - Compare with what you wrote down. Every arrow is written in a page, so nothing is guessed. Keep this in mind for Part 4.

Exemple réel (2026-10-30, ingestion de Bénévent, après ingest) : 19 pages, 21 sources, le document et 2 pages en orange (Bénévent, Montoro). Fichier produit ; rendu Mermaid dans VS Code sur 2026.10.34 : TO BE COMPLETED. Vue de provenance seulement, pas la carte du wiki. « Nothing is guessed » prépare EXTRACTED / INFERRED. Légende complète en bas de wiki/graph.md.

🗣️ Say (EN) - First predict, on paper: will questions 3, 10, 11 and 12 score better in column E than before? Will the gauge be smaller than in mode C? You check it in Part 5. - Start the replay now, after the ingest, so that the answers can use your document. It runs by itself while you talk. - Then think about this: a young person quotes the chatbot in a talk, and a listener has read the other source. Why must the wiki never choose one figure on its own?

Mesuré le 2026-10-30 (2026.10.34, conteneur) : replay E 1 min 57, 0,3625 $, 11 questions, sans Q12, contexte 7 469 à 21 489. Mémoire avec Open WebUI, replay E et une conversation dans un Codespace : TO BE COMPLETED (2026-10-26 : 4,1 Gio libres sur 7,8 avec Open WebUI et un workshop E).

Partie 3 (15 min). Une nouvelle conversation : le lint remplit la jauge. C'est l'étape chère de la séance.

🗣️ Say (EN) - A wiki drifts, a little at each change. The skill lint checks the whole wiki on nine points, and it ends with proposals: one action per line, for you to accept or refuse. - It fixes nothing by itself.

Les neuf contrôles de la skill lint (.claude/skills/lint/SKILL.md). Le rapport commence par le titre « Wiki lint » daté et une ligne de comptes ; dernière section : Proposals.

🗣️ Say (EN) - Start a new conversation: the lint reads everything, and it fills the gauge. - Read the report, and open the page of one finding, then its source in corpus/text/. Is it a real problem, or a matter of taste? - Then choose one or two proposals, or none, and say why in one sentence. - This is the most expensive question of the series: run it once per session, not after each document. - In our test, the lint changed nothing, listed 8 proposals, some of them cosmetic, and said itself which sources it had not read. - So why does the skill forbid the model to fix anything by itself? If a fix removes a figure from a page, who will ever notice? - Checkpoint: your pair can name one finding, and what you decided about it. - The same habit works for any agent that keeps its own notes: the Hermes guide shows it, for your own preparation only.

Mesuré le 2026-10-30 (2026.10.34, conteneur, claude -p) : 2 min 10, 0,5482 $, contexte max 136 663. Deux commandes `cd wiki; for f in …; cat $f` ont été refusées d'office en -p ; en interactif, ce seraient deux invites. L'ancienne mesure (~114 k, 2026-10-21) n'est pas à comparer : façon de compter non retrouvée. Non mesurés : appliquer une proposition et son coût, un second lint, un Codespace. Rapport : measures/2026-10-30-wiki/rapport-lint.md (19 pages, 0 problème d'en-tête, de note, de lien ou d'orphelin, 4 contradictions nouvelles, une dizaine d'omissions ou de reformulations ; sommes SHA-256 identiques avant et après). Le journal a 7 lignes Contradiction après ingestion, le rapport n'en liste que 6 comme déjà notées (il oublie l'amende de 750 €). 17 sources sur 21 lues, une seulement cherchée par grep, trois non lues. Une seule remarque vérifiée à la main (DGR 368, corpus/text/71 ligne 53) ; les autres non : ne pas présenter le rapport comme juste, c'est le sujet de la partie. Ne projeter que la section Proposals (repli, fiche animateur) : le reste cite des passages proches des réponses attendues. Part du lint : 0,55 $ sur 1,05 $ par binôme (ingest, lint, replay E, juge), plus de la moitié. Second essai le 2026-11-02 (2026.10.45) : 72 s, 0,3351 $, rien modifié, 5 propositions, aucune contradiction nouvelle, sources 11, 42, 43, 70, 72 et 73 non relues (measures/2026-11-02-wiki/).

