Tuning Quote Generator

Quote Generator is tuned by its tenant, not by support. This article lists every lever an administrator can move, where it lives, what it ships as, what it changes, and what to look at afterwards to tell whether the move helped.

Before you begin

You need administrator access. The Admin entry appears in the sidebar only for administrators, and a box a tenant can never use is not drawn at all, so your Admin page may be shorter than this article.

Which Lever Changes Which Step

A change applies to quotes built after it

Nothing already on a quote is re-scored, re-seeded or re-priced when you move a lever. To see what a change does, build a new quote from the same documents and compare.

Model Binding

Open Admin → Operations → Prompt Management. Under Knobs, the Model binding card holds six settings and says what it is for: “how the model binds document lines to catalog parts, and what it may spend doing it”.

Every setting works the same way. Type a value, click Save, and the row is badged overridden. Click Reset to default to put it back. A value outside its range is refused before anything is written, and the refusal names the range.

Setting	Ships as	Range	What it changes
Confidence flag threshold	0.7	0 to 1	The bar under which a bind is marked “worth a closer look”. A flag, never a filter.
Tie margin	0.05	0 to 1	The gap a tie’s winner must clear over the runner-up before it is chosen without you.
Searches per line	4	0 to 12	“Catalog searches the model may make for one line before it must decide.”
Lines per model pass	12	0 to 60	“Document lines one model pass works through before handing the rest to the next pass.”
Drafted lines per unit	40	0 to 80	“The most lines the model may propose for one unit when it drafts a parts list.”
Services per opportunity seed	12	1 to 40	The most services one opportunity seed may propose. Shown only where prices are given rather than built from parts.

Confidence flag threshold: what to watch

Raising it flags more binds; lowering it flags fewer. It removes nothing either way, so it cannot make a quote wrong. What it decides is how much of the quote asks for a second pair of eyes. Build a quote and count the flagged lines: if almost every line carries the flag, the bar is above what your catalog can support and you are training estimators to ignore it.

Searches per line: what to watch

This is the spend lever. Each search is a round trip, so lowering it caps what one difficult line can cost, and raising it gives the model room to prove a part is absent rather than guess. Read the declines on the AI Record afterwards. A decline that names what it searched is worth more than a bind the model was pushed into.

Lines per model pass: what to watch

A pass is a batch, not a second opinion. A line decided in one pass leaves the queue and is not looked at again, so every pass after the first is deciding lines the first never saw. A small number means a long package runs many passes, and every pass boundary costs time that produces no decisions.

This setting also caps how many lines one quote can reach

A chain runs at most eight passes, so the ceiling is eight times this number: at the shipped 12 that is 96 lines, and a package with more than that leaves the rest undecided. On large document packages, 24 is the measured setting: it halves the number of passes and doubles the ceiling, and it was tested against the time one pass is allowed with room to spare. Most quotes never run a second pass and are unaffected.

What to watch. After raising it, check that a large build decided every line it extracted, and that the build did not get slower. Raise this on its own: moving it and Searches per line in the same change makes the result unreadable.

The Preset Seeding Floor

On the same page, below the Model binding card. It is a whole number entered as a percentage, and the field says what it is: “% match required to load a preset BOM”.

The card states the rule: “A unit only loads its matched preset’s BOM when the match scores at or above this floor”. Below the floor the unit arrives bare with its extracted specifications and waits under Needs attention as Units with no parts list yet, where the preset stays a suggestion under Change base.

What to watch. Lower it and more units arrive carrying a parts list nobody chose. Raise it and more units arrive bare for the model to draft. Build the same package at two settings and compare how many lines came from the preset against how many the model drafted. The AI Record names the source of every line.

Prompt Blocks

Same page, under Prompt blocks. Five pieces of prose reach the model as it reads your documents: honesty, structural_focus, line_focus, not_a_line_item and unit_framing. The last two are shared by every stage of the read. Edits are audited and confirmed, and a reset returns the block to the composed default for your tenant’s shape.

Prose is the weakest lever on this page

Telling the model to be careful changes how it words an answer far more reliably than it changes what it decides. Where a rule has to hold every time, it belongs in the catalog, a kit or a preset, not in a paragraph.

Products: Families, Presets and Quote Defaults

Admin → Products. A family is the shape of a thing you sell and its Specifications are its fields. A preset pins some of those fields to real values and carries the parts and labor a proven design starts from.

• Families and Specifications decide what the extraction can recognise on a page. A specification nothing declares cannot be read off a document.
• Presets decide what a matched unit arrives carrying. They are versioned: every edit creates the next version, and a quote keeps the version it was built with.
• Preset mining proposes presets out of the quotes you have already built.
• Quote default values, on a family, is what the proposal prints for every unit of that family. A value set on the unit itself always wins, so per-quote edits keep working and quotes already built do not move.

