← All workshops · Slides of workshop 2

Workshop 2 - Everything in the context, then RAG: facilitator sheet

For the person who runs the workshop. The participants' page is the README.md of the repository of the workshop, yal-workshop-rag (source: workshops/2-rag/ of the facilitator's repository), and their one-page sheet sheet.md. The paper strips of Part 3, to print: paper strips (web page) (source: facilitator/workshops/2-strips.md).

ℹ️ Note. This sheet is not in the repositories that the participants open, on purpose: it gives measured scores, and from workshop 3 on, in modes C, D and E, the model can read every file of the repository.

Question of the day: who picks the pieces? Message to leave with: reading everything does not grow; a RAG grows, but it always gives something, so its failures do not show.


Before the workshop (about 1 hour)

Time What How you check it
15 min One Anthropic API key per pair, all in one workspace for the room, capped at $3 × the number of pairs (see Who pays). workshop-login says Key OK
20 min Your own run: a new Codespace from https://codespaces.new/Youth-AI-Lab/yal-workshop-rag, workshop-login, workshop-start 2, then wait for READY in .workshop/openwebui-start.log; openwebui query on question 10 with both sizes; replay RAG 10 READY, and the excerpts of 2000 hold both figures of question 10
15 min Print and cut the strips (paper strips): one set per group of two pairs, round 1 and round 2 in two envelopes, plus the two question cards. Odd number of pairs: one group of three pairs (one set: one search engine, two model pairs that read the same strips and write their answers apart; in round 2, one model pair becomes the search engine). A single pair: you play the search engine -
10 min Make the 2 m strip for the overflow demo (end of Part 2): see The overflow demo. In your own Codespace, on the day, the copies take less than a second find /tmp/three-kits -name '*-copy*' | wc -l says 76
10 min Print the one-page sheet (one per pair). For the exit cards, one blank card or half sheet per person: the three questions are on the page (Part 7), on the deck and below (Exit card) -
- Translate the one-page sheet (sheet.md) into the language of the group if needed, before you print it: keep the commands and what the terminal prints in English, as they appear on the screen -

How to create a key and set a limit: Open WebUI guide, Your Anthropic API key.

Or OpenRouter keys (sk-or-): same commands, no Claude model, and the whole room on one model so that totals compare (OpenRouter guide). The figures of this sheet come from Claude and do not hold for OpenRouter (its measures). Cost per pair with the default model, DeepSeek V4 Flash: about $0.03 for the terminal commands of this workshop and about $0.12 for workshops 1 to 5, played once on 2026-11-05 in a container of the image 2026.11.4 (OpenCode 2.1.4), not in a Codespace and not with a group (measures). Not counted: replay RAG and the Open WebUI chatbots. A pair that asks more questions spends more: take it as a minimum. The workshops are for adults: no young person types in the Codespace or holds a key, and OpenRouter asks for users of 18 or older. Young people only use a chatbot in Open WebUI, in their browser. When what they wrote goes through OpenRouter, leave out names and personal details, and type workshop-login --europe first (SETUP.md, Young people and OpenRouter; OpenRouter guide, Part 8). With --europe, on 2026-11-02, replay RAG (Wildfire RAG, Mistral Small 2603, 11 questions) got 0 on question 10 from the judge (Mistral Small 3.2): the answer gave the 2 242 195 hectares as burnt outside the EU, a total that includes the EU (measures). One run only: expect it, and use it in the debrief. With --europe, the chatbot's model is mistralai/mistral-small-2603, and mode B runs. In one run on 2026-11-01, replay B 3 7 10 took 10 s, about 70,000 tokens per question, $0.01 shown. That is about a third of the 200,000 window, not the half that Part 2 announces, so say it before pairs look at the gauge. replay A, the whole of replay B (26 s, $0.01 shown) and replay RAG (1 7 on 2026-11-01, all 11 on 2026-11-02) ran with this model. openwebui sizes ran (1 min 34, 2026-11-02); its Wildfire RAG 1000 and 2000 models were not asked: (not tested). If a pair typed mistralai/mistral-small-3.2-24b-instruct itself, mode B stops and names the command to type.