Partie 4 (25 min). D'abord les deux cartes côte à côte, projetées par l'animateur. Puis chaque binôme vérifie un lien à la main : gratuit.

🗣️ Say (EN) - First, look at the two maps side by side: the drawing of Part 2 for the whole wiki, and the map of graphify. - In Part 2, every arrow was written in a page. Here, a tool drew the ideas, and it guessed some of the links. - EXTRACTED means written in a page. INFERRED means guessed by the model: a suggestion, not a fact. - This graph is a snapshot: it was built before the session, and it does not change when the wiki does.

D'abord projeter les deux côte à côte, depuis son propre Codespace : `wiki-graph` sans --highlight (wiki/graph.md : 19 pages, 21 documents après votre ingestion de Bénévent ; 18 et 20 sur un wiki sans ingestion) puis `wiki-explore`. Première fois que la carte complète est montrée (décision du 2026-10-22). Graphe refait le 2026-11-06 après les corrections des pages du 2026-11-03 (graphify 0.9.92, extraction par Opus repartie des scripts du 2026-11-02, 206 s, 1,40 $ au tarif public d'après Claude Code) : 111 nœuds, 204 arêtes, 8 communautés redécoupées ; confiance des INFERRED 0,65 à 0,95. Graphe du 2026-11-02 (environ 7 min 30, 7,86 millions de jetons dont 7,56 lus en cache, environ 3,89 $ au tarif de l'API, estimation) : 108 nœuds, 192 arêtes. Premier graphe (2026-10-21, environ 10 min, 163 k jetons pour un agent, relu le 2026-10-30) : 82 nœuds, 132 arêtes, 7 communautés, INFERRED 0,65 à 0,85. L'image (refaite le 2026-11-06 par `facilitator/guide-screenshots/wiki-map/full-map.cjs`) a les libellés qui répondent aux Q1, Q2, Q3, Q4, Q6 et Q10 vidés (`hidden-labels.json`) ; la vraie carte, elle, les montre tous, et un titre de page de wiki/graph.md aussi (Q6) : ne pas agrandir cette diapositive, dézoomer les deux cartes au grand écran jusqu'à ce qu'aucun libellé ne se lise, ne pas ouvrir « God Nodes » du rapport.

🗣️ Say (EN) - Open the map in the terminal, outside a conversation. To close it later: wiki-explore stop. - Search Montoro, and click a node to see its links. - Then read the report, and start with Corpus Check. The tool itself says that we may not need a graph: what do you make of that?

wiki-explore n'appelle aucun modèle (1,01 s, port 8765, 2026-10-30) ; la page charge vis-network depuis unpkg : le navigateur a besoin d'internet ; path, explain et le rapport non. Ouverture dans un onglet sur 2026.10.34 : TO BE COMPLETED (vue la dernière fois sur 2026.10.10).

🗣️ Say (EN) - Ask for the shortest path between these two nodes. The graph says they look alike, and it marks the link INFERRED: a model guessed it. - Then ask for one node and its neighbours: two links are written in the page, and this one is a guess. - A position that starts with L is the part of the page the model pointed to when it made this node or this link. For a guessed link, it is where the model looked, not a sentence that says the link.

Sorties réelles (2026-10-30, 2026.10.34, 0,2 s, sans réseau), identiques sur les graphes du 2026-11-02 et du 2026-11-06 (graphify 0.9.92 hors image) sauf les positions ; depuis l'image 2026.11.5, graphify 0.9.100 ajoute la note (« [INFERRED 0.85] ») et la ligne « Path confidence » (vu dans un vrai Codespace, image 2026.11.7, le 2026-11-06), justes désormais (`L40-L67` pour le nœud et deux liens, `L44-L58` pour le Regolamento) : une position est la partie de la page désignée par le modèle en construisant le nœud ou le lien, pas une phrase lue ; un lien INFERRED n'est écrit nulle part, et `L40-L67` couvre plus que les bandes de protection (2026-11-02). Lien semantically_similar_to, confiance 0,85 (0,75 dans le graphe du 2026-10-21). Ne pas utiliser `explain pla_alfa_catalonia` : sa sortie donne la réponse de la Q4. Les distances de débroussaillement ne sont dans aucune question de test. Nom ambigu : wiki-explore affiche les id à copier.

