Gdot-america

Every cropped image from the cropping run, page by page, with what the model read off it: view, objects, measurements, notes and labels. Click a crop to see it beside its JSON.

One set of distinct items across every page. The same bent read as “BENT”, “Interior bent” and “Bent No. 2” is one item here, carrying every page and crop it was seen in. One model call over all of step 1’s readings; the page mapping is resolved from its grouping rather than retyped by it.

Read the crops in step 1 first — there is nothing to group yet.

Every item from step 2 compared against the project’s engineering ontology, and then, for each one that lands in a class, a take-off of what that class asks for. The class list goes up as a closed vocabulary, so regional wording resolves without a synonym table — BENT to a pier system, PSC BEAM to a prestressed girder. Any class id that is not in the ontology is rejected rather than stored. The take-off reads every crop behind an item, derives what the drawings imply, shows its arithmetic, and says plainly when a number is not there.

Map the items in step 2 first — there is nothing to classify yet.

Everything steps 1 and 2 know, written to the vector store for a chat agent to answer from. Cut fine — a point for each object, each measurement and each note, not one per crop — with the page, the crop and the drawing written into every one, so a retrieved number arrives with enough context to be an answer. The readings as a whole and step 2’s items are indexed too, for the questions that are about a drawing rather than a figure.

Re-runnable at any time: each point is keyed to the record it came from, so re-indexing replaces what changed, adds what is new, and drops what the drawings no longer support. Re-read a crop or correct one by hand, then press this.

Nothing to index yet — read the crops in step 1 first.

Collection mindmap_vectors, filtered on projectId = __project_4.