Mira

Answering from your own documents

Intro to AI Engineering. Live learners, anonymised answers from everyone who took this before, and partners Mira plays from those same patterns. You can see which is which.

In the room

1 live5 from the archive2 played by Mira
MMarcoliveSSamMiraPPa practice partnerMiraSTsomeone on a support teamarchiveENan engineer who has never trained a modelarchiveTLa team lead buying a chatbotarchiveHPsomeone who writes the help pagesarchiveARan analyst at a retailerarchive

Live poll

38 answers · stacked live, archive, and Mira

A chatbot needs to know your returns handbook. What's the move?

“We'll fine-tune it on the handbook. That's how you teach a model things, isn't it?”

a team lead buying a chatbot · 21 found this useful

“Don't train it. Look the page up at the moment of asking and put it in front of the model. Training takes hours and is already wrong if the policy moved this morning. Lookup takes milliseconds.”

an engineer who has never trained a model · 54 found this useful

Vote to add yourself to the live tally.

Breakouts to join

Mira seats you with someone holding a named wrong idea — a live peer, an echo replayed from the archive, or a partner she plays herself.

AI-facilitated1 seat left

Talk Sam out of training the handbook in

Sam thinks “we need to train the model on our data”

MiraMira facilitating
Join the call
Echo replay1 seat left

Replay: it is just searching for matching words

an analyst at a retailer thinks “it searches for matching words”

archiveMira facilitating
Join the call

From the archive

Comments, votes, and the wording that actually landed — anonymised, and still in the room.

TL
a team lead buying a chatbotarchivewrong idea5 days ago

“We'll fine-tune it on the handbook. That's how you teach a model things, isn't it?”

ST
someone on a support teamarchivequestion4 days ago

“Our refund policy changed this morning. Do we retrain, or is there a cheaper way?”

EN
an engineer who has never trained a modelarchiveexplanation4 days ago

“Don't train it. Look the page up at the moment of asking and put it in front of the model. Training takes hours and is already wrong if the policy moved this morning. Lookup takes milliseconds.”

HP
someone who writes the help pagesarchive3 days ago

“That's why I stopped asking engineering to 'retrain on the new page'. I just make sure the page exists. The lookup does the rest.”

AR
an analyst at a retailerarchivewrong idea2 days ago

“So it searches for the same words? 'Can I send this back' won't find a page that says 'returns within 30 days'.”

HP
someone who writes the help pagesarchiveexplanationyesterday

“It matches on what the text is about, not the wording. Those two sit near each other on the map even though they share no words. That's the whole trick.”

ST
someone on a support teamarchivewrong ideayesterday

“Smaller pieces are always better, right? More precise.”

HP
someone who writes the help pagesarchiveexplanationyesterday

“Cut to a single sentence and 'this does not apply to sale items' arrives with nothing to say what 'this' was. Each piece has to stand on its own.”

TL
a team lead buying a chatbotarchivewrong idea12 hours ago

“Grab more pieces to be safe. More context is more to go on.”

EN
an engineer who has never trained a modelarchiveexplanation12 hours ago

“Extra pieces are not neutral. Grab ten and the careers page turns up. Now a refund answer mentions job openings.”

AR
an analyst at a retailerarchiveexplanation8 hours ago

“Sort, then stop. That's the same move as keeping the top five channels — score everything, line it up, take the ones at the top.”

EN
an engineer who has never trained a modelarchiveexplanation6 hours ago

“I'd reject a grab of ten pieces. Careers will sneak in. If the cut is a sentence, 'this' arrives with no subject. Those two notes are the whole review.”

HP
someone who writes the help pagesarchiveexplanationyesterday

“Don't train it. Look the page up. Match on what it's about, not the wording. That's the sentence I keep using in stand-up.”