AI for CAD imports

The module AI CAD BoM import (ai_mrp_cad_bom) helps with the part of a CAD import that the rules cannot do. A drawing that only has position labels (“R1 / 100x100x4”, “PM 2 / 300x300x10”, “Bolt / M 16”) tells you which members exist but not how many there are or how long they are, because the same member is labelled in several views. The AI reads the drawing again, including its dimensions, and suggests values that you accept or reject. It installs itself when the AI module and the CAD import are both installed.

Important

The AI never writes a quantity, length, section or product on its own, and it never creates products or bills of materials. A suggestion changes the import only when you accept it, and the weights, the totals of the two options and the material summary change only then. Products and bills of materials are still created by the normal flow of the CAD import.

Requirements

  • A chat model is configured (see Providers and models).

  • You are in the group AI: User, the capability Ask is on for you (see Who can use what) and you may edit CAD imports (Manufacturing / User).

  • The import is Parsed or Reviewed.

Analyse an import

Open a parsed import and click Analyse with AI in the form header. The analysis runs in the background, so you can keep working. The form shows the progress (it refreshes every few seconds) and you are notified when the suggestions are ready or when the analysis failed. Only one analysis runs for an import at a time, and while it runs you can neither parse the import again nor accept or reject suggestions.

The job runs in your own environment: your language, your company, your access rights. It goes through these steps and commits each one separately, so a failure keeps the suggestions of the steps that were finished and shows the error under AI analysis failed: and in the chatter.

  1. Reading the drawing. The drawing is read again for its geometry (see below).

  2. Interpreting the lines. The AI checks every line: its kind, family, designation, grade, whether it is a purlin and a short description. It decides the ambiguous AxBxC (plate or hollow section), reads the general notes of the drawing (grade, bolt class, standards) and explains conflicts such as the same mark with different sizes.

  3. Estimating quantities and lengths. For the lines that came from labels, the AI proposes the number of pieces and the length of one piece (or the width and length of a plate), with its reasoning and the evidence. A quantity or a length that is written on the drawing is never estimated again. A line whose quantity and length are both stated is not sent to the model at all. When only one of the two is stated, it is passed to the model as known, and the answer cannot change it. Lines that were read from a parts list or from block attributes are not estimated.

  4. Matching products. Lines without a product get a suggested existing product, chosen among a few candidates that the server selected.

  5. Checks. Rule-based checks look for gaps (see below). This step does not use the AI.

  6. Review note. A short note in your language with the totals now and the totals if you accepted every suggestion, the main materials and the checks. It is shown on the tab AI review and posted in the chatter.

Every answer is parsed as JSON and checked. Whatever is not valid is dropped: unknown lines, products that were not offered as candidates, designations the parser cannot read, negative values and views that do not exist.

How the drawing is read

The geometry step collects the dimensions of the drawing (linear and aligned dimensions of model and paper space), splits each space into views and sheet frames, and finds the title of each view (SECTIUNE, PLAN, DETALIU, SECTION, DETAIL, METSZET, ALAPRAJZ, …), the cut markers and the sheet numbers. The scale of a view comes from its dimensions: a drawing made at 1:10 with a dimension scale factor of 10 is detected as such, and lengths are given to the AI in millimetres.

For each line the AI receives its labels per view, with the position of each label, the leader line that touches it, the segments the leader points to (the member), the nearest lines and the nearest dimensions. The AI is told that the same piece is labelled in several views, so pieces are counted once (usually from the plan) and lengths are taken from sections and details.

Review the suggestions

The tab AI review appears after an analysis. It has:

  • the AI review note and the Drawing notes read by the AI;

  • the list Suggestions to review: for each line the current quantity and length, what the AI proposes (AI quantity, AI length (mm)), its Confidence, the reasoning, the evidence (views and values) and a note; and

  • the list Checks.

Use the buttons Accept and Reject on a line (the ticks in the list of lines do the same), Accept all above the threshold and Reject all. Run the checks again repeats the checks without the AI, for example after you corrected a line. Suggestions can be accepted or rejected while the import is parsed or reviewed.

Accepting a suggestion applies:

  • a changed designation (the line is parsed again: family, kind, dimensions, plate size), otherwise the kind;

  • the grade (it is resolved through the material master, so “OL37” becomes S235JR), the description and the purlin flag;

  • the quantity, and the length of a linear member or the width and length of a plate when missing; and

  • the suggested product, as if you had chosen it by hand.

The line’s source becomes AI suggestion and the confidence is stored. A line whose confidence is below the threshold stays flagged To review after you accept it. The threshold is the field AI confidence threshold on the tab (0.6 by default, from the settings), and Accept all above the threshold accepts only suggestions that reach it.

A product that you accept from the AI, or that you choose by hand on an analysed line, is remembered as an alias, so the next import finds it without the AI.

screenshot: manufacturing-cad-import-ai-review
menu
Manufacturing app --> CAD imports --> (an import) --> AI review
shows
The AI review tab with the review note, the suggestions list with the proposed quantity, length, confidence and reasoning, and the checks list.
highlight
The confidence column and the Accept and Reject buttons (red frames).
data
The staircase import after an analysis, with one low-confidence suggestion.
module
ai_mrp_cad_bom
notes
English UI, light theme, 1440px width.

Checks

The checks never change data. They report: bolts without nuts or washers; a plate with no bolt or hole label near any of its labels; purlins without sag rods or tie rods (information); a missing grade; open conflicts, with the AI’s explanation; a linear member without a length and without an estimate; a piece longer than the stock bar (information); a line whose weight is far above the others (more than eight times the median and more than 40 % of the total); and the number of suggestions below the threshold (information). Each check is a warning or an information.

Quick prompts

On a CAD import the AI panel (see The AI panel) offers the quick prompts Explain this import (in place of the generic Summarise), What still needs review? and Material totals by type. The panel sees the lines, the material summary and the checks of the import, as far as your own access rights allow.

Choose the model

The analysis is interactive: you wait for the suggestions. Use a fast chat model with a low cost per call (see Providers and models).

  • Model variants that run as a batch (names ending in :batch) do not suit this analysis, because their answers arrive much later.

  • Very slow reasoning modes can take longer than the timeout of the provider. If you use one, raise the Timeout of the provider (see Configuration).

Cost, limits and data sent to the model

  • Every call goes through the normal AI machinery: the capability Ask, the rate limit, the blocked patterns and the AI log with its cost (see Monitoring). In the background job the per-minute rate limit is waited for instead of failing, but the daily limits stop the job at once.

  • The size depends on the drawing. The analysis of a staircase drawing with 260 labels, 17 groups of members and 787 dimensions took seven calls and about 22,000 input tokens. A very large drawing is limited to twenty estimation requests (roughly 100,000 tokens); the lines beyond are reported as not estimated.

  • The model receives the line data and texts of the drawing, the general notes, the geometric context (numbers, view titles and sheet numbers), the totals and check messages for the note, and, for matching, the names and internal references of the candidate products. It receives no customer, project or import name and no other business data.

Settings

In Manufacturing app ‣ Configuration ‣ Settings, block CAD BoM import, under AI suggestions:

  • AI confidence threshold: estimates below it stay flagged to review and are left out of Accept all above the threshold. Default 0.6.

  • Analyse with AI after parsing: queues the analysis automatically after each parse when label lines need a review and you may use the AI; otherwise it is skipped silently. Off by default.