An estimating assistant where the model names the parts and code sets the prices

A proposal goes in as a PDF or Word document; payment terms and conditions are cut out before the model sees it. The model extracts into a strict schema and is told to copy numbers exactly and never compute them. Code then rebuilds the bill-of-materials tree, reconciling every line to its parent, every parent to its section and every section to the subtotal and the ten per cent tax. A document that does not reconcile is flagged for review rather than quietly saved, and a person confirms before anything is stored.

The model reports indentation, and code trusts the resulting tree only if every parent equals the sum of its children. If it does not (a bundle shown both whole and in parts, a per-unit assembly), the tree is rebuilt by recursive descent from the sums alone. Hours are derived from labour dollars only when the result lands on a sensible step and reproduces the printed figure within one per cent.

Each job is embedded and a new brief is matched on a hybrid score: weighted cost-driver parameters, semantic similarity of the scope (falling back to keyword overlap when no embedding key is set) and overlap of equipment categories. The closest three are returned with a breakdown and plain-English reasons, so the estimator can see why a past job was chosen.

The model turns the brief into sections and components, each marked new, retained or removed, and then chooses parts, but only from a list of candidates drawn from the price book. Anything it returns that is not on that list is dropped. A deterministic function does all the pricing: price-book lookup times quantity, hours times the rate card. No price the model wrote ever reaches an estimate.

Each priced line carries its basis (price book, matched job, rate card, ratio estimate or 'needs price'), with a confidence level and the source job, shown as a badge. Low-confidence rows are tinted, and a part nothing can price goes in at zero with a warning rather than at a guess. Cabling, consumables and freight are ratio estimates from the best-matching section, and the demo says so on screen.

The scope, assumptions and exclusions text is drafted last, with an instruction to mention only equipment that is in the estimate and to list what it mentions. Those mentions are checked against the lines, and any that do not match raise a warning; the model's description is audited by the same data it describes.

One real job is held out of the library, the price book and the knowledge layer entirely. A comparison view sets the draft against that quote section by section, shows what is missing from the draft and what the draft has that the real quote did not, and compares labour hours.