1. Keep exact words searchable
OCR and highlighted text preserve names, error codes, amounts, and phrases. SQLite FTS5 is the planned full-text index for these fields and user notes. Current Python text search uses literal matching; the website uses authored keyword matching. Neither is an implemented production FTS5 index.
2. Find visual meaning with vectors
The Python prototype already uses compatible CLIP image/text encoders to compare a written description with image embeddings. Keep model identity, preprocessing, and vector dimensions explicit. Evaluate sqlite-vec against exact scanning before choosing the production vector index.
3. Retrieve the meaning of your notes
A separate text-embedding model can match a request such as “make the purchase button less prominent” to a note like “this CTA is too loud.” That note encoder still needs evaluation. Do not compare vectors from unrelated model spaces directly; combine the ranked candidate lists instead.
4. Use the links that make context precise
The lightweight ontology tells search and assistants which note refers to which region or selected words, and which source/collection the image belongs to. Explicit links and filters can narrow eligible records without asking a model to invent relationships.
5. Return evidence, not just a similarity score
Fuse ranked text, visual, and note candidates and apply authorized-source filters. Return the matched record, selected text, note/region IDs, source, and ranking-method evidence. An assistant requests the original or crop separately. It receives readable context and images, not the raw embedding database.
Update and evaluate
Editing a note refreshes its text and vector representation. Changed originals create a new image revision rather than moving old highlights onto new pixels. Use labeled exact, visual, note-paraphrase, and no-match queries; record quality, latency, resource costs, and model revisions. Ranking scores are not calibrated probabilities.
This preview
The website searches authored sample text and notes. It uses authored text boxes for highlights and stores annotations in your browser. No vector model, OCR engine, personal-file indexing, or assistant connection runs on this page.