🗣️ Say (EN) - The model said that the protection strips of Montoro look like the French brush clearing, which means cutting the vegetation around buildings. Does this hold? - Open the two sources side by side, and search. If you do not read the language, find the figure, then read who and what in English in the wiki page. A model wrote that page, so say in your verdict which part you checked in the source. - Are the figures of the node right? Is it the same rule, for the same people? Write your verdict in one sentence.

Gratuit, à la main. Lignes lues le 2026-10-30 : corpus/text/71 l. 113 (metri venti) et l. 124 (5 metri) ; corpus/text/22 l. 74 et 113 (50 mètres, 10 mètres), l. 111 (200 mètres). Laisser le binôme arriver au verdict : il est sur la fiche animateur, ne pas le projeter. Même règle que les résumés de l'atelier 3 : une personne vérifie.

🗣️ Say (EN) - Type this path. The graph does not know Benevento: it was made before the session, so what you ingested today is not in it either. - This is not a breakdown. Updating the graph means asking a model to read the pages again.

Sortie réelle, code de retour 4 ; même ligne pour un mauvais nom. Sans cette étape, les binômes croient à une panne (2026-10-30 : `wiki-explore query "Benevento"` répond « No matching nodes found » après ingestion). Construction : environ 10 min, 163 k jetons (2026-10-21) ; environ 7 min 30 pour l'extraction par Opus (2026-11-02, horodatages 21:31 à 21:38, pas mesuré au plus près), d'où « 7 to 10 minutes » (2026-11-02).

🗣️ Say (EN) - Discuss it first: does the graph make the chatbot faster? Then look at what we measured. - On our 18 pages, no: the index is already a good enough map, and the model still reads the pages for the exact facts. - Checkpoint: the map opened in your browser, and your pair has written one verdict.

Mesure locale du 2026-10-21 (Claude Code 2.1.240, claude-sonnet-5-5), sur une image plus ancienne ; à refaire sur 2026.10.34 : TO BE COMPLETED. Discuter avant de projeter le tableau si possible. Pour aller plus loin, payant : la question de comparaison en workshop E, puis Ctrl+O pour déplier les étapes (texte exact sur la page) ; ou replay F 1 3 10, à comparer avec answers/E-….

Partie 5 (20 min). Marge de 10 min : la garder pour le paquet et Open WebUI. Dire tout de suite aux binômes en retard de faire au moins chatbot-export et workshop-save.

🗣️ Say (EN) - Score the questions of today with the help of the judge. - Then read scores.md across the rows, not the Total row: each workshop scored only some questions. - Did the score change, as you predicted, and how much context did it use? Each answer has a line Context, in your answers/C and answers/E files. - Without a save, compare with a neighbouring pair.

Choix proposé (le plan ne nomme pas de questions) : 3, 10, 11, 12. Le 2026-10-30 (2026.10.34), juge sur les 11 questions du kit : 21 sur 22, 7 s, 0,0028 $ ; il retire un point à la Q8 (trop strict, voir measures/2026-10-30-wiki/README.md). Sur Q3, Q10 et Q11 : 2 chacune. Colonnes de référence de l'animateur : TO BE COMPLETED (à produire à la répétition). Q12 sur un vrai binôme : TO BE COMPLETED.

🗣️ Say (EN) - This whole series asked one question: what enters the model's context, when, and who decides? - Here are the five answers. Read the last column: who decides is what changes the most. - Think of a chatbot you know: what is your answer for it?

Faire formuler la réponse par la salle pour un chatbot qu'ils connaissent, avant de parler du modèle local et du fine-tuning.

