The specs are in the PDFs. The PDFs are on the website. Nobody can search either one.

The situation

Every answer a dealer needs is locked inside a PDF

The manufacturer makes industrial valves and fittings sold through a network of 40+ dealers. Product pages on the website show a photo, a model number, a one-sentence description ("High-performance butterfly valve for industrial applications"), and a link to download a PDF spec sheet. The spec sheet contains the actual data — bore sizes, pressure ratings, material options, end-connection types, torque values, and compatibility notes. But that data only exists inside the PDF.

When a dealer needs to quote a job, they download three or four spec sheets, open each one, and manually compare dimensions and pressure ratings. When a procurement AI agent tries to find compatible products, it hits a product page with no structured specs and a PDF it cannot reliably parse — tables in the PDF are formatted inconsistently across product lines because each line came from a different acquisition, and the PDFs were never standardized.

The internal sales team has the same problem. A rep fielding a call about which valve fits a specific application has to open PDFs one at a time or rely on tribal knowledge about which model numbers map to which specs. New reps take months to learn the catalog because there is no searchable, structured reference — just 500+ individual documents.

The work

What Taz actually does with this catalog

Taz processes every product page and its linked PDF spec sheet. The extraction is systematic and covers the full catalog, not a sample.

• Spec extraction from PDFs: Every spec sheet is parsed and its data extracted into structured records — dimensions, pressure ratings, material grades, temperature ranges, end-connection types, torque values, weight, and compliance certifications. Where PDF tables are inconsistent (merged cells, footnotes embedded in headers, specs split across pages), the extraction handles each layout and flags ambiguities for human review.
• SKU normalization: Model numbers, naming conventions, and product-line hierarchies are standardized across the catalog. "BV-400-SS-150" and "Butterfly Valve 4in Stainless 150#" become the same record with a canonical name, a sortable model number, and all legacy aliases preserved. Duplicate and superseded models are identified.
• Comparison table generation: Within each product family, comparison tables are built from the extracted specs — sortable by bore size, pressure class, material, or connection type. A dealer looking at 4-inch butterfly valves can see every option side by side instead of opening six PDFs.
• Compatibility mapping: Where spec sheets note compatible actuators, gaskets, or mounting hardware, those relationships are extracted into a structured compatibility index. "BV-400 is compatible with actuator models ACT-200 and ACT-350" becomes a queryable record, not a footnote on page 2 of a PDF.
• Agent answer pack: A structured JSON catalog with every product spec, comparison table, and compatibility record. Sourced entirely from the extracted PDFs and verified product pages. This is what a procurement AI or dealer portal would use to answer "show me all 6-inch valves rated for 300 PSI in 316 stainless."

The deliverable

What the manufacturer receives

• Structured product database — Every SKU in a normalized table: model number, canonical name, all extracted specs, product-line membership, status (active/superseded/discontinued), and a completeness score per record. Exportable as CSV or JSON.
• Comparison tables by product family — Ready-to-publish HTML tables for each product line, sortable by the specs dealers actually filter on. Includes a recommended "quick spec" summary for each product page.
• Compatibility index — A cross-reference of which products work with which accessories, actuators, and mounting hardware. Structured as a queryable dataset, not a document.
• Implementation ticket list — Prioritized changes to product pages: which specs to surface on the page itself, which PDFs have data quality issues to resolve, and where comparison tables should be linked. The manufacturer's web team works through the list without interpretation.
• Agent answer pack (JSON) — Machine-readable product specs, comparison data, and compatibility records. Every value traces back to a specific PDF and page number.