About the implementation and examples
Engineering notes draw on author-supplied MailTrace development reports through 13 October 2026. They describe the documented revision, including local work, rather than independently verified live deployment. Experiment designs and architecture proposals describe work to evaluate. Examples are educational, use synthetic data and are not copied production code.
Model the preparation stages separately
The supplied local implementation update reports import and upload, previews and downloads, saved titles, source links, filters, reversible exclusions, retries and retrieval eligibility checks. It explicitly leaves OCR, image analysis and audio/video transcription unavailable. These are development-status statements, not a claim that the evidence changes are deployed.
Text extraction can prepare a text-bearing PDF for search. A scan needs an OCR path or another documented extraction method. A photo may need image analysis, and a recording needs transcription before text retrieval can cover its spoken content. Those stages have different failure modes and should not be hidden behind one “processed” badge.
Describe what an item can currently support
This illustrative representation separates availability of the file from availability of its text. A user can still view or download a scan while the application explains why its contents are not included in an answer.
{
"source_id": "pdf-example",
"file_available": true,
"preview_available": true,
"text_status": "unavailable",
"reason": "scan_requires_ocr",
"eligible_for_text_retrieval": false
}Add extraction without losing provenance
A future OCR integration should save extractor identity, version, page references and the relationship between extracted text and original pages. A vision summary is a derived interpretation and should not silently replace the text or image it describes. Errors in names, dates and reference numbers need review paths.
Evaluate with text PDFs, scans, mixed documents, rotated pages and low-quality images. Test resource limits before expensive processing, then explain failures and retries in the interface. The objective is a reliable route from an answer citation back to a page or source segment, with honest coverage when some media cannot yet be searched.