Supplier Catalog Extractor

1. What are you trying to achieve?

Turn a supplier catalog into records you can trace back to the exact source, review, and pass to a sourcing decision. Start with the changed document, not a guessed company match.

2. User intents

Primary intent: prepare versioned supplier, model and specification candidates. Secondary intents: review changed fields, locate missing units and compare catalog revisions. This is not supplier qualification, a binding quotation, or proof of origin.

3. Inputs

The proposed adapter accepts an authorized PDF, HTML snapshot or existing extraction output, together with a document identifier, capture time when known, version and target fields. A Website Change Monitor review packet can identify which source and fields need attention. This page does not upload or fetch documents.

4. Outputs

The proposed output is a review bundle: candidate model records, raw field text, normalized values when justified, original units, source locations, unresolved conflicts and a revision relationship. Extraction confidence and human review state are separate. A missing value stays missing.

5. How it works

Choose an existing extraction engine for the document type. Preserve its output and source coordinates; map fields to the catalog schema; flag uncertain entities and units; ask a reviewer to accept or reject candidates. Only reviewed records may enter a future capability graph. These execution steps are proposed.

6. Example workflow

Illustrative case: a fictional AX-12 revision changes payload from 5 kg to 7 kg. Keep both excerpts and document versions, associate each value with its page or element, and flag the change for review. Do not conclude that a specific mission can carry 7 kg or merge AX-12 with another supplier model.

7. Integration Stack

Integrate document conversion and web extraction where they already exist. Benchmark a managed layout service against the same labeled catalog set. Keep source evidence separate from the recommended execution stack; no provider is connected to Smart Tools in this release.

Integrate

• Docling — Candidate for an initial local PDF adapter. PROPOSED
• Unstructured — Alternative when an existing document pipeline already uses its elements. PROPOSED
• Diffbot Product API — Candidate for suitable web product pages rather than scanned engineering catalogs. PROPOSED

Benchmark

• Azure Document Intelligence — Compare managed output against the same labeled pages before selecting deployment.

Evidence sources

P1 provider documentation and release records. Supplier origin/capability claims need their own evidence.

Do not rebuild

Do not rebuild general OCR, PDF layout analysis, commodity web crawling or a provider-specific product parser. Add the supplier/model review and decision contract around an existing engine.

Smart Tools owns

Requirement context, provenance, reviewed entity mapping and downstream decision routing.

8. Build / Integrate / Benchmark Matrix

Compare local document processing, document partitioning, managed layout analysis and web product extraction. These options solve different input problems. Selection depends on document layout, language, deployment constraints and measured field quality; there is no universal accuracy ranking here.

Market Reality

9. What Smart Tools owns

Why Smart Tools exists: a parsed table still does not say which supplier and model it belongs to, which revision applies, what is uncertain, or how it affects a buyer requirement. Smart Tools should own the reviewed relationship between those facts and the next decision, rather than a replacement OCR engine.

10. Works best when

Start with a small, authorized catalog set that has recognizable model identifiers and reviewable source pages. Define required fields and a reviewer before choosing an adapter. A useful pilot includes clean tables, merged cells, scanned pages and conflicting revisions.

11. Not suitable when

Not suitable for inferring absent certifications, assuming a distributor is the manufacturer, guessing units, merging similar model names, or silently publishing confidential drawings. Use the original source when the document cannot support a reliable field location.

12. Technical architecture

Proposed architecture: controlled intake → document/version fingerprint → selected extraction adapter → immutable raw output → field mapping → conflict and entity review → approved export. Keep parsing workers separate from the website. Record provider and model versions so a rerun can be compared.

13. API / SDK / OSS options

Docling, Unstructured and Diffbot are integration candidates; Azure Document Intelligence provides a managed-layout comparison. The downloadable contract defines candidate fields and review states, not a working Smart Tools extraction endpoint. No credentials or paid calls are required to read this specification.

14. Data / privacy / deployment

Keep confidential catalogs private by default. Before execution, choose local or hosted processing, retention, reviewer access and deletion behavior for that document. The current page receives no catalog content; anonymous page/link events exclude documents, excerpts, source URLs and API keys.

