Public PDF to JSON with Page-Cited RAG Chunks

Convert text-based PDFs — from a public URL or pasted as base64 — into clean, page-cited JSON and deterministic RAG chunks ready for a vector database or an AI agent — without installing PDF libraries or paying for an LLM.

Why use it

• Citations built in — every chunk carries its page plus exact start/end character offsets, so answers can cite the page.
• Deterministic — the same PDF and settings always give byte-identical chunks (stable IDs for upserts), with a SHA-256 of the input.
• Predictable cost — one flat price per document, no per-token or per-page surprises.
• Safe by design — isolated parser, strict size limits, no network access.

Typical uses: RAG ingestion of public reports, filings, manuals and papers; agent tools that need "read this PDF" with citations; deduplication by hash.

Input

Field	Description
publicDataOnly	Must be true (see Responsible use).
pdfUrls[]	Public http(s) URLs of PDFs to fetch (see Fetching from URLs).
chunkSize, chunkOverlap	Optional defaults for URL PDFs: chunk size 250–4000 chars (default 1500), overlap 0–1000 (default 150, less than chunkSize).
documents[]	Optional pasted PDFs: objects with pdfBase64 (required, base64 of the PDF) and optional per-document chunkSize / chunkOverlap.

Supply at least one of pdfUrls or documents; 1–5 items in total per run (URLs + pasted documents).

Limits: 1 MiB decoded PDF, 1–20 pages, 200,000 text characters. Encrypted, malformed and image-only (scanned) PDFs are rejected uncharged — there is no OCR, table or image extraction. JSON input under 8 MB, output under 1 MB per document. Fixed 256 MB memory. Each document is parsed in an isolated child process (no Apify credentials, 160 MiB memory, 15 s CPU / 20 s wall-time limits). Nothing in the input is executed.

Example input (a public W3C test PDF URL) is prefilled in Console and in INPUT.example.json:

{"publicDataOnly": true, "pdfUrls": ["https://www.w3.org/WAI/ER/tests/xhtml/testfiles/resources/pdf/dummy.pdf"]}

Fetching from URLs: safety and limits

The Actor downloads only what you list, from the main process; the parser still runs in an isolated, network-free child process.

• Public http/https URLs only, ports 80 and 443 only, no credentials in the URL (user:pass@).
• The hostname is resolved and every resulting address must be publicly routable. Localhost, private (10/8, 172.16/12, 192.168/16), link-local/cloud-metadata (169.254/16), CGNAT (100.64/10), multicast, reserved and IPv6 loopback/ULA/site-local addresses are refused (URL_BLOCKED). The connection is pinned to the vetted address, so DNS rebinding cannot redirect it.
• Up to 3 redirects, each hop re-validated the same way. 10 s connect timeout, 20 s total per URL. No cookies, no authentication headers, no proxies.
• Responses are capped at 1 MiB (gzip transfer encoding accepted; the decompressed size is capped too, and the download is aborted once the cap is exceeded).
• Failed fetches (URL_BLOCKED, FETCH_FAILED, FETCH_TIMEOUT, TOO_LARGE, WRONG_CONTENT) are listed in OUTPUT.errors with documentIndex, the error code and the URL, and are never charged. Only successfully processed items are charged.
• documentIndex counts pasted documents first, then URL entries. Every output item has sourceUrl (null for pasted documents) and fetchedUrls; OUTPUT.fetchedUrls is the number of successful downloads in the run.
• Requests identify as ZendrenApifyActor/0.2 (+https://apify.com/zendren). Only fetch content you are permitted to process.

Output

One dataset item per processed PDF:

{"documentIndex": 0, "kind": "pdf", "sha256": "3df79d34…", "pageCount": 1, "characterCount": 14,
 "sourceUrl": "https://www.w3.org/WAI/ER/tests/xhtml/testfiles/resources/pdf/dummy.pdf", "fetchedUrls": 1,
 "pages": [{"page": 1, "text": "Dummy PDF file"}],
 "chunks": [{"page": 1, "start": 0, "end": 14, "text": "Dummy PDF file"}]}

The key-value store record OUTPUT summarises the run: processed and fetchedUrls counts, an errors list (with documentIndex, error code, message and, for URL items, the url) and stoppedForSpendingLimit. Invalid documents are reported there, are not written to the dataset and are not charged; valid siblings still process. Check OUTPUT even when the run status is Succeeded.

Pricing

Pay per event, no subscription, no platform-usage pass-through:

Event	Price	When
processed-document	$0.008	One PDF successfully processed and saved to the dataset
apify-actor-start	$0.00005	Once per run (Apify synthetic start event)

Example: 5 documents in one run = $0.04005. Invalid documents are free; the start event still applies to runs whose input is rejected. Set Maximum cost per run in Console or ?maxTotalChargeUsd= in the API — the Actor stops before parsing a document it cannot pay for.

Use it from the API, Make, n8n or an AI agent

curl -X POST "https://api.apify.com/v2/acts/zendren~public-pdf-json-chunks/run-sync-get-dataset-items?maxTotalChargeUsd=0.05" \
  -H "Authorization: Bearer $APIFY_TOKEN" -H "Content-Type: application/json" \
  -d @INPUT.example.json

Send the token in the Authorization header, never in the URL. The Actor is also usable as an MCP tool through Apify's MCP server (mcp.apify.com).

Responsible use

Supply only public data you are permitted to process, with no personal data. publicDataOnly is a policy acknowledgement, not an automated privacy detector. The Actor only downloads the exact public URLs you list (no crawling, no logins, no proxies) and never calls external AI services.

Support

Found a bug or need a format supported? Open an issue on the Actor's Issues tab. Fixture attributions: see NOTICE.md in the source.