Inspiration

People can arrive at an appointment with observations scattered across dates and topics. VisitWeave explores a private way for people to choose what they want to discuss, while keeping their original wording attached to each question. VisitWeave is informed by AHRQ's Questions Are the Answer and the WHO Primary Care Checklist. Neither organization has evaluated or endorsed this prototype.

What it does

VisitWeave turns a person's own paragraphs into a short, editable visit agenda. Its interactive website accepts notes, displays up to three source-linked cards, and allows the person to edit, reorder, remove, add, and copy questions. A fictional sample lets judges try the complete flow without entering health information. Every selected card quotes the exact note it came from.

How it was built

The browser-only extractive NLP engine tokenizes notes, weights terms with TF-IDF, measures cosine similarity, and builds a graph among notes. Thirty PageRank-style iterations rank recurring threads. A diversity penalty keeps near-duplicate notes from filling the list. The interface uses semantic HTML, CSS, and vanilla JavaScript with no runtime service, account, or paid model. Note text is processed in the browser tab; the app has no backend endpoint for notes and does not store them in browser storage.

Codex produced the concept framing, submission text, implementation, visual design, and tests with the participant's review and selection. The participant has reviewed the code explanation and confirmed that they understand the prototype and can explain its implementation. Automated checks were run by Codex; they are not described as manual testing performed by the participant. No remote AI model processes a user's notes. The source code and three engine tests are public.

Design challenges and testing

The hard part was keeping the ranking transparent and medically modest. A similarity score can be mistaken for clinical importance, so the UI labels it relative prominence and explicitly says it is not medical priority. The generated prompt is intentionally generic; the user supplies the actual question. Three automated engine tests check exact source-text traceability, generic question prompts and score bounds on sample inputs, and empty/unrelated input handling with the 5,000-character limit. They passed in the current review. These narrow tests do not establish clinical safety or usability.

Limits and next steps

VisitWeave does not interpret symptoms, assess urgency, triage, diagnose, or recommend treatment. It is not clinically validated. Further work would test readability and usefulness with consenting users and clinicians, improve multilingual tokenization, and conduct a privacy and accessibility review.