The Documents Corpus
The documents corpus is the PDF subset of the saudicompute.com source layer. It comprises 307 documents totaling over 2.3 million words and 11,659 PDF pages, with full-text search across the corpus. Documents are the highest-density single source class on the platform: a single PIF annual report or SDAIA strategy publication can carry the equivalent analytical weight of dozens of press articles, and the documents layer is therefore disproportionately represented in the platform’s analytical claims.
Documents are different from web sources in three ways. First, they are typically primary disclosures — strategy publications, annual reports, regulatory filings — that have been through the disclosing entity’s own internal review and that carry institutional weight that web sources rarely match. Second, they are version-controlled at the document level, with explicit publication dates, version numbers, and revision histories that make temporal analysis tractable. Third, they are structurally rich, with embedded tables, charts, footnotes, and cross-references that the platform’s extraction pipeline parses into structured records.
Document Categories
The documents corpus organizes around eight primary categories.
Sovereign and government strategy documents include the Vision 2030 framework documents, the National Transformation Program publications, the Saudi National Strategy for Data and AI (2020), successive SDAIA strategy refreshes, the Saudi Cloud Computing Strategy, and the broader Vision 2030 program-delivery plans. These are the highest-weight documents on the platform.
PIF annual reports and investor publications include the PIF annual report series (2017 onward), the PIF strategy refresh publications, and the broader investor-disclosure cycle. PIF documents are the primary source for the AI-attributable subset of the fund’s $930 billion AUM and the sovereign-capital architecture that anchors the buildout.
Aramco and SABIC primary disclosures include the Aramco IPO prospectus, the Aramco annual reports and earnings transcripts, the SABIC annual reports, and the broader sovereign-corporate-primary disclosure set. These documents are the source material for the energy-systems backbone behind the Saudi data-center fleet.
Ministry and regulator publications include MCIT publications, CST regulatory frameworks, SAMA cloud-computing guidance, CMA capital-markets publications, MoH health-data residency frameworks, and SDAIA’s PDPL implementing regulations and guidance.
US-side regulatory documents include the Federal Register publications relating to the AI Diffusion Framework, BIS license actions and policy publications, CFIUS publications relating to Saudi-counterparty transactions, SEC filings of US-listed counterparties, and the broader Treasury, Commerce, and State Department documentation that surfaces the bilateral architecture.
Partner-corporate filings include NVIDIA, AMD, Qualcomm, Microsoft, Alphabet, Amazon, and Oracle 10-Ks, 10-Qs, and 8-Ks where they surface Saudi-relevant disclosures; the broader investor-day materials; and the partner annual reports.
Conference materials include LEAP keynote materials, FII publication packages, US-Saudi Investment Forum communique documents, and the broader conference-disclosure document set.
Academic and analytical PDFs include peer-reviewed publications on Saudi AI policy, sovereign-wealth analytical reports, sell-side equity research that surfaces in PDF form, and the broader analytical-literature subset.
How Documents Flow Through the Platform
Documents enter the platform through the ingestion pipeline. Each document is captured at its source URL, OCR-processed where necessary (Saudi government PDFs are sometimes scan-based rather than text-native), parsed for structured content (tables, headings, dates, named entities), and ingested into the platform’s underlying database. The platform extracts five primary record classes from each document: full-text content, structured tables, named-entity mentions (entities, money, GPU, capacity), citations and cross-references, and metadata (publication date, version, source URL, capture timestamp).
The extracted records flow into the platform’s analytical layer through three paths. The full-text content is searchable through the documents index and is referenced from the per-entity and per-section landing pages. The structured tables surface in the /tables/ section landing and are also surfaced inline on the relevant analytical pages. The named-entity mentions feed the entity-profile pages and the intersection layer.
Notable Documents
A small number of documents are referenced repeatedly across the platform’s analytical layer.
The 2020 SDAIA National Strategy for Data and AI is the strategic anchor for the AI program. Its $20 billion commitment, four-pillar framework, and 2030 horizon are referenced across nearly every section landing.
The PIF annual report series is the primary source for the sovereign-capital architecture. The most recent annual report carries the AUM disclosure, the portfolio-allocation breakdown, and the strategic-direction commentary that anchors the capital section’s analytical layer.
The Vision 2030 program-delivery plans are the macroeconomic frame inside which the AI program sits. Their non-oil GDP targets, project pipeline disclosures, and KPI tracking are referenced from the geopolitics, capital, and timeline sections.
The Saudi Data Intelligence Nuclear Brief — a multi-hundred-page reference document covering Saudi data infrastructure, AI strategy, and adjacent national programs — is the densest single document in the corpus. It is referenced from the methodology, infrastructure, and policy sections.
The PDPL implementing regulations and SDAIA guidance series are the operational reference for the data-residency architecture. Every claim about cross-border transfer, personal-data classification, or cloud-resident processing flows from this document set.
The BIS Federal Register publications relating to the AI Diffusion Framework are the operational reference for the Tier-2 silicon access regime. The publications, alongside the underlying license actions, are the source material for the silicon and policy sections.
How to Use the Documents Index
The documents index supports three primary use patterns.
Reference look-up: a reader who wants to verify a specific claim on the platform can trace it back through the in-page citation to the underlying document, retrieve the document from the index, and read the relevant passage in original form.
Topic exploration: a reader who wants to develop deep familiarity with a topic — say, the Saudi PDPL framework — can pull the document subset relevant to that topic from the index and read across the documents in order. The platform’s editorial layer is designed to be a navigation aid into the documents, not a substitute for them.
Comparative analytical work: a reader doing comparative analysis across sovereign-AI programs can pull the equivalent document classes from the platform’s tracked corpus and run their own comparative reading. The PIF annual report alongside the Mubadala annual report alongside the Temasek review supports a comparative sovereign-capital analysis that the platform’s analytical layer also runs at the structured-data level.