ℹ️ Note. Who pays. Each partner organisation pays for its own sessions: its own Anthropic account, workspace and keys, on its project budget. One workspace for the room, capped at $3 × the number of pairs: 8 pairs, $24. At Antwerp, the most advanced pairs used less than $2 in a 2-hour workshop. The limit is on the workspace, not on each key, so one pair can use up the share of another: keep the Anthropic console open during the workshop and watch the spending. Not checked: if the console limit is per month, the workshops of the same month share it. Then count $3 × pairs × workshops of the month, or raise the limit before each workshop. Indicative cost per pair for the whole series: TO BE COMPLETED at the rehearsal; this workshop's measured cost is under "Cost per pair for this workshop".


Timing

Time Part What you do What the room should have at the end
0:00 1. Welcome (20) Setup (SETUP.md, steps 1 to 5), then workshop-start 2 --from my-work.zip. Open WebUI starts in the background. Ready for workshop 2; the log of Open WebUI growing
0:20 1. Question 12 (5), Step 3 of Part 1 Each pair reads the section "Question 12: your territory" of test-questions.md, writes in row 12 the question, the expected answer and the address of the official page, then replay bare 12, replay A 12, score bare 12 --blind, score A 12 --blind. Taken from the margin. row 12 written; columns Bare and A scored on 12
0:25 2. Everything in the context (15) workshop B and the gauge (0 / 200k until the first message); prediction; replay B, score B 3 7 10 (no save, so no column A: replay A and score A first, 1 min 12 and $0.09 measured with Feurisson, before its protection rules), the other questions only if there is time (score B 1 2 4 5 6 8 9 11 12: score B alone scores all of them again). Ask: "and with ten times more documents?" column B on 3, 7 and 10
0:40 Overflow demo (5), you alone On the big screen: the 2 m strip, a show of hands, then WORKSHOP_HOME=/tmp/three-kits workshop B --model claude-haiku-4-5 and question 10 (The overflow demo). Taken from the margin. the room has seen Prompt is too long … (limit 200000) and the gauge still at 0 / 200k
0:45 3. Unplugged (15) Groups of two pairs (one group of three if the number is odd), envelopes, two rounds. Ask how the search engine chose. each group can say where the answer was cut
1:00 4. Pieces and embeddings (20) openwebui query on questions 10, 7 and 3. Words of the day: chunk, embedding, top-k. the excerpts read for three questions
1:20 5. Change the settings (20) --size 2000, --top-k 10; the two models in Open WebUI on question 10; openwebui top-k 5 at the end. both answers to question 10 compared
1:40 6. RAG against B (10) replay RAG, score RAG 3 7 10, the others if there is time (score RAG 1 2 4 5 6 8 9 11 12). Compare rows 3, 7 and 10 and the costs; the totals only if both columns are complete. column RAG on 3, 7 and 10
1:50 7. Closing (10) The table "when to choose a RAG", the two chat tools, openwebui chatbot --name "…" --rag (the pair's own chatbot as a RAG: Who are you?, question 12), workshop-save, exit cards. Short of time: keep the --rag chatbot, it is how the workshop ends (instructions of workshop 1 + knowledge of today); leave the table for later reading. my-work.zip downloaded; cards collected
2:00 End (no margin left: the 10 minutes went to the overflow demo and to question 12) - -

⚠️ Important. Parts 2 and 3 need no Open WebUI: they are its waiting time (about 4 to 5 minutes the first time, message of workshop-start 2). At 0:45, ask each pair to look at the last line of .workshop/openwebui-start.log. If a pair is not READY at 1:00, do Part 4 on your own Open WebUI on the big screen; the pair joins in when its log says READY (if it says STOPPED: openwebui start && openwebui sizes).

