Machine-Spec Pipeline
A source-grounded machine-spec dataset behind a product configurator: an LLM pipeline that lets the model emit a value only if it can quote the source, enforced by the same check in the writer and the auditor.
The problem
A product configurator needs accurate machine specs, but the data is scattered across manufacturer sites and PDFs, and an LLM will confidently invent a plausible number. The dataset has to be trustworthy, not just complete.
What it does
- Extracts specs (weight, flow, capacity, dimensions) from manufacturer sources with an LLM.
- Lets the model emit a value only if it can quote the source; otherwise the field stays empty.
- Runs the same grounding check in the write guard and the audit.
- Publishes honest coverage per category, not a flattering single number.
The core idea: grounding as an invariant
Quote or null. The model may only emit a value it can quote from the source; if it can't, the field stays empty, and the same check runs in the write guard and the audit, so there's one definition of 'grounded'. The gold-standard categories, Skid Steer and Compact Track Loader, reach 98.6% core-field fill across 172/172 models and are cross-checked against a second source; overall readiness is a published 34%, not a rounded-up claim.
How a gap gets filled
Missing specs don't wait on a research project. A request form captures the gap and, with Jira connected, files it as a ticket; an agent then picks it up, opens the manufacturer source, extracts the fields under the same quote-or-null rule, and writes back only what it can ground, with the ticket as the audit trail. Filling a hole in the dataset becomes a request the pipeline closes itself, and anything it can't ground stays open for a human rather than being guessed.
how it fits together
Extract, then prove every value
LLM extraction with a grounding gate: no value is kept unless a real source quote backs it.
- Catalog spine78 brands / 11 categories / 1,314 models in scope.
- Batched discoveryOne web session per (manufacturer, category), not per model, to keep extraction cost down.
- LLM extractionPulls weight, flow, capacity, and dimensions, each value carrying its source text.
- Grounding gateThe model may only emit a value it can quote; ungrounded values are nulled, never written.
- Provenance-tagged dataset628 models with data, 4,447 verified specs, each tagged with its source.
- Streamlit dashboardHonest coverage by percent of required fields filled; requests are logged and filed as Jira tickets when connected.
product screens
key engineering decisions
Quote or null, never guess
No value is written without a verbatim quote from its source; ungrounded values are nulled. Data-lineage governance implemented in code.
One check for write and audit
The write guard and the audit share a single grounding implementation, which is what makes it an invariant rather than two checks that can disagree.
results & outcomes
- Feeds a product configurator specs it can trust, because no number is stored unless the source is quoted, so a customer never sees an invented figure.
- A spec request is captured and, with Jira connected, filed as a ticket an agent then fulfills, scraping the source and grounding each value, so closing a data gap is a request, not a manual research project.
- Publishes honest readiness per category, so the business knows exactly which product lines are ready to configure and which aren't.
- One grounding check shared by the writer and the auditor, so 'verified' means the same thing everywhere and can't quietly drift.