PDF Evidence Graph API

Turn public, text-layer PDF documents into deterministic, citation-ready JSON for AI agents, retrieval pipelines, audits, and document-processing systems.

This Actor accepts only public HTTPS PDF URLs. It extracts normalized page text, stable evidence spans, native bookmarks, document metadata, and link relationships. Each evidence span has a reproducible citation identifier tied to the exact PDF bytes, page number, span index, and text hash.

Why use it

• Stable citations that can be stored and checked later.
• Predictable JSON for API, MCP, RAG, and automation consumers.
• Page-level and span-level hashes for provenance checks.
• Native PDF structure only; it does not invent headings.
• Outbound links are reported but never followed.
• One bad document does not discard successful results from the same run.

Input

{
  "urls": ["https://example.org/public-document.pdf"],
  "include_page_text": true,
  "include_links": true,
  "include_outline": true
}

Submit between 1 and 20 unique public HTTPS URLs. The Actor deliberately does not accept authentication material or custom request headers.

Output

Successfully processed document records are written to the default dataset with exactly one workload-banded value event. Failed document records are stored separately in the ERRORS key-value-store record and are never written as billable dataset items. The OUTPUT record contains the run summary and delivery count, not a free duplicate of successful value records.

Key fields include:

• document_id, binary_sha256, and content_sha256
• normalized pages with page hashes
• deterministic evidence_spans with character offsets and citation IDs
• native outline entries, where present
• internal and outbound link_graph edges
• bounded, customer-safe error codes

Example evidence span:

{
  "citation_id": "pdf:v1:<binary-sha256>:p0001:s0001:<text-hash-prefix>",
  "page_number": 1,
  "span_index": 1,
  "character_start": 0,
  "character_end": 142,
  "text": "Extracted evidence...",
  "text_hash": "<sha256>"
}

Determinism and limitations

The same PDF bytes and product version produce the same document identity, page hashes, citation IDs, spans, native outline, and link graph. Runtime duration and timestamps are operational metadata and are not deterministic.

The MVP supports PDFs with an embedded text layer. OCR, scanned-image interpretation, passwords, signed download links, non-standard ports, private-network targets, and target-site authentication are intentionally unsupported. Some PDFs encode visual reading order ambiguously; extracted order follows the PDF text representation. The Actor does not provide legal verification or claim that document content is true.

Security and privacy

• HTTPS only, port 443 only.
• DNS resolution and every redirect are checked against non-public network ranges.
• Download count, size, redirect, page, text, link, and runtime bounds are enforced.
• No browser, proxy network, paid API, external AI model, or persistent copy of input files is required.
• PDF actions and embedded JavaScript are not executed.
• Outbound links found in PDFs are never visited.
• Temporary files are removed after processing.
• Full input contents are not written to application logs.

Pricing

This public Store Actor uses pay per event. The final validated event prices are:

• document-small: $0.005, up to 10 pages and 2 MiB.
• document-medium: $0.020, up to 100 pages and 10 MiB, excluding small documents.
• document-large: $0.050, all other supported documents within the 250-page and 25 MiB hard caps.

Exactly one workload-band value event is associated with each successfully delivered document. Failed documents do not generate a document-processing charge. Platform usage is included in the event price. A small synthetic Actor-start event may appear once per run under Apify's platform accounting; it is not a document-processing event.

Support

Report reproducible failures using the public source URL, the safe error code, and the Actor version through the Apify Issues workflow. Never send access credentials, private URLs, signed links, or private documents. The public production Actor remains limited to public HTTPS, text-layer PDFs and the documented resource and security limits. itations that can be stored and checked later.

• Predictable JSON for API, MCP, RAG, and automation consumers.
• Page-level and span-level hashes for provenance checks.
• Native PDF structure only; it does not invent headings.
• Outbound links are reported but never followed.
• One bad document does not discard successful results from the same run.

Input

{
  "urls": ["https://example.org/public-document.pdf"],
  "include_page_text": true,
  "include_links": true,
  "include_outline": true
}

Submit between 1 and 20 unique public HTTPS URLs. The Actor deliberately does not accept authentication material or custom request headers.

Output

Successfully processed document records are written to the default dataset with exactly one workload-banded value event. Failed document records are stored separately in the ERRORS key-value-store record and are never written as billable dataset items. The OUTPUT record contains the run summary and delivery count, not a free duplicate of successful value records.

Key fields include:

• document_id, binary_sha256, and content_sha256
• normalized pages with page hashes
• deterministic evidence_spans with character offsets and citation IDs
• native outline entries, where present
• internal and outbound link_graph edges
• bounded, customer-safe error codes

Example evidence span:

{
  "citation_id": "pdf:v1:<binary-sha256>:p0001:s0001:<text-hash-prefix>",
  "page_number": 1,
  "span_index": 1,
  "character_start": 0,
  "character_end": 142,
  "text": "Extracted evidence...",
  "text_hash": "<sha256>"
}

Determinism and limitations

The same PDF bytes and product version produce the same document identity, page hashes, citation IDs, spans, native outline, and link graph. Runtime duration and timestamps are operational metadata and are not deterministic.

The MVP supports PDFs with an embedded text layer. OCR, scanned-image interpretation, passwords, signed download links, non-standard ports, private-network targets, and target-site authentication are intentionally unsupported. Some PDFs encode visual reading order ambiguously; extracted order follows the PDF text representation. The Actor does not provide legal verification or claim that document content is true.

Security and privacy

• HTTPS only, port 443 only.
• DNS resolution and every redirect are checked against non-public network ranges.
• Download count, size, redirect, page, text, link, and runtime bounds are enforced.
• No browser, proxy network, paid API, external AI model, or persistent copy of input files is required.
• PDF actions and embedded JavaScript are not executed.
• Outbound links found in PDFs are never visited.
• Temporary files are removed after processing.
• Full input contents are not written to application logs.

Pricing intent

The intended Store model is pay-per-event, charged atomically with storage of a successfully parsed document. Workload events are document-small (up to 10 pages and 2 MB), document-medium (up to 100 pages and 10 MB), and document-large (within the 250-page/25 MB hard caps). Failed documents produce no value event. Prices remain a pre-publication hypothesis until a zero-cost owner-approved Apify container benchmark measures actual platform usage.

Support

Report reproducible failures using the public source URL, the safe error code, and the Actor version. Never send access credentials. This development version is local and unpublished.