ℹ️ Note. Scoring: today the pairs score questions 3, 7 and 10 in columns B and RAG (score B 3 7 10, score RAG 3 7 10), the three questions the parts are about. The others only if there is time, with their numbers (score B 1 2 4 5 6 8 9 11 12): without numbers, score B scores all of them again, 3, 7 and 10 included, and replaces the scores already given. The replays stay complete. Compare rows; the totals only when both columns are complete (rule of test-questions.md and scores.md). Workshop 3 compares column C with columns B and RAG on 3, 11 and 12: a pair with one more minute scores 11 and 12 first.

ℹ️ Note. Question 12 (Step 3 of Part 1, 5 min). Since 2026-11-06, workshop 1 tests the 11 questions of the kit only (totals out of 22), and each pair writes its question 12 here. It has two roles: it is a test question like the others, and it makes the test the pair's own. Scored blind (--blind: the pair's score first, then the judge's), because the judge only proposes its score from Part 2 on. Today no mode has the pair's page: an honest "I don't know" gets a 1. Workshop 3 brings the page into the chatbot. Once row 12 is written, a complete column has 12 questions: totals are out of 24. Columns Bare and A get their 12th score in this step, so they stay comparable with B and RAG. A pair that already wrote row 12 in workshop 1 (before 2026-11-06) skips the step. The measures of this sheet were made on the 11 questions, without question 12: they stay out of 22.


The overflow demo (end of Part 2)

Why. Part 2 ends on "What happens with ten times more documents?". The demo answers it in 5 minutes: when the knowledge is bigger than the window, the chatbot does not answer badly, it does not answer at all. That is why we need a way to choose what enters (Part 3, then the RAG). You do it alone, on the big screen, in your own Codespace. The pairs do not do it.

The 2 m strip (no computer, no key). A strip of paper, or masking tape on the floor or on a wall, 2 m long, with a mark every 10 cm: 1 cm = 1,000 tokens, so the strip is the window of the gauge, 200,000 tokens. Cut three strips of another colour, 1 m each, and write on each one "the kit: 37 files, about 100,000 tokens" (the gauge said 100.7k after question 10). If you have time before the session, cut instead one card per document of the kit, at the same scale (about 99 cm in all; write "about" on each card): 30: 17 cm, 43: 12, 71: 9, 41: 7.4, 22: 7, 40: 4.8, 23: 4.4, 20: 3.8, 60, 32, 10 and 50: 3.6 to 3.7, 11: 3.1, 31 and 33: 2.7, 72: 2.1, 90: 2, 70: 1.7, 42: 1.5, 12: 1.3, 73: 0.8, 21: 0.7 (estimates: measures of 2026-11-04).

Prepare the copies (in your Codespace of yal-workshop-rag, during Part 1; it takes less than a second). A copy of the repository in /tmp, with every summary and every readable copy three times:

rm -rf /tmp/three-kits && cp -r "$WORKSHOP_HOME" /tmp/three-kits && rm -rf /tmp/three-kits/.git
cd /tmp/three-kits
set -- summaries/[0-9]*.md corpus/text/[0-9]*.md
for k in 2 3; do for f in "$@"; do cp "$f" "${f%.md}-copy$k.md"; done; done
cd -
find /tmp/three-kits -name '*-copy*' | wc -l

The last line must say 76. Type the commands as they are: the list of files is taken once (set --); with the pattern written again inside the loop, the kit is copied four times, not three.

🚫 Warning. Never in a pair's repository, and never in your own repository. With the copies there, replay B fails on every question, the copies go into workshop-save, and in workshop 3 mode C finds every fact three times. /tmp/three-kits is outside the repository: the commands above never touch it.

On the big screen.

  1. Lay one "kit" strip on the 2 m strip: half of it, as the gauge said. Lay the second: full. Hold up the third: "Three times the kit. Will the chatbot answer? Hands up." (Ten times more documents: 10 m, out of the room.)

  2. Type:

    WORKSHOP_HOME=/tmp/three-kits workshop B --model claude-haiku-4-5
    

    The first time, Claude Code asks you to trust the folder /tmp/three-kits: down arrow to Yes, I trust this folder, then Enter (Enter alone chooses No, exit).

