Improving document ingestion
This page distinguishes implemented Tychora behavior from upstream capabilities worth evaluating. Upstream references were checked on 2026-10-13. These are proposed improvements, not configuration switches already exposed by Tychora.
Current pipeline
The Docling adapter accepts PDF, DOCX and supported images, retains structured JSON and source geometry, extracts tables, and generates page/picture assets. It uses a pinned local layout model, CPU execution and explicitly provisioned model artifacts. Remote Docling services and external plugins are disabled.
The standard Docling pipeline sets do_ocr=False. Separate selected-region OCR and optional Anthropic vision exist, but a scanned document does not automatically receive a complete Docling OCR pass. Optional Anthropic region interpretation is also different from converting entire pages with Docling's VLM pipeline.
Highest-value experiments
| Improvement | Why it matters | Work needed before shipping |
|---|---|---|
| OCR as an explicit extraction policy | Recover text from scans and image-only regions | Language/engine selection, pinned assets, confidence and exact provenance; scan gold set |
| Heading hierarchy recovery | Better sections, context groups and useful neighbor boundaries | Enable the upstream hierarchy stage, preserve its provenance, test mixed numbering/bookmarks/styles |
| Table extraction controls | Improve difficult cell alignment and choose accuracy/runtime trade-offs | Compare FAST/ACCURATE and cell matching on a labeled table set; preserve fragments and row locators |
| Layout and resource controls | Improve reading order and predictable throughput | Versioned options for layout rules, render resolution, accelerators and limits; memory/time measurements |
Docling's advanced options expose model prefetching, table behavior, image scaling and execution limits. Tychora should expose a small validated subset per profile, include it in extraction fingerprints, and retain the effective settings with each extraction. Simply passing arbitrary upstream arguments would weaken reproducibility.
The heading hierarchy stage can use document structure and layout cues. It requires HeadingHierarchyOptions(enabled=True) and parsed-page information. Its choices can change parentage and chunk context, so compare section boundaries and citation association before enabling it by default.
Enrichments and full-page vision
Docling enrichments offer code/formula and picture processing. These can improve formula retrieval, diagram descriptions and figure search. Store generated text as derived evidence with model, prompt/configuration and source dependencies; preserve the original crop as the citation target. Additional enrichments add model downloads, latency and memory, so activate only what the corpus needs.
The VLM pipeline is a separate conversion path worth testing for layouts where standard extraction fails. Evaluate retained source geometry, hallucinated text, reading order and cost. A locally run model needs asset and hardware qualification; a remote model also needs the existing upload-consent boundary. Do not route private source bytes through a new provider implicitly.
The model catalog is useful for selecting task-specific candidates. Tychora's asset manifest should pin model identity, bytes, license and expected runtime rather than selecting whichever upstream alias is newest at execution time.
Audio and video
Docling audio/video processing makes transcripts and media-derived structures possible, but Tychora does not currently support those source formats. Useful evidence would need time ranges, transcript alignment, frame identities and source-media retention. Speaker identity, ASR uncertainty and frame sampling also need explicit policies.
Treat this as a separate adapter/provenance extension. A generic document-kind definition alone cannot add media decoding, ASR assets or timestamp citations.
Acceptance criteria
For each experiment, keep the original immutable and compare the same labeled source set. Measure text accuracy, heading boundaries, table-cell/row reconstruction, figure association, citation validity, retrieval recall, wall time and peak memory. A better-looking Markdown export is insufficient if locators or retrieval quality regress. See evaluation and architecture.