Updates and Cadence
The documents corpus is updated continuously. New documents are ingested as they are published; document refreshes (a new PIF annual report, a new SDAIA strategy version) trigger downstream analytical-layer refreshes. The platform maintains the historical document set so that temporal-analysis use cases are tractable: a reader can pull the 2020 SDAIA strategy alongside the most recent SDAIA strategy and read the trajectory of the program over the intervening period.
Documents are not retracted from the corpus. Where a document has been retracted at its source, the platform retains its capture and flags the retraction at the document-record level. This is a deliberate editorial choice: temporal analytical work depends on having the original document available even when the source has subsequently retracted or amended it.
Full-Text Search Across the Corpus
The 11,659 PDF pages in the documents corpus are indexed for full-text search. Subscribers and readers can search across the corpus by keyword, phrase, named entity, date range, document class, or any combination thereof. The search interface returns the matching pages in original PDF form alongside the surrounding context, with the document’s provenance metadata attached to each result.
Full-text search is the foundation under several common analytical workflows. A researcher tracing the evolution of a specific concept across PIF annual reports can search for the keyword across the report series and assemble the temporal trajectory. A compliance analyst tracking the propagation of a specific PDPL provision across SDAIA implementing regulations can search across the regulation series. A policy researcher surveying Saudi-government-side framings of a specific bilateral relationship can search across the broader government-publication subset.
Document Lifecycle and Re-Ingestion
Documents enter the corpus once and are then maintained on a continuous basis. New versions of documents (e.g., a new PIF annual report each year, a new SDAIA strategy publication on its release cadence) are ingested as new records, with the prior versions retained for temporal-analytical work. The relationship between versions is captured in the provenance-metadata layer, so that a reader can navigate from the prevailing version to its predecessors directly.
The re-ingestion workflow handles three classes of update. Major-version updates (e.g., a strategy refresh, an annual-report cycle) trigger full ingestion of the new version with the prior version retained. Minor-version updates (e.g., an addendum or correction to an existing document) trigger an addendum record attached to the original. Errata (e.g., a typographical correction in a previously-ingested document) trigger a metadata-only update with the original full-text retained.
OCR Quality and Edge Cases
A meaningful share of the Saudi government PDF corpus is scan-based rather than text-native, particularly for older documents and for documents originally produced in the print-publication workflow. The platform’s ingestion pipeline applies OCR to scan-based documents and produces a searchable text layer alongside the original image. OCR quality varies; the pipeline applies confidence-scoring at the page level and surfaces low-confidence pages for editorial review.
Arabic-language OCR is structurally more challenging than English-language OCR, particularly for older typesetting conventions and for documents with complex layout. The platform’s pipeline uses Arabic-aware OCR engines and applies post-processing for the regional typesetting conventions that occur in the Saudi government corpus. The Arabic-language full-text search is therefore available across the relevant subset of the corpus, with the same confidence-scoring discipline as the English-language subset.
How to Search Effectively
Effective search across the documents corpus benefits from a small number of practices. Starting with named entities (Humain, NVIDIA, SDAIA) typically returns higher-precision results than free-text concept searches. Combining a named entity with a date range narrows the result set to operationally relevant documents. Filtering by document class focuses the search on the specific source type that is analytically relevant. The platform’s search interface supports each of those practices and surfaces query-refinement suggestions where the initial query returns a noisy result set.
Citation Discipline
The platform’s citation discipline ties every analytical claim to a document or web source where applicable. Within the documents layer specifically, citations surface in three forms: in-page citation that points to the document and the relevant page, footnote-style citation in long-form analytical pieces, and per-record provenance metadata in the underlying data layer. The discipline applies uniformly across the platform’s analytical output.
The citation discipline supports two reader workflows. Audit-trail workflow lets a reader trace any analytical claim back to its underlying document; the platform’s commitment is that any claim is auditable in this way. Source-pursuit workflow lets a reader who has read the analytical claim navigate to the underlying document for further reading; the platform’s commitment is that the documentent is available in the corpus and is reachable through the citation.
Bilingual Document Handling
The Saudi government documents corpus includes a meaningful fraction of dual-published documents (Arabic and English versions), Arabic-only documents, and English-only documents. The platform’s ingestion pipeline handles each pattern through the appropriate processing path. Dual-published documents are paired in the corpus so that readers can navigate between the language versions; Arabic-only documents are processed with Arabic-aware OCR and full-text indexing; English-only documents follow the standard processing path.
The bilingual handling is operationally consequential because primary Saudi government publications are increasingly dual-published, with the Arabic version typically the legally-authoritative version and the English version the practically-consumed version for the platform’s primarily English-reading audience. The platform’s editorial layer cites the English version for accessibility while preserving the Arabic-version provenance metadata for legal-authoritative reference.
Document Storage and Access
The 307 documents in the corpus are stored in a structured archive that pairs the original document file with the extracted full-text, the structured-data extracts (tables and named-entity mentions), and the provenance-metadata layer. Subscribers at the institutional and enterprise tiers of the Sovereign Compute Terminal can access the original document files directly through the documents index; analyst-tier subscribers and public readers can access the extracted full-text and the analytical layer that consumes it.
The storage architecture is engineered for durability and reproducibility. Document captures are preserved in their original form; extracted derivatives are versioned alongside the originals; provenance-metadata records are immutable once created. The architecture supports the platform’s commitment to long-horizon analytical reproducibility: a claim made in 2026 should remain auditable against the original document captures decades later.
For deeper reading:
- Sources — the broader source-domain layer of which documents are a subset
- Tables — the structured-data extracts surfaced from the documents corpus
- Methodology — how documents flow into the SCS framework
- About — platform mission and editorial standards