P2M AI Feeder – PDF to Markdown Converter

What does P2M AI Feeder do?

P2M AI Feeder converts uploaded PDFs and public PDF URLs into clean Markdown, deterministic chunks, optional image descriptions, and optional structured Docling JSON. Its fast defaults read native PDF text and use fast table extraction. Opt-in local OCR handles scans and sufficiently large bitmap text, while a bundled local vision model can turn photos, diagrams, figures, and charts into searchable descriptions for AI agents, retrieval-augmented generation (RAG), search indexing, and document-analysis pipelines.

The Actor performs OCR, image understanding, and table extraction locally with Docling. It does not send document content to an external OCR, LLM, or vision API. Running it on Apify adds API access, schedules, integrations, run monitoring, and durable run-scoped storage.

Why use P2M AI Feeder?

• Convert native-text PDFs, scanned pages, embedded screenshots, diagrams, charts, tables, and multi-column layouts through one input contract.
• Replace unhelpful <!-- image --> markers with locally generated descriptions of detected pictures.
• Enable local OCR for scans and text-bearing bitmap regions when the native text layer is insufficient.
• Consume chunks directly from the default dataset or retrieve complete artifacts from the key-value store.
• Preserve heading context in each chunk for better retrieval and agent grounding.
• Process large PDFs in bounded page ranges while reusing one isolated Docling worker instead of reloading its models for every range.
• Persist each completed page range immediately, preserve usable Docling partial output, and report failed ranges without discarding the rest of the PDF.
• Automatically retry a timed-out enrichment range without image descriptions and then without table structure before recording a page-level failure.
• Isolate per-document errors by default, or stop immediately with fail-fast mode.
• Protect the downloader with file-size limits, PDF signature checks, bounded retries, redirect validation, and blocking of private network targets.

How to convert PDF files to AI-ready Markdown

1. Open the Actor's Input tab.
2. Upload one or more PDF files, or paste public HTTP(S) PDF URLs.
3. Leave OCR and Describe images off for ordinary digital PDFs. Enable OCR for scans and image-based text; enable image descriptions only when semantic visual content is required.
4. Enable Force full-page OCR only when a scan has a broken or misleading hidden text layer.
5. Adjust chunk size and overlap for your model's context window.
6. Click Start.
7. Read document and chunk rows in the dataset, or open the complete files from the run output.

No browser, proxy, or external API key is required.

Input

See the Input tab for all configuration options. A typical API input is:

{
  "sources": ["https://example.com/sample-report.pdf"],
  "doOcr": false,
  "forceFullPageOcr": false,
  "describeImages": false,
  "imageDescriptionPrompt": "Describe the image for retrieval. Extract important readable text and explain the main visual structure. Do not guess.",
  "extractTables": true,
  "tableMode": "fast",
  "pageBatchSize": 6,
  "chunkSizeChars": 6000,
  "chunkOverlapChars": 500,
  "includeMarkdownInDataset": false,
  "includeChunksInDataset": true,
  "exportDoclingJson": false,
  "maxDoclingJsonPages": 50,
  "maxPages": 250,
  "maxFileSizeMb": 50,
  "doclingTimeoutSeconds": 540,
  "workerTimeoutSeconds": 600,
  "failFast": false
}

chunkOverlapChars must be smaller than chunkSizeChars. workerTimeoutSeconds must be at least 30 seconds greater than doclingTimeoutSeconds, giving Docling time to serialize cooperative partial output before the subprocess is stopped. The older documentTimeoutSeconds input remains a compatibility alias for the hard worker timeout.

Full Markdown is always stored as a file; includeMarkdownInDataset only controls whether it is duplicated into the document dataset row.

