RAG Corpus Provenance Packager

Turn explicit, rights-declared HTML, Markdown, text, and born-digital PDF sources into deterministic, deduplicated corpus pages, traceable chunks, and a portable evidence package.

What this does not do

• No open-web crawling or link following.
• No paywall or authentication bypass.
• No OCR in version 1.
• No vector embedding generation.
• No legal conclusion about reuse rights.
• No LLM summaries or synthetic citations.
• No billing by both page and chunk.

Sources and authorization

Every source must be one explicit HTTPS URL or one record from an Apify key-value store you can read. Every source also requires a rightsBasis declaration with evidence text or an evidence URL. The declaration and any detected indicators are preserved as provenance; they are not legal advice and do not establish permission. A metadata_only declaration is refused because this Actor's output contains source text.

HTML uses direct HTTP responses only. Markdown and plain text must be UTF-8 or UTF-16. PDFs must contain extractable text; image-only pages return an uncharged ocr_required result. HTML selectors are explicit and bounded. The Actor never follows links discovered in a source.

Input example

{
  "sources": [
    {
      "url": "https://flintglade-tools.github.io/apify-fixtures/rag/sample.html",
      "rightsBasis": {
        "type": "user_owned",
        "evidence": "Flintglade-controlled public fixture"
      },
      "formatHint": "html",
      "excludeSelectors": ["footer"]
    }
  ],
  "chunking": {"strategy": "character", "size": 1000, "overlap": 150},
  "dedupe": {"mode": "exact"},
  "retainRaw": false,
  "concurrency": 1
}

Outputs

The default dataset contains one result per page plus one uncharged run summary. Admitted pages include stable document/page IDs, canonical and raw hashes, rights evidence, extraction indicators, chunk IDs, and invalidation links. Exact duplicates and refusals remain visible but uncharged.

Child chunks are written to a run-specific named dataset and are included in the page price. Every chunk carries its page and document IDs, source and page locators, character range, text hash, rights declaration, and chunker version.

The default key-value store receives:

• CORPUS_MANIFEST.json
• EXTRACTED_DOCUMENT.jsonl
• CORPUS_PACKAGE.zip
• VERIFICATION.json
• ARTIFACT_MANIFEST.json
• generated RAW_SOURCE_* records only when retainRaw is explicitly enabled

Example admitted page fields:

{
  "status": "ok",
  "charge": {"event": "packaged-page", "chargeable": true},
  "data": {
    "pageId": "page_...",
    "canonicalHash": "sha256:...",
    "chunkCount": 2,
    "duplicateOfPageId": null,
    "rightsBasis": {"type": "user_owned", "evidence": "Flintglade-controlled public fixture"}
  }
}

Pricing

packaged-page costs $0.02 per admitted source page. Generated chunks are included and are never charged separately. For example, three admitted pages cost $0.06 in Actor events, plus Apify platform usage. Blocked URLs, source outages, unsupported media, oversized sources, parse failures, rights refusals, OCR-required pages, exact duplicates after the first page, and charge-limit rows are uncharged.

Baseline behavior

Without baselineStoreId, each run is stateless. With an existing authorized KVS, admitted pages compare their canonical hash to the prior page record and report baseline_created, unchanged, or changed. Changed pages list the superseded hash and prior chunk IDs. Baseline state is committed only after Apify confirms the paid page event, and a concurrent baseline change is preserved rather than overwritten.

Limits

At most 100 sources, 50 MB per source, 2,000 PDF pages per run, and 250,000 chunks per run. Configured source concurrency is bounded to four; the runtime may reduce it to one so the next source is not scheduled after the paid-event budget is exhausted. The run timeout is 900 seconds, with 2 GB default memory.

Support

See SUPPORT.md for reproducible issue-report details. Never include private source content, signed URLs, storage tokens, or credentials in a support request.