Web UI (P3) and semantic recall (P4)

[jackeydou]

Implements P3 and P4 of the foundation spec. Decisions made along the way are recorded as dated notes in the spec.

P3: web UI

• distill ui serves a local web app at http://distill.localhost:<ui.port>. It listens on loopback only, checks Host and Origin, and requires a session cookie that a one-time link sets. distill ui stop stops it. distill mcp starts it in the background before handing out note links; ui.autostart = false turns that off. Details: docs/web-ui.md.
• Web app (apps/web, React 19 + TanStack Router/Query + Tailwind v4), embedded into the binary:
  • The sidebar has Distill and Review tabs, with every tag listed below them.
  • Distill is a timeline of notes grouped by day: each entry shows the question and the first two lines of the conclusion, and expands in place to the full note. A tag filters it.
  • Review is a dashboard: week-over-week notes, repeat rate, annotation coverage, active days, a 26-week activity calendar, notes per week, most-asked and unannotated topics, duplicate suggestions, and breakdowns by tag, project and agent.
  • The note page (add/edit/delete annotations, reopen the original session), the topic page (rename, merge) and the sync-conflict page remain.
• Core: new operations for topic rename and merge, annotation edit and delete, conflict resolution (the losing copy is moved to <data dir>/discarded/, not deleted), a paged timeline, and extra stats.
• API types are generated from Rust with ts-rs into apps/web/src/api/generated. CI fails if the committed files drift.

P4: semantic recall

• distill model pull downloads a local multilingual embedding model: paraphrase-multilingual-MiniLM-L12-v2, quantized, about 257 MB, run through fastembed on ONNX Runtime.
• With the model installed, recall merges the keyword and embedding rankings, so questions asked in other words or another language are found. Without it, recall stays keyword-only.
• Vectors are stored and searched in sqlite-vec vec0 tables with cosine distance. Registering the extension needs one unsafe call, so the workspace lint changes from forbid to deny, with a single exception in crates/distill-core/src/index/vec.rs.
• The web UI suggests topic pairs that look like the same question, with merge and dismiss. No background LLM is involved.
• Thresholds were measured on 12 Chinese/English question pairs: rewordings scored 0.40–0.74 and different questions up to 0.54, so recall keeps a low floor (0.40) and merge suggestions use 0.65. See docs/recall.md.

Things to know

• Building now needs Node 22 and pnpm. mise run install builds the web UI, then installs the binary with it embedded.
• The release binary grows from about 12 MB to about 42 MB, mostly ONNX Runtime.
• The model tests are #[ignore]d and CI does not run them. Run them locally after pulling the model:
  DISTILL_TEST_MODEL_HOME=<home with model> cargo test -p distill-core --test semantic -- --include-ignored

Testing

• mise run check passes: Rust fmt, clippy, tests, plus Biome, tsc, Vitest and a build of the web app.
• The P4 exit criterion is a test: two rewordings that keyword recall misses are found by semantic recall. It passes against the real model.
• Clicked through in the in-app browser against a seeded vault:
  • authorize a browser, open a note, add an annotation;
  • jump from a tag to other notes with that tag;
  • expand a timeline entry, filter the timeline by tag;
  • view the Review dashboard, merge a suggested duplicate;
  • dark theme and phone width.
• Not verified yet:
  • opening an agent-given link from a real Codex / Claude Code session;
  • the deep link from a note back to its original session;
  • *.localhost resolution on Linux and Windows;
  • CI with the new Node and ONNX Runtime steps (first run is this PR).

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
🔒 Security Review	✅ Completed 2026-09-26T22:18:19.366395Z	96fbf77	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.