15. Industrial use cases

Manufacturing: preserve part/model codes, dimensions, units and catalog revisions. UAV: distinguish a vehicle, payload and configuration rather than inheriting a whole-company claim. Composites: record a stated material or test method without treating an equipment list as a measured material result.

16. Internal links

Previous: review a meaningful source change with Website Change Monitor. Next: hand reviewed candidates to Supplier Capability Matcher; it remains planned. RFQ Requirement Parser is a separate planned input path for buyer requirements. Use the workflow map to inspect availability before continuing.

17. Cross-site routing

Sites provides company and source context; BE PASSIVE provides original catalog access and a contextual requirement route. The BYTE source gallery is an existing engineering experience, not a live demonstration of this extractor. A dedicated extraction Demo and cross-site completion measurement remain unavailable.

18. External evidence / references

The references below are official documentation reviewed on 2026-10-08. They establish documented provider capabilities, not integration tests, tenant entitlement or measured accuracy on this catalog set. Each external reference keeps its direct URL and a stable tracking ID.

1. Docling official documentation

   Document conversion, table structure and JSON export; local execution is documented.

   P1 · 2026-10-08 · EV-SCE-DOCLING
2. Unstructured official documentation

   Partition documents into typed elements; PDF table extraction has strategy-specific options.

   P1 · 2026-10-08 · EV-SCE-UNSTRUCTURED
3. Azure Document Intelligence official documentation

   Layout analysis returns document text, tables and source-region information.

   P1 · 2026-10-08 · EV-SCE-AZURE-LAYOUT
4. Diffbot Product API official documentation

   Extract product identifiers, specifications and other data from commerce product pages.

   P1 · 2026-10-08 · EV-SCE-DIFFBOT-PRODUCT

19. SEO

Search intent: evaluate supplier catalog extraction and field-level provenance before implementation. English and Traditional Chinese pages have reciprocal language links and distinct canonical URLs. Original company sources remain their own pages; this specification does not replace them.

20. Structured data

This page is a TechArticle with official citations and breadcrumb data. The JSON spec exposes the product status, proposed contract, integration comparison, related routes and acceptance gates. It does not advertise a working SoftwareApplication or fabricated ratings, prices or extraction results.

21. Analytics / tracking

Available measurement covers specification views, downloads, official references and next-step clicks. Execution, extraction quality, useful-result feedback and RFQ completion are unavailable until the corresponding functionality is connected. A reference click is not a successful extraction or inquiry.

22. MVP

This release delivers the bilingual product specification, proposed field contract, integration research and semantic handoff links. Next implementation slice: select one adapter and review ten authorized catalog pages against a labeled field set. Do not start several provider connectors at once.

Product specification and bilingual page

AVAILABLE

Provider adapter and catalog execution

PROPOSED

Private review storage and reviewer authorization

PROPOSED

Supplier capability matching and graph writes

PLANNED

Dedicated extraction Demo and RFQ outcome

NOT_CONNECTED

23. Acceptance criteria

Execution acceptance gates: every accepted field has a source locator; missing values and zero remain distinct; units are preserved; duplicate models and contradictory versions remain separate until reviewed. Test rotated scans, merged tables, malformed output, timeouts, duplicate callbacks and access-controlled deletion. Publish measured coverage and unresolved cases, not an unsupported accuracy target.

24. Product graph position

Supplier Intelligence: monitor → catalog candidates → capability matching. Manufacturing Source-to-Quote can reuse the reviewed catalog with a separate buyer requirement. Shared product nodes retain one identity across suites; this page adds one specification, not another set of product counts.

1. Supplier Catalog Extractor · SPECIFICATION_AVAILABLE;PROPOSED:catalog-extraction
2. previous: Review the source change first · AVAILABLE
3. next: Inspect the capability-matching handoff · PLANNED
4. alternative: Start from buyer requirements instead · PLANNED
5. evidence: Explore the company evidence network · AVAILABLE
6. evidence: Open the original BYTE catalog library · AVAILABLE
7. commercial: Discuss a component requirement · AVAILABLE
8. experience: Explore the existing engineering source gallery · AVAILABLE

25. Next related product

Continue to Supplier Capability Matcher only after supplier/model identity, required field provenance and buyer requirements are available. If those inputs are missing, return to the source or prepare the RFQ input. The next product remains explicitly planned.