Some questions are too big for one answer. “What most explains the difference between our regions?” could be prices, the product mix, returning customers or the season, and each needs its own look. /solve works such a question in parts. BumbleB Crunch™ splits it into smaller questions as it needs to, answers each one with its figures recounted, puts the answers back together, and brings the result to your chat. Where a part couldn’t be read, the answer says so instead of leaving it out.
This is in beta. Splitting a question into parts, and Dig deeper, haven’t yet been through the checks that a workflow’s figures have, so treat an answer as a careful first pass and read the figures that matter.
What you’ll need
- A chat with at least one dataset chosen. Click + and choose Choose datasets; see Datasets.
Start a solve
- In the message box, type
/solveand your question, such as/solve what most explains the difference in order value between regions. Without a question it says “Say the question after /solve.” - A card asks “Solve this by splitting it into parts?”. It repeats your question and, after Reads:, the datasets it will use. With none chosen it says Choose a dataset first and offers no Start.
- Set the budget after Up to. It starts at 400 units; a unit measures the model work the parts do, and this is the most the solve may spend. It takes a whole number from 1 to 100,000, and Start waits until it is one.
- Click Start, or Cancel to drop it. Nothing runs until you click Start.
Already have an answer you want to go further with? Click Dig deeper under it. The same card opens with that answer’s question.
Follow it in the chat
A card stays in the chat while the solve works. It shows its state, how many parts have answered and how many are running, and the units spent of your budget. Open the run shows the whole thing on its own page; Stop ends it.
On that page, the parts are listed as a tree, each indented under the question it came from, with a line joining it to that question. Above it, four tiles count the Budget, the Agents (answered, running and waiting on you), the Depth it has split to, and the Coverage: how many parts were read. Click a part to see its question, Why it split:, Closes when: (what would answer it), and its Share of units and what it spent. Stop this part ends that part and everything under it; the rest goes on without it.
A question splits at most three levels deep, into two to eight parts each time.
Answer the split gate
Small splits go ahead on their own. A split waits for you when its parts would use more than 100 units in all, or when it proposes a part your question didn’t ask for. The card (and the run’s page) then says which part “wants to split” and why, lists the parts with the units each would use, and marks assumed on any part no words of your question asked for. It ends with the total, such as “160 units in all, over the 100 a split may use without asking.”
- Approve the split runs the parts.
- Answer without it has that part answer as it stands, without splitting.
- Stop this part ends it; the rest of the solve goes on without it.
If the solve waits on something else, the card says Waiting on you with a link to answer it in the run.
Read the answer
When the solve ends, the card says “Bringing the answer back here…”, and the answer arrives in the chat as a new turn headed with your /solve question:
- a line for each finding, with its figures;
- Not read, listing each part that wasn’t read and why, for example because its share of units was spent;
- a line such as “3 of 4 parts read; 1 not read”. A part Crunch could read only some of counts as read and is said, as in “3 of 4 parts read (1 in part); 1 not read”. The Coverage tile and the run’s state count the same way;
- a Weakest link: line summing up the marks, such as “certified” or “certified, except 2 figures: 1 flagged, 1 not checked”.
Each figure is recounted independently as it’s found, and the answer carries the same marks Verify puts under a chat answer. Certified means the figure is the value of the count it states; where your question could be read another way, the marks say so.
If the solve was stopped or failed, the answer says that, with the reason and what it had found by then. Ask your next question as usual: the chat carries the solve’s findings.
Change one part: What if
Once a solve has ended, open the run, click a part and choose What if…. Change that part’s question and its Budget for this part (at most the share of the part above it). Crunch shows what is Changed, what Runs again, what Joins again, and what is kept as it was with no units spent again. Run the what-if starts a new run; the original stays as it was.
Good to know
- A part that is not read doesn’t fail the solve. The answer is built from what was read, and the Not read list says what is missing. A bigger budget usually reads more.
- When a solve ends with some parts unread, its state shows how many were read, such as “done · 1 of 4 read”, in amber rather than as a clean finish. A part read only in part keeps it amber too: “done · 4 of 4 read (1 in part)”.
- The marks say a figure was computed as stated, and that a different reading is said. They don’t promise every figure is right.
Related
- Workflows — the steps a solve runs in, and making your own.
- Check an answer’s numbers with Verify — what each mark means.
- The + and / menus — where Solve sits.
- Agents — a job you save and run again.
- Pieces — turn what a run found into a report or dashboard.