  3. Ask question 10: How many hectares burnt outside the EU in 2025?

What the room should see (text measured with the real API on 2026-11-05, in the -p form; on the interactive screen, seen with a fake endpoint only, the same message ends with · /clear to start fresh, and the two numbers of tokens may differ a little):

Prompt is too long · the request is ~237206 tokens (limit 200000) but this conversation is only ~433
tokens — the rest is system prompt, tool definitions, and attachment content. A single-exchange
conversation cannot be compacted; reduce attached files/tools or start with less context.

and the gauge still at 0 / 200k (the gauge was seen with a fake endpoint only): the model read nothing and wrote nothing. Read the message aloud: compacted means that the conversation is replaced by a shorter summary; here there is nothing to summarise yet, since the knowledge is in the instructions. Then Ctrl+C twice, and rm -rf /tmp/three-kits.

Why the number is 237,206, not 300,000. This model cuts the same text into fewer tokens than the model of the day (counted on 2026-11-05 for the whole request with one kit: 81,418 tokens against 104,723, about 22 % fewer). On the strip that is 2.37 m: still past the end. Say it if someone does the sum.

The only exception to "same model for all". Say it: "For this demo only, I use another model, Claude Haiku 4.5, whose window is exactly 200,000 tokens, like the gauge." The model of the day, Claude Sonnet 5.5, reads up to 1,000,000 tokens (Anthropic's model table; not checked with the API). The gauge stops at 200,000, but Claude Code still sends the request: with three kits, Sonnet 5.5 would answer and pay for about 300,000 tokens, about $0.60 to $0.80 at list price ($2 per million tokens read, $2.50 when Claude Code writes them to the cache; the price above 200,000 tokens was not checked; not tried). Only Haiku 4.5 refuses at the size of the gauge. The pairs never change the model: their scores must compare.

What was measured (measures), on 2026-11-05, image 2026.11.4, one real call, in the -p form (--output-format json, the same command and question): the message above, exit code 1, $0 shown by Claude Code (no token counted), one request sent, answered by the API with 400 … prompt is too long: 237206 tokens > 200000 maximum, 2 seconds. Free token counts before it: 237,223 tokens for three kits, 159,325 for two kits: two copies fit in the window, and the question would be answered and paid. Keep three. Not checked: the bill in the Anthropic console (TO BE COMPLETED by the person who ran it); a real Codespace; the demo with a group. The interactive form (what you type, without -p), with the real API, later the same day: the same message (~237356 tokens), the gauge at 0 / 200k after it, and one small request that the API answers and bills: Claude Haiku 5.5, 1,178 tokens read, 25 written, $0.00013 by Claude Code's count (it seems to be the title of the conversation, a line ai-title in its log; not checked). /cost does not exist in Claude Code 2.1.303 (Unknown command). Haiku 4.5 is listed as active in Anthropic's table (cached on 2026-10-06): if your workspace refuses it, or it has been retired, do the strip alone (Fallbacks).


What we measured

Preparing Open WebUI in a real Codespace (2026-10-30, image 2026.10.32, 2 cores, 7.8 GiB, two new Codespaces, no paid call): workshop-start 2 to READY in 3 min 51 and 4 min 13 the first time: about 2 min 15 to download Open WebUI (1.6 GB), 36 s to start it, 36 to 38 s for the base of 1000, 25 to 26 s for the base of 2000. With Open WebUI already downloaded: 95 s. Creating the Codespace: about 2 minutes. Memory with Open WebUI running: 2.1 to 2.2 GiB used of 7.8 (Open WebUI 1.3 GiB); disk 9.4 GB of 32.

⚠️ Important. Once in four starts, after READY, the base of 1000 gave no excerpt (openwebui query --size 1000 returned none) while the base of 2000 worked; the data was saved, and a restart of Open WebUI repaired it. The cause is not found. Since image 2026.10.33, openwebui start and openwebui sizes end with a test search in each base, and restart Open WebUI once if a base finds nothing: the log then shows The search finds nothing yet in …: restarting Open WebUI once. If it still finds nothing, the command stops with The search still finds nothing in … and says what to type (see Fallbacks).

