Elysia Tools
Navigation
Workflow Playbook
Turn PDFs into clean, safe, citation-ready inputs for LLM summarization, embedding, search, and RAG pipelines.
Hubs
A PDF is rarely a clean LLM input by default. It may contain scanned pages, repeated headers, legal redlines, hidden layers, tables, captions, or pages that do not belong in the knowledge base. This workflow keeps those concerns in one sequence so the final model input can be reviewed instead of guessed.
Use encrypted-pdf-converter only when you are authorized to open the document, then narrow the workload with pdf-page-range-extractor. Page scoping matters because irrelevant front matter, appendices, and duplicated forms can distort embeddings and make citations harder to trust.
For scanned or image-heavy documents, pdf-ocr-text-layer and scanned-pdf-ocr-to-markdown help turn visual pages into searchable text. For digital PDFs, inspect semantics with tagged-pdf-inspector and pdf-to-json-structure-explorer, then export usable content through structured Markdown, table extraction, and image-caption extraction.
Before any vector ingestion, remove repeated page furniture with pdf-header-footer-noise-remover, review strikethrough remnants with pdf-strikethrough-review-extractor, and run pdf-prompt-injection-scanner to surface hidden or off-page text. Only after that should you use pdf-to-clean-text-for-llm for direct model work or pdf-rag-chunker-citation-pack for citation-grade retrieval.
Workflow playbook
Open authorized password-protected files, select only the pages needed for the knowledge task, and avoid pushing irrelevant appendices or cover pages into the model context.
Add or verify an OCR text layer for scanned PDFs, then export OCR-backed Markdown when the original document is image-heavy or copied text is unreliable.
Compare tagged structure against layout heuristics, explore headings and node metadata, and extract Markdown, tables, and image captions that need to remain available to the LLM.
Remove repeated page furniture, check struck-through review text that may still be legally relevant, and scan hidden or off-page content before trusting the extraction.
Export clean plain text for summarization or package heading-aware chunks with page references, bounding boxes, and citation metadata for retrieval systems.
Use the access and OCR tools first when the document is password-protected, scanned, or missing a reliable text layer; otherwise start with structure inspection and conversion.
Clean text is enough for summarization and drafting, while RAG ingestion should keep page numbers, headings, tables, captions, bounding boxes, and safety findings.