OCR and image descriptions are different

• doOcr recognizes readable characters in scanned pages and sufficiently large bitmap regions. The conservative area threshold avoids sending tiny icons and decorations through CPU OCR.
• forceFullPageOcr replaces the native text layer with OCR on every page. It is useful for scans with a bad hidden text layer, but it is slower and can reduce accuracy on normal text PDFs.
• describeImages uses the bundled SmolVLM-256M model to explain a picture's visible content and relationships. It is disabled by default because CPU inference can dominate runtime. Small pictures are skipped, descriptions are capped at 100 new tokens, and the generated description is included in Markdown (and Docling JSON when that export is enabled) instead of returning only an image placeholder.
• imageDescriptionPrompt lets you focus descriptions on labels, numbers, workflow relationships, chart values, or other domain-specific details.

OCR and vision descriptions are machine-generated. Verify critical numeric, medical, legal, or financial information against the original document.

Output

The dataset starts with a document summary and, when enabled, follows it with chunk rows:

[
  {
    "type": "document",
    "documentId": "a1b2c3d4e5f607182930",
    "fileName": "sample-report.pdf",
    "status": "partial",
    "pageCount": 24,
    "convertedPageCount": 22,
    "imageCount": 6,
    "describedImageCount": 6,
    "chunkCount": 9,
    "failedPageRanges": ["13-18"],
    "warnings": [
      {
        "code": "PAGE_RANGE_FAILED",
        "message": "Pages 13-18 could not be converted after bounded attempts."
      }
    ],
    "markdownUrl": "https://api.apify.com/..."
  },
  {
    "type": "chunk",
    "documentId": "a1b2c3d4e5f607182930",
    "chunkIndex": 0,
    "headings": ["# Sample report", "## Findings"],
    "text": "# Sample report\n\n## Findings\n\nExample content..."
  }
]

You can download the dataset in various formats such as JSON, HTML, CSV, or Excel. Complete Markdown, chunk JSON, durable page-range Markdown, and optional per-range Docling JSON use the Apify-compatible document- key prefix. The OUTPUT key is updated after every processed range and contains the current manifest, part URLs, warnings, and completed-page progress.

Data table

Field	Type	Description
type	string	document, chunk, or error row
documentId	string	Stable SHA-256-derived identifier for the PDF bytes
status	string	succeeded, partial, or failed
pageCount	integer	Source PDF page count
convertedPageCount	integer	Pages that produced usable output
imageCount	integer	Picture elements detected by Docling
describedImageCount	integer	Pictures enriched with a local SmolVLM description
chunkIndex	integer	Zero-based chunk position
headings	array	Active Markdown heading hierarchy
text	string	Agent-ready Markdown chunk
markdownUrl	string	URL of the complete Markdown artifact
chunksUrl	string	URL of the complete chunk JSON artifact
doclingJsonUrl	string or null	URL of structured Docling JSON when enabled
warnings	array	Non-fatal timeout and controlled-degradation notices
failedPageRanges	array	Page ranges unavailable after bounded fallback attempts
error	string	Sanitized per-document failure information

How much does it cost to convert PDFs?

This Actor uses transparent pay-per-event pricing based on the conversion work that produces usable output:

Event	Price	When it is charged
Actor start	$0.0025	Once per memory-scaled Actor start event
PDF page processed	$0.008 per page	After a successfully converted page batch has been saved as Markdown
Image described	$0.015 per image	Only for a non-empty local vision-model description saved in the Markdown

Dataset document and chunk rows are not the billing unit, so changing chunkSizeChars or enabling chunk rows does not increase the conversion price. A 10-page PDF with no described images costs approximately $0.08 plus the memory-scaled start charge. Enabling image descriptions adds $0.015 for each image that receives a saved description.

The Actor checks the user's maximum run charge after every persisted page batch. If the limit is reached, it stops before starting more conversion work and keeps all completed page-range artifacts available. For predictable spend, test a representative PDF first and leave describeImages disabled for text-only collections. The Apify Console estimate is authoritative for the memory-scaled start charge and any future pricing-tier discounts.