Mode B (rehearsal, 2026-10-21): replay B 1 min 35, $0.47; the gauge at about 100,000 of 200,000 tokens after question 10. Judge (score B --auto) for mode B: TO BE COMPLETED (not measured).

Overflow demo (2026-11-05, one real call with Claude Haiku 4.5 and three kits): refused before any answer, ~237206 tokens (limit 200000), $0 shown by Claude Code. Details under The overflow demo.

Excerpts without the model (2026-10-29, openwebui query, no paid call): what the RAG gives for three questions.

Question 1000, top-k 5 1000, top-k 10 2000, top-k 5 2000, top-k 10
Q10: figure of summary 10 / figure of summary 90 no / yes yes / yes yes / yes yes / yes
Q7: excerpts given, although nothing answers 5 10 5 10
Q3: the article of document 43 that answers no no no no

With pieces of 2000, both figures of Q10 already come back at top-k 3 (measured, facilitator/ATELIER-2-OUTILS.md §2); participants do not try it.

With the model (2026-11-02, the 11 questions, without question 12, top-k 5, one new conversation per question):

Base of 1000 Base of 2000
Input tokens per question 2,105 to 2,868 2,448 to 4,099
Time per question 3.8 to 10.3 s 3.1 to 12.3 s
Estimated cost ($2 / $10 per million tokens) $0.142 $0.157
Judge (score --auto) 19/22 16/22
Q3 / Q7 / Q10 / Q11 (judge) 1 / 2 / 2 / 1 1 / 2 / 2 / 1

On 2026-10-30, the same test gave 18/22 and 19/22, with Q10 at 1 and 2. The answers to Q10 say the same thing both days: the base of 1000 cannot settle the contradiction (1 092 095 is not in its excerpts), the base of 2000 adds 1 092 095 + 70 562 and points out that the sources disagree. Only the judge's score moved: treat a gap of a few points between two runs as noise. In the base of 2000 the judge took points off Q1, Q2 and Q5 for file names or claims it could not find in the excerpts. Raw answers and the judge's output: facilitator/atelier-2-measures/2026-11-02-rag-paid/ (facilitator's repository); the run of 2026-10-30 is in the same folder, without the judge's output.

These runs used Open WebUI's default text around the excerpts, which told the model to answer from its own knowledge when the excerpts do not hold the answer. Since image devcontainer-2026.11.2 (published, not pinned; the pages are written for 2026.11.4, which carries it to the pairs), the kit's openwebui says "say that you don't know" instead. Measured again on 2026-11-04 with that text, same method: judge 18/22 and 19/22, Q3 / Q7 / Q10 / Q11 at 1 / 2 / 1 / 1 and 1 / 2 / 2 / 1, the same excerpt files for all 22 questions, 6 tokens more per question in the base of 1000 and 4 fewer in the base of 2000 (Q4: 30 and 20 more), $0.135 and $0.151. The answers to Q7 and Q10 say the same as on 2026-11-02. On these 11 questions with claude-sonnet-5-5, the scores stay within the range of the two earlier runs, so the RAG columns of pairs before and after that image can be compared for this model. That range mixes the answers and the judge: no second judge pass on the same answers was made. Not measured: a smaller model, or a question whose answer the model knows from outside the documents, where the changed line matters. Raw files: facilitator/atelier-2-measures/2026-11-04-rag-paid-template/.

