Skip to content

TODO.md

Latest commit

 

History

History
57 lines (36 loc) · 3.03 KB

TODO.md

File metadata and controls

57 lines (36 loc) · 3.03 KB
Raw

DoctorTalk TODO

Process: Anything not finished in the current change belongs here. Do not rely on memory or chat for follow-ups.

API: Gemini output length

  • Simplify uses a high max_output_tokens (8192) so long clinical notes are not cut off mid-sentence. Earlier default was 1000 and truncated the model response.
  • If check-similarity JSON is incomplete on very long texts, raise max_output_tokens in _generate for that route (currently 900).

UI: light and dark mode

  • Theme toggle in the nav (persists doctortalk-theme in localStorage, respects prefers-color-scheme on first visit). Legacy medbridge-theme is migrated once.
  • data-theme="dark" / "light" on <html> with CSS variables in static/css/style.css.
  • Optional polish: theme-colored chart fills in app.js canvas helpers (drawHeat, drawFlagList) when data-theme is light.
  • Optional: sync <meta name="color-scheme"> with the active theme for form controls.

DEFT corpus plus patient context (before or inside Gemini)

Phase 0: Licensing and format

  • Obtain the canonical DEFT source (paper, supplement, or author release) and confirm redistribution terms.
  • Normalize to an internal table or file, for example { term, plain_phrase, notes }.

Phase 1: Retrieval before generation

  • Build a matcher over clinician input (exact phrase, then fuzzy or n-gram fallbacks).
  • Inject a compact DEFT block into the simplify system prompt (cap rows to a token budget).
  • Log which DEFT rows were attached per request (audit trail).

Phase 2: Personalized patient data (already partly in app)

Existing: FlaggedWord, Session.word_frequencies, Session.hover_times, comprehension, level.

  • Formalize a patient context object from the DB (top flagged terms, weighted by word_frequencies, trend from sessions).
  • Append Patient context to the Gemini system prompt (same route as today’s history_ctx, but structured and bounded).
  • Optional later: store accepted simplify pairs and retrieve similar past lines (embedding search).

Phase 3: Generation contract

  • Keep a single simplify entrypoint: system + DEFT_RETRIEVED + PATIENT_CONTEXT + user text.
  • Version prompts in code or config for reproducibility.

ClinicalBERT (or clinical embeddings) after generation

  • Sentence-split original and simplified text with the same rules.
  • Align sentences (greedy or DP by cosine similarity).
  • Load medicalai/ClinicalBERT (or a clinical sentence embedding model) via transformers, mean-pool hidden states, compute cosine similarity per pair.
  • Aggregate (mean and min) and return alongside optional Gemini narrative check.
  • Add torch / transformers to the project and cache the model in process.

Honest labeling

  • UI copy: distinguish embedding similarity from LLM self-judgment and from “clinical safety”.

Ops

  • Feature-flag heavy models so dev machines can run UI without GPU.