Ask ARGUS
Asking questions in your own words, checking the answers, following up, letting the assistant propose changes you confirm, and using it on the phone.
The detailed manual is currently available in English.
Ask ARGUS answers questions in your own words, in any language, from the workspace’s equipment, tickets, documents, the text of attached files and this guide.
Ask
- Open Ask ARGUS and ask: “Which ion pumps are on the linac?”, “What went wrong with the gun vacuum before, and how was it fixed?”, “How do I export a workspace?”.
- You see the lookups as it works (Searching equipment, Reading what is written, Opening a record). Each record key in the answer opens its record.
- Follow up in the same conversation (“which of those are in AC1?”): earlier turns are its context. New chat starts afresh. Conversations are kept, for you only.
An answer built from no lookups is marked as such: do not treat it as coming from the workspace’s records. When something is not found, it says what it looked at.
Check an answer
Open the linked records and check that they say what the answer says, about the right equipment, at the right revision or date. Lookups do not make every conclusion correct.
Not found can mean the information is not there, that you may not read it, or that it has not been indexed yet. In a follow-up, give asset keys, dates and the part of the machine.
Rate an answer
Under each answer, 👍 says it was useful; 👎 asks, optionally, what was wrong or missing. Each answer is rated once. The rating is kept without your name: the question, the answer, the workspace, the model, the ARGUS version, the client and the lookups it made (not what they returned). An answer rated not useful becomes an Argus Knowledge ticket for the people who improve Ask ARGUS (the ARGUS Knowledge feedback workspace, seen only by them and administrators).
In the comment, say what is wrong, the records that show it and what it should have said. Leave out secrets and unrelated personal information.
Let it make changes for you
Ask it to create, change or relate records: “Create screen station SCN01, composed of camera … and motor …”. It proposes; it never changes anything by itself.
- It looks up the records and checks its proposals: the type exists, the records exist, the attributes belong
to the type, the relations are allowed. New records it proposes are called
new:1,new:2… so they can be related in the same answer. - Proposed changes appears under the answer, with a box for each.
- Apply all of them or the ones ticked, or Discard. Applying goes through the same checks as the forms, your rights, and your name in the history.
- Each then shows applied (with a link to the record), failed (with the reason) or discarded. A relation that needs a record whose creation failed fails too.
You are only offered what your rights allow you to do by hand.
If only some changes succeed
Read each result and open what was created or changed. Correct the ones that failed and retry only those: do not create again what already succeeded. When you apply a subset, include the creations the selected relations need.
Ask it how to do something
“How do I import an EPIK8s configuration?”, “step by step, how do I move local workspaces to production?”: it answers from this guide, one step at a time if you like. The Ask ARGUS to walk me through this links in Help start the same.
On the phone (ARGUS Field)
Ask is a tab of its own in the mobile app.
- Type a question, or press the microphone and say it: when you pause, what was heard is sent as the question.
- You see the lookups while it works. Answers are formatted (lists, bold, tables), and the record keys they cite open the record. Follow-ups continue the conversation; New conversation starts afresh; Earlier conversations reopens past ones, the web’s included.
- The speaker under an answer reads it aloud; 👍 / 👎 rate it (see Rate an answer).
- Hands-free (the headset icon): a spoken conversation. See Hands-free below.
Hands-free
The orb at the bottom shows what ARGUS is doing, in movement and shape as well as words: a slow glow while it waits (Ready when you are), rings that follow your voice while it listens (Listening), traces running into the core while it looks things up (Looking it up…), waves leaving the core while it answers (ARGUS is answering), and a still, angular shape with a message when something went wrong.
- What you are saying shows under the orb as it is heard; it is sent when you pause.
- Answers are spoken short: the essential in a sentence or two. The whole answer and its sources stay on the screen. A longer answer ends with Do you want the details?: say yes for the rest, or ask something else.
- An ambiguous question gets one clarifying question back.
- Say repeat (ripeti), slower (più lentamente), show the source (mostra la fonte) or stop.
- Silence does not end it: it keeps listening, and after a few minutes of nothing it waits for a tap on the orb. End or stop ends it; so does leaving the screen.
- To interrupt an answer, tap Interrupt or the orb. Interrupting by speaking is not possible: while the phone speaks, its speech recognition cannot listen without hearing the phone itself.
- A problem says which it is (ARGUS unreachable, microphone not allowed, speech not recognised) and offers Retry; the conversation is kept.
- With the phone’s reduce motion setting on, the orb is drawn still, with an icon for each state.
Spoken questions are turned into words by the phone’s speech-recognition service (on Android usually Google’s, on iOS Apple’s), which may process the audio under its own terms; only the recognised words go to ARGUS. Answers are read aloud by the phone. The first time, the phone asks to allow the microphone. Check equipment codes in what was heard before relying on the answer.
More than questions, on the phone
The + beside the question box offers:
- Register equipment from a photo: the camera opens on the nameplate, the form is filled from it, and you check and register (see The mobile app (ARGUS Field)).
- Find or register by its label: scan a QR code or barcode, read a printed serial, or type it; if nothing carries it, register a new unit.
- Write a document by voice: pick the document type and record it; the workspace’s speech-to-text model transcribes it and the assistant writes it up with the type’s template (see Dictate a document in The mobile app); you review the draft and what to check; the transcript is kept with it, the recording only if you choose.
- Write a document from a photo: photograph a page; its text is read on the phone; the photo is kept with the draft.
When it is unavailable
Ask needs the workspace’s AI endpoint to be set up and checked; searching what is written by meaning also needs the knowledge index. Writing a document by voice needs a speech model. See AI settings.