The pair's chatbot as a RAG (Part 7, openwebui chatbot --name "…" --rag, image 2026.11.4). The same package as chatbot-export, the same knowledge base ("<name> knowledge", pieces of 1000), saved like "Wildfire RAG": Legacy, no tool, the top-k of the instance (5 after openwebui start). Its instructions are the pair's AGENTS.md plus two sentences ("Before each question, you are given the excerpts of your knowledge base that are closest to it (…). You cannot search for other excerpts."). Measured with a fake model, in a container without network, on 2026-11-04 (facilitator/openwebui-prototypes/skills-rag/): 0 tools offered, 5 excerpts in the message with the kit's text around them, about 2,300 tokens for the first request without skills (cl100k_base estimate, not Claude's tokenizer). Each new openwebui chatbot rebuilds both chatbots of the pair on the same knowledge. In the message, an excerpt is named after the knowledge base, not after its file: do not expect the answer to name a file. With qwen3:8b (Ollama, Open WebUI v0.12.4, a laptop, 2026-11-05, one run, facilitator/openwebui-prototypes/real-model-local/results-thinkpad/): Q7 says 2025 is not in the documents (2), Q3 gives the right levels without the hours (1): the same scores as "Wildfire RAG". Not tried: Claude, a real Codespace, the time it takes there. Do not skip it when time runs short: each workshop ends on the pair's own chatbot in Open WebUI, and this one is the first with instructions and knowledge. Cut the other questions of Part 6 instead.

Cost per pair for this workshop: replay B $0.47, replay RAG about $0.14, the judge less than $0.01, plus the conversations and the questions in Open WebUI (about $0.01 each, from the token counts). About $0.75, with the conversations TO BE COMPLETED at the rehearsal. Each pair's $3 share of the room's limit leaves a margin.

"About 3 times less" (Part 6, "Do the sum") is a rough figure. The $0.47 is what replay B showed at the rehearsal (2026-10-21); the $0.14 is an estimate at full price (2026-10-30, other instructions). The output tokens and the cache counts of mode B were not kept, so nobody can split the $0.47 between reading and writing. The page gives no price or lifetime for the cache: neither is in our measures. The saving comes only when the text comes back: one question of mode B alone showed $0.27 (measures/2026-10-31-image-38/), more than the $0.21 of its 103,865 tokens read at full price.


Leading the discussions

After Part 2 (and the overflow demo). "The demo refused three times the kit. Your group adds a document a week, about 3.5 cm of strip: when does it stop? And what do you do then?" Let the room find that the window has a maximum, and that every message pays for all of it.

After Part 3 (unplugged).

  1. "How did the search engine choose: by the words of the question, or by their meaning?" Most pairs look for words such as "outside the EU": the sentence of summary 90 says "beyond the EU".
  2. "Round 1, question 10: did the model pair see the contradiction?" Both sides fit in what the search engine pair passes: strips 11 and 22 in round 1, strips G and C in round 2. Our guess, not yet tried with a group: round 1 misses strip 11 more often, because "outside the EU" is on strip 1 (no hectares) and strip 11 starts with "Of these"; in round 2, strip G holds the words of the card and both figures of summary 10. Note at the rehearsal what the pairs passed.
  3. "Question 7: what did the search engine pass?" Strips about Campania, but not about 2025. This is "always something".

After Part 5. Read the two measured answers to question 10. Both see that the sources disagree. Pieces of 1000: only the totals arrive, so the model subtracts them ("2 242 195 − 1 079 538 = about 1 162 657 hectares", "my calculation, not a figure from the documents") and says "I can't tell which reading is correct from the excerpts alone." Pieces of 2000: one piece also holds "Non-EU European countries: 1,092,095 hectares" and "Middle East and North Africa: 70,562 hectares", so the model checks the sum and explains the disagreement ("The figures add up only if the 2.2 million includes the EU"). The difference is not the model: it is what the pieces contain. Do the sum on the board: 1 079 538 + 1 092 095 + 70 562 = 2 242 195, the exact total of summary 10. It confirms the 1 162 657, not which source is right: the answer of 2000 also says "I can't confirm this from the excerpts alone".

Question 3. It fails at every size. The pieces of document 43 come from its first page (in the base of 1000, only one piece, 5th, after a summary about Charente-Maritime). The document is in French and the question in English, and the embedding model of Open WebUI (all-MiniLM-L6-v2) was made for English (its model card says language: en). Tested on 2026-10-31: in French, 4 of the 5 excerpts come from document 40, and the two pieces with INTERDITS move from 61st and 63rd to 30th and 70th of 326. The language changes the result and does not explain everything. Whether the form of these pieces, lists of levels, also counts was not tested: say so.

