PDF Text Extractor — RAG Chunks & Page Links

Turn public PDFs into text chunks for AI search and document Q&A, with links back to the original pages. Supports PDFs with selectable text; no OCR.

Each chunk retains its PDF page, exact character offsets and stable identifiers. Full page text and extraction reports are included.

No model key or embedding service is needed. This Actor extracts existing text; it does not perform OCR, summarize documents or reconstruct tables.

Quick start

1. Enter up to ten direct public HTTPS PDF links and choose the page and chunk limits.
2. Leave Maximum cost per run at $0.10. Pricing is $0.005 per complete PDF ($5 per 1,000); the ten-document input limit means at most $0.05 in document events per run. This version requires an explicit event cap no greater than $0.10.
3. Start the run. Under Full document output, download the OUTPUT JSON to check document statuses. Use Chunks for table viewing, export or indexing.

Example input:

{
  "urls": ["https://www.w3.org/WAI/ER/tests/xhtml/testfiles/resources/pdf/dummy.pdf"],
  "maxPages": 50,
  "chunkSize": 1500,
  "chunkOverlap": 100
}

Use direct file links without query parameters. Login-protected, signed and private-network URLs are unsupported.

Outputs

Output	What it contains
Chunks	Default dataset of chunks from complete, successfully charged documents. Export these rows as JSON or CSV.
Full document output	Canonical OUTPUT JSON in the run's key-value store: documents, pages, chunks and extraction statuses. Includes available partial results for inspection.
Run summary	Final SUMMARY JSON with document outcomes and the run's billing totals.

The full-output and summary views list their JSON file with a download control. They do not mix billing-recovery records into the results list.

Each chunk contains text, sourceUrl, citationUrl, pageNumber, documentId, chunkId, charStart and charEnd. For example, the sample PDF produces this text and citation:

{
  "text": "Dummy PDF file",
  "pageNumber": 1,
  "citationUrl": "https://www.w3.org/WAI/ER/tests/xhtml/testfiles/resources/pdf/dummy.pdf#page=1",
  "charStart": 0,
  "charEnd": 14
}

Chunks overlap but never cross a page. Offsets count Unicode code points in the emitted page text after Unicode/newline normalization. They are not byte offsets in the PDF. Identifiers are deterministic for the same PDF bytes and extracted chunks; they do not deduplicate charges across separate runs.

For an indexing workflow, use the chunk's text and retain the citation, page number and identifiers as metadata. The Actor does not call a vector database or an AI model.

Pricing and incomplete runs

One event is charged only after a complete PDF with usable text is saved and read back from OUTPUT. Duplicate bytes within the same run, failed downloads, empty-text documents and partial extraction are not charged. Platform usage is included in the listed event price; there is no startup or per-chunk event fee.

The Actor stops before fetching another document if the remaining event budget cannot cover it. A successful run can therefore contain fewer documents than requested. Check document counts and billing totals in Run summary.

Full output is saved before a charge is attempted. Dataset export happens at the end. If a run is interrupted, Chunks or Run summary may be missing or incomplete while Full document output remains available. Use that canonical record to recover delivered results. Uncertain billing or dataset writes stop processing; they are not blindly retried. Starting a separate run is a new billable extraction.

Limits and data handling

• At most 10 unique URLs, 10 MiB per PDF, 100 pages per PDF and 100 pages across the run.
• At most two million emitted page characters across the run. Additional page, chunk and output limits can produce an explicit partial result.
• Public HTTPS on port 443 only. No cookies, custom headers, proxies, passwords or access-control bypass. Private addresses and unsafe redirects are refused.
• Text-layer extraction only. Image-only pages may need another OCR tool. Existing OCR text and PDF reading order can be wrong; successful extraction does not certify the document's accuracy or visual layout.

Outputs are stored in the run's Apify storage. No document content is sent to an AI provider. Process only documents you are allowed to use. Extracted text remains untrusted source material for downstream agents, including passages that look like instructions.

Help

Use this Actor's Issues tab for problems. Include the error code and a public sample URL only if you are allowed to share it. Do not post credentials, private documents or signed links. Keep the affected run's output available while the issue is reviewed.