Elysia Tools
Navigation
Workflow Playbook
Clean PDFs and Word files, remove extraction noise, scan hidden prompt risks, split text, score chunk quality, and package citation-ready RAG inputs.
Hubs
RAG quality starts before the first vector is created. A corpus may include searchable PDFs, scanned pages, Word documents, repeated page furniture, hidden instructions, appendices, tables of contents, and long sections that do not split cleanly. This workflow keeps those preparation decisions visible so indexing does not become a black box.
Start by narrowing the source with pdf-page-range-extractor, then choose the extraction path. Use pdf-to-clean-text-for-llm for selectable PDFs, scanned-pdf-ocr-to-markdown for image-heavy pages, and word-text-extractor for DOCX sources. This gives the team one comparable text layer before chunking begins.
Before accepting that text, inspect risk. pdf-header-footer-noise-remover helps find repeated page furniture, tagged-pdf-inspector shows whether semantic structure survived extraction, and pdf-prompt-injection-scanner catches hidden or off-page instructions that should not enter an automated prompt or retrieval corpus.
Use ai-token-estimator to check whether planned chunks fit the target model and retrieval policy. Then use sentence-splitter or text-splitter to create candidate boundaries. Finish with ai-rag-chunk-quality-scorer to compare coherence, coverage, overlap health, and size consistency, and use pdf-rag-chunker-citation-pack when chunks need page-aware citations. The final handoff should include accepted chunk settings, rejected cleanup risks, and enough citation metadata to trace answers back to source pages.
Workflow playbook
Select the pages and document sources that belong in the retrieval corpus, then convert PDFs, scans, and Word files into reviewable text or Markdown.
Check whether headers, footers, weak PDF structure, or hidden prompt-injection text will pollute embeddings or downstream answer grounding.
Estimate token size, split text by meaningful sentence or delimiter boundaries, and create chunk candidates that match the retrieval system's limits.
Compare candidate chunking schemes for coherence, coverage, overlap health, and size consistency, then prepare page-aware citation packs for indexing.
Use PDF cleanup for selectable PDFs, OCR-to-Markdown for scanned pages, and Word extraction for DOCX material so each source enters the same reviewable text pipeline.
Use citation-pack generation when answers must trace back to pages or regions; use general text splitting and quality scoring when you are comparing chunk policies before indexing.