Tips and advanced options

• Keep image descriptions disabled for text-only PDFs.
• Keep OCR disabled when PDFs have reliable embedded text; enable it for scans or text-bearing images.
• Use forced full-page OCR only for scans or broken hidden text layers.
• Customize the image prompt for diagrams, charts, product screenshots, or scientific figures.
• Use fast table mode for simple, high-volume documents.
• Set extractTables to false for plain-text documents that do not need row-and-column reconstruction.
• Enable accurate table mode only for documents whose complex tables justify the extra CPU time.
• Use smaller pageBatchSize values (2–4) for scanned or unusually slow PDFs. The default of 6 is intended for ordinary digital or mixed documents.
• Enable Docling JSON only when downstream consumers need the full layout/provenance model. When disabled, the Actor skips building and transferring that structure. It is also skipped when the PDF exceeds maxDoclingJsonPages; Markdown and chunks continue normally.
• Keep full Markdown out of the dataset for large PDFs and follow markdownUrl instead.
• Disable chunk dataset rows when another service will retrieve chunksUrl.
• Use asynchronous Actor runs for large PDFs; synchronous API callers may time out before Docling finishes.
• doclingTimeoutSeconds is a cooperative limit for one page range. workerTimeoutSeconds is the larger hard boundary that can terminate image enrichment or another stage that does not stop cooperatively. Download requests are separately capped at 120 seconds.
• The Actor logs a heartbeat every 30 seconds and reports converter construction versus conversion/export timing for each page range. Docling initializes most models lazily, so the first range can include model startup.
• Multiple PDFs and ranges remain sequential to protect memory. The persistent worker retains a compatible Docling converter between ranges; after a hard timeout, the next attempt starts a clean worker.
• CPU inference threads follow APIFY_DEDICATED_CPUS when available. On Apify, 8 GB currently corresponds to 2 CPU cores, matching this Actor's default memory allocation.

For a fast diagnostic run, use one small PDF with OCR, image descriptions, and Docling JSON disabled:

{
  "sources": ["https://example.com/sample.pdf"],
  "doOcr": false,
  "forceFullPageOcr": false,
  "describeImages": false,
  "tableMode": "fast",
  "exportDoclingJson": false,
  "includeMarkdownInDataset": false,
  "includeChunksInDataset": false,
  "maxPages": 20,
  "pageBatchSize": 6,
  "doclingTimeoutSeconds": 240,
  "workerTimeoutSeconds": 300
}

Then enable OCR, accurate tables, Docling JSON, and image descriptions one at a time and compare each document row's durationMs.

Local development

Use the Apify CLI so local storage and Actor environment variables match the platform:

$apify run

Place local input at storage/key_value_stores/default/INPUT.json. Local run output remains under storage/ and is not synchronized to Apify Console.

FAQ, disclaimers, and support

Can it read password-protected PDFs?

Not currently. Remove the password before uploading the file.

Can it read private intranet URLs?

No. Remote downloads intentionally reject private, loopback, link-local, reserved, and other non-public IP addresses. Upload the PDF through the Input tab instead.

Is document content sent to an LLM?

No external service receives the document. Docling, EasyOCR, and SmolVLM run inside the Actor container. The required model files are downloaded while the Actor image is built and the runtime is configured for offline inference. The resulting content is written only to the run's Apify storages.

Are image and chart descriptions exact?

No vision model can guarantee exact interpretation of every image. EasyOCR extracts readable text, while SmolVLM supplies semantic descriptions. Native PDF tables continue through Docling's structured table extractor. For charts, always verify important values against the source image.

Only process documents you have the legal right to access and transform. PDFs may contain personal, confidential, copyrighted, or regulated information; you are responsible for lawful input, retention, sharing, and downstream use. Review your Apify storage access settings before processing sensitive material.

Use the Actor's Issues tab for bugs, conversion samples, or feature requests, and the API tab for programmatic integration.