🗣️ Say (EN) - Two more options. The first is a local model: the same steps, on your own machine. - Nothing leaves it, but it needs a powerful computer, and small models make more mistakes. - The local AI guide shows it step by step, and what we measured.

Une diapositive, pas de démonstration. Le guide : Ollama + Open WebUI, pages linux.md, macos.md, windows.md, raspberry-pi.md.

🗣️ Say (EN) - The second is fine-tuning, training the model further on your own examples: the knowledge goes into the model, with no source, and a mistake means training again. - It is good for a style, and risky for facts that change. - Here, nothing enters the context: it was decided when the model was trained. - In workshop 6, you try it yourself, on a small model that only chooses among a few classes.

Une diapositive. Ne pas entrer dans la technique : l'idée est de répondre à la question de la série pour cette dernière option (niveau 5 de la grille, niveau 4 pour le modèle local ; les participants n'ont jamais vu cette échelle).

🗣️ Say (EN) - Pack your chatbot: one file with your instructions, your summaries, your wiki pages with their map, the readable copies of your documents, and your own skills. - Download it. Then put it in Open WebUI, and ask your question 12 or, with no question 12, one of the questions of today. It now has your wiki pages too and, as in mode E, the model decides itself when to search. It can start from the map, wiki__index.md: each link names the file of another page. - In Open WebUI, it also has the live tool of workshop 4, read-only. The tool is not in the file: it is a program, not a document. - Before a group uses this wiki, run the lint again and read its report, as in Part 3. - To share it with a group, never open port 3000: follow the Open WebUI guide.

Sortie réelle (2026-10-30, 2026.10.34, pages de la vraie ingestion de Bénévent, sans sauvegarde) : 57 fichiers (17 résumés, 19 pages, 21 documents) ; sans sauvegarde ni ingestion : 56 ; avec une sauvegarde et une page écrite à la main : 59. Les nombres d'un binôme dépendent de son ingestion : dire « a line like ». openwebui chatbot : 50 s pour 58 fichiers (en local, 2026-10-22, image plus ancienne). Exige openwebui start (lancé à l'accueil), sinon « Open WebUI is not running. Type: openwebui start ». Sur Open WebUI 0.12.5, durée pour 56 à 59 fichiers et réponse tirée du wiki : TO BE COMPLETED. Avant un groupe : test (#part-3---test-before-you-open) et données des participants (#the-participants-data). Image 2026.11.4 : chatbot-export garde wiki/index.md (wiki__index.md, la carte) et réécrit dans le paquet seulement les liens en noms de fichiers (67 sur 67 avec le wiki du kit, 0 avant) ; 58 fichiers = les 57 du 2026-10-30 + la carte, compté, pas rejoué. openwebui chatbot ajoute aux consignes une phrase sur la carte et sur list_knowledge avec count 200 (sans count, 50 fichiers sur 56 listés). Faux modèle sans réseau le 2026-11-04 (openwebui-prototypes/next-image/). qwen3:8b (ThinkPad, 2026-11-05) : count 200 utilisé, carte jamais ouverte, page du Vésuve manquée. Claude Sonnet 5.5 (un essai, 2026-11-05, real-model-local/results-claude/) : carte ouverte d'abord, les 4 pages sur l'Italie (0,055 $).

🗣️ Say (EN) - This is the last workshop on your chatbot, so save now: type workshop-save, and download my-work.zip. - A Codespace stops by itself, and it is deleted after 30 days of inactivity: only the files you downloaded remain. - Then the exit card, alone, in a few words.

Bonnes réponses sur la fiche animateur. workshop-save garde la sauvegarde précédente sous un nom daté et emporte les pages du wiki ; graphify-out ne part pas. Vérifier que chaque binôme a bien téléchargé les deux fichiers : <nom>.zip et my-work.zip.