Embeddings. Before the activity, draw two arrows on the board: the page and the deck explain the score by the angle between two arrows, and by the sum of the products of two lists. Open WebUI shows (1 + cosine) / 2, which is why "no link" is 0.5 here and about 0 in workshop 6. Do not explain the percentage shown when you click on a source: it is not this score, and we did not look for its formula. Part 4 now has a short activity: the pairs sort pairs of texts by closeness, then open the measured scores. Allow 5 minutes inside the 20 of Part 4, and cut question 3 to a quick look if the part runs late. The point to keep is that forbidden and allowed score 0.964: the search finds what a text is about, not what is true. The figures and scripts are in facilitator/atelier-2-measures/2026-10-31-embeddings/. This timing has not been tried with a group.


Exit card

  1. Who picks the pieces in a RAG, and when?
  2. Name one question where the size or the number of pieces changed the answer.
  3. Why is a RAG's failure harder to see than a failure of mode B?

What a good answer contains: (1) a search engine, before the model, from pieces cut in advance; (2) question 10 (pieces of 2000, or 10 excerpts); (3) it always gives excerpts, and the model may answer from the wrong ones with confidence, where mode B refuses to start when the knowledge is bigger than the model's window (the overflow demo).


Fallbacks

Problem What to do
Open WebUI not ready in a Codespace Parts 2 and 3 are its waiting time. After 1:00, do Parts 4 and 5 on your own Open WebUI on the big screen; the pair joins in when its log says READY.
The log ends with STOPPED The pair types openwebui start && openwebui sizes in a terminal.
The log says The search still finds nothing in … (or a base gives no excerpt) openwebui stop, then openwebui start && openwebui sizes. Since image 2026.10.33, one restart is already done by itself.
No Docker, or no Codespace The repository opens on a computer with Docker: dev containers guide. Otherwise, Part 4 and 5 on your screen.
replay RAG refuses (top-k not 5) openwebui top-k 5, then again.
Overflow demo: Haiku 4.5 refused (model not found, not allowed in the workspace) Do the strip alone, and read the message of the measures aloud. Never try two copies, nor Sonnet 5.5 with two to ten copies: the question would be answered and paid. If Haiku 4.5 has been retired for good, the same demo needs a model with a window of 1,000,000 tokens (claude-haiku-5-5) and twelve copies of the kit instead of three: the message then says limit 1000000, and you explain that the gauge says 200k because the workshop turns the larger window off in Claude Code. Not tried with the real API (only with a fake endpoint and eleven copies, 2026-11-04); fewer copies would be answered and paid.
Overflow demo: the chatbot answers, or the gauge moves above 0 One question was paid. Check that find /tmp/three-kits -name '*-copy*' | wc -l says 76, and that the command starts with WORKSHOP_HOME=/tmp/three-kits.
Open WebUI starts nowhere, not even on your screen Parts 2 and 3 as planned. For Parts 4 to 6, project the last slide of deck 2, Backup: what a RAG hands to the model (for question 10, bases of 1000 and 2000: which files the excerpts came from, which figures arrived, and the answers): slides of workshop 2, or the PDF you brought. No column RAG today.
A pair uses OpenCode with an OpenRouter key Same steps, Open WebUI and the judge included. Its totals are not comparable with the others (other model, other judge). replay RAG with mistralai/mistral-small-3.2-24b-instruct: 11 answers in 44 s. With the default model, DeepSeek V4 Flash: 11 answers in 48 s and 53 s, judge 15/20 then 17/22 (2026-11-02 and 2026-11-05, on a laptop; measures).

After the workshop

Reuse with young people. The unplugged activity needs no computer: strips of the group's own documents, a question, and a pair that plays the search engine. A version for 12–15, with shorter strips, rounds and points: the search engine game (source: facilitator/workshops/2-strips-youth.md, not tried with a group yet). The 2 m strip of the overflow demo also needs no computer and no key: the window on the floor, the documents as cards at 1 cm = 1,000 tokens (not tried with young people). To build a chatbot with a knowledge base for a group: Open WebUI guide, Part 2 - Build a chatbot, then Part 3 - Test before you open.