What to watch. After changing a family default, build a quote and preview the PDF rather than reading the field back. The default is inherited by new units only.

Scope of work: what the proposal prints under a unit

The scope a unit is performed under reaches the customer’s paper directly, and it is composed rather than typed as one block. A section is a heading, an introduction and any number of bullets, and a section may have any of the three: a heading with bullets and no introduction is a section, and a blank field is not stored, so it prints nothing rather than an empty line.

It is written in two places, and they are the same editor:

• On a family, under Quote default values, as the wording every unit of that family starts from.
• On the quote, on the unit’s own card. The control is collapsed and says what it holds: “Scope of work” with the number of sections, or “No scope of work” when there are none.

Each section has a Heading and an Introduction above its bullets, and the fields say what they are for: “Heading — e.g. ELECTRICAL SPECIFICATIONS (optional)”, “A sentence or two under the heading (optional)” and “One line of scope”. Sections can be reordered with Move up and Move down, and removed.

A unit that has nothing of its own shows the family’s wording and says so: “Showing {the product family}’s wording. Editing here writes a version for this row only”. Removing every section from a unit gives it back the family’s.

On paper, consecutive units that resolve to the same scope print it once rather than repeating it per unit, and the block says what it covers. Up to three units it names them, “Applies to {A, B} and {C}.”; beyond that it counts them, “Applies to the {N} units above.”. A single unit gets no such sentence, because it would only repeat the name printed above it.

A unit whose scope differs breaks the run

The fold is over units standing next to each other, so changing one unit’s scope splits the block around it. If you expect a fold and do not get one, compare the wording of the units either side, and check that none of them is hidden from the customer.

The older one-paragraph form still works

A tenant whose template binds a single scope paragraph keeps printing it, and clearing the sections does not delete that paragraph. The two revert independently.

What to watch. Author the family wording once and preview the PDF rather than reading the field back. Then edit one unit on one quote and preview again: a unit’s own scope always wins, so a family edit never rewrites a quote somebody has already built.

The Catalog, Kits and Aliases

Admin → Catalog. This is the lever with the largest effect on binding, and the least glamorous one.

• Parts and services is what the model searches. A part that is not here cannot be bound to, however the request is worded.
• Kits let one line arrive as an assembly with its pieces. Where a document asks for something your shop builds from several parts, a kit is the answer, and the model can choose one.
• Matching aliases is a library of learned equivalences, gathered from your own confirmations and your own documents. An alias raises candidates at match time. It never binds a part and never prices a line.
• Manufacturers and Manufacturer tiers decide which brands are preferred where several parts would fit.
• Spec database and Enrichment queue are where a part’s specifications get filled in, and specifications are what the model matches against.

What to watch. Count the declines on the next build. A decline that proves the catalog does not hold the part is a catalog task, and it is the clearest signal here of work worth doing.

Pricing

Admin → Pricing. None of this changes what lands on a quote. All of it changes what the lines cost.

• Margin tiers: each tier names a margin and a quote picks one. Exactly one tier is the default for new quotes.
• Allowances: an amount per size, for a cost that is not built from lines. A size with no entry stays unpriced and says so. Nothing is interpolated.
• Formulas and Settings and constants: the math, and the named numbers the formulas read.
• Contingency, Generic costs and Fee standards: the additions that are not parts.

What to watch. A total that looks low is usually a quote saying it is incomplete. Look for unpriced lines before you look at the margin.

People and Roles

Admin → Operations → People, which describes itself as “Users, named roles, and who is allowed to sign quotes.” Permission roles say what someone may do; quote roles say who is named on a quote and whether they sign; signing entities say who the paper is from. Being named on a quote grants nothing on its own. The full treatment is in Quote Roles and Who Signs.

Finding a Setting Without Knowing the Map

The Admin page has a search. Type what you would say out loud rather than the name of the box: threshold, seeding, margin, who can sign. Each result names the box the setting lives in and opens it directly. A setting your tenant cannot use is not offered, so the search never sends you to a page that would refuse you.

Troubleshooting

I changed a setting and the quote looks the same

A change applies to quotes built after it. Existing quotes are never re-scored or re-seeded. Build a new quote from the same documents.

Save is disabled, or the page refuses my value

The value is outside its range, or is not a whole number where one is required. The message names the range and what the setting bounds. Nothing is written when a value is refused, so the setting keeps the value it had.

A setting in this article is not on my page

A setting a tenant’s shape can never use is removed rather than explained. Parts settings do not appear where quotes are priced as given services, and Services per opportunity seed appears only where they are.

I do not know which lever to move

Read the quote first. Reading the AI Record says which hand placed the line you are unhappy with, and that decides which lever is the right one.

Related Articles

• How a Quote Gets Built
• Reading the AI Record
• Optional Units and Lines
• Quote Roles and Who Signs