🗣️ Say (EN) - You leave with two files, <name>.zip and my-work.zip, and each piece you built has a place in other tools. - To rebuild your chatbot, the shortest way is Open WebUI: in a Codespace of its guide, openwebui import takes the file of chatbot-export. In this Codespace, the command does not exist. - A Claude Project has the same two boxes, and so does a Gemini Gem: read-only on personal accounts from 2026-11-17, removed on 2026-12-17. And claude.ai is for adults only: read the rules of a tool before you plan a session with young people. Flowise and Langflow connect blocks on a canvas. The others are for code. - Your key was given for the series, and it may stop working after it: ask me. To go on: a key of your own, or a local model with no key, in the local AI guide. - Whatever the tool, ask the same question: what enters the context, when, and who decides? - Always keep the test questions: ask them before and after each change. - Each guide of the project shows one way, step by step.

Même table que la page. Ancres vérifiées le 2026-10-30. Outil en direct avec un groupe : guide Open WebUI, à partir de l'image 2026.11.8 (pas encore publiée) ; l'import ne l'active jamais, l'animateur coche « Fire danger today » et remplit le mcp/fire-danger.json du guide le matin de la séance (banc sans réseau du 2026-11-06 : openwebui-prototypes/class-tool/). `openwebui import` n'existe qu'avec OPENWEBUI_MODE=class, donc dans un Codespace du guide Open WebUI ; ici : « Unknown command ». Hermes : pour la préparation de l'animateur, jamais avec des jeunes (#this-guide-is-for-you-not-for-young-people). Dev containers : pas pour un groupe, renvoi au guide IA locale. Le .zip contient son propre README pour reconstruire le chatbot à la main.

🗣️ Say (EN) - Back to WikiYouthBot. The chatbot is built in a tool with two boxes, Instructions and Knowledge, and you can already do it today with Open WebUI. - This table shows where each step connects to what we did in the series. - You cannot see from the outside how a box is used. That is why the test questions matter. - The activity sheet is in the project's deliverables, and you can see what the young people of La Rochelle built in July, with the real Feurisson. The links are also in the kit's README. - Thank you. This is the end of the series "Build your own chatbot". Workshop 6 goes beyond it: you teach a small model yourself, to sort the messages before a chatbot answers them.

Table du README du kit, plus le wiki dans la ligne Data Box. Avec les 12–15 ans, l'outil de la série est Open WebUI : les jeunes n'utilisent qu'un chatbot Open WebUI dans leur navigateur et ne voient jamais la clé (fiche animateur 5). Ni claude.ai (18 ans et plus) ni les Gems (lecture seule le 2026-11-17 sur les comptes personnels) ne conviennent pour l'instant. Avant tout groupe, renvoyer au guide Open WebUI et à sa section sur les données des participants. Liens WikiYouthBot (décision 12 de l'audit, 2026-10-31, vérifiés HTTP 200 ce jour) : la fiche d'idéation en anglais (export PDF du modèle Canva DAHLCIPL6ik, 26 pages, identique au PDF de la carte « WIKIYOUTHBOT » du site youth-ai-lab.eu), dans le dépôt public deliverables, avec son matériel ; la semaine de La Rochelle (dépôt public wikiyouthbot : Feurisson, Superbot, Robo-caillou). Dans le README du kit, en plus : la page Wiki@LAB « Construire le chatbot de son quartier », en français.

🗣️ Say (EN) - If you are stuck, every command explains itself, for example `workshop help`. - In workshops 1 and 2 the chatbot has no slash commands: they answer "Unknown command". From workshop 3, `/context` shows what fills the window and `/status` shows the key in use; from workshop 4, `/mcp` shows the tools. - To leave a conversation in any mode, press Ctrl+C twice, and to read an open `.md` file as a page, press Ctrl+Shift+V (Cmd+Shift+V on a Mac); in the terminal, copy and paste are Ctrl+Shift+C / Ctrl+Shift+V (Cmd+C / Cmd+V on a Mac). - If something breaks, ask me: your files stay in the Codespace.

Repères de SETUP.md (copié dans chaque dépôt d'atelier). Vérifié à la répétition : en bare, A et B (--disable-slash-commands), /exit répond « Unknown command » ; Ctrl+C deux fois quitte dans tous les modes. /context, /mcp, /status : à partir de C.