PhilAI

A journal that answers you with books.

Write about your day. PhilAI finds the lines that matter and links each one to a passage from public-domain literature: Dickinson, Marcus Aurelius, Brontë, Pascal. Then, if you want, you can talk about it with strangers who felt the same thing, anonymously.

Try it live → philai.vercel.app

Built with Meta Muse Spark and Muse voice transcription · Next.js · Supabase · pgvector

Why

Journaling apps either say nothing back, or reply with a chatbot's generic encouragement. But almost anything you could feel at 2am has already been written down, often beautifully, by someone a hundred years ago. PhilAI points you to the passage that was written for the line you just wrote. It doesn't summarize or advise. It hands you a book.

Reading something that names your feeling makes you want to talk about it, so every theme also has a forum where you can share a passage and the line it answered, with no names attached.

What it does

• Write or speak. Type an entry, or dictate it. Audio goes to Muse transcription and is thrown away.
• Line-level matches. PhilAI picks up to three key lines in your entry and highlights each one, along with the exact phrase in the passage that echoes it. A one-line note explains the connection.
• "Also close." Each match comes with runner-ups from other authors, each taking a different angle on the line.
• It can say "no match." If nothing in the library really fits, PhilAI shows nothing instead of a weak passage.
• It learns your taste. "Spoke to me", "Not for me" and saves shift which authors you see in later matches.
• Saved library and author pages. Keep passages, and browse everything the library holds by one author.
• Anonymous forums. 16 theme rooms (grief, belonging, ambition, regret…). Threaded replies, likes, and notifications.
• Safe sharing. Before you post, PhilAI catches names, schools, places, emails and phone numbers so you can take them out. Posts are moderated, and three reports hide anything until it's reviewed.
• Crisis care. If an entry suggests self-harm or passive ideation, the page shows the 988 Suicide & Crisis Lifeline and locks sharing for that entry.

How matching works

flowchart LR
  E[Journal entry] --> A["Muse Spark: analysis<br/>key lines, emotions, themes,<br/>what each line is really about"]
  A --> R["Retrieval<br/>MiniLM embeddings + pgvector<br/>30 candidates per line"]
  R --> S["Rerank<br/>themes, emotions,<br/>your feedback"]
  S --> C["Muse Spark: compare<br/>reads the actual passages,<br/>scores fit and insight 0–3"]
  C -->|both ≥ 2| M["Match + echoing phrase<br/>+ why-this note"]
  C -->|too weak| N[No match]

1. Analyze. One Muse Spark call reads the entry and returns its key lines (quoted exactly), emotions from GoEmotions' 28 labels, themes, and a short reading of what each line means. A risk check runs here too.
2. Retrieve. Each line, plus that reading, is embedded with all-MiniLM-L6-v2. The model runs in-process through transformers.js, so embedding needs no API call. pgvector returns the 30 closest passages.
3. Shortlist. Theme and emotion overlap and your own feedback narrow the 30 down to 10.
4. Compare. One Muse Spark call per line reads the shortlisted passages and scores fit (is this about the same thing?) and insight (does it show the writer something new?). A passage has to score at least 2 on both to be shown.
5. Verify. Every highlighted line and phrase is checked against the original text character by character before it's stored, so a highlight can't show text the model made up.

More detail: docs/insight-matching.md and docs/richer-matching.md.

How we use Meta Muse

Muse model	Used for
muse-spark-1.3 (falls back to 1.2, 1.1)	Entry analysis, passage comparison and explanations, tagging the library, scrubbing personal details, moderation
muse-voice-transcribe-1.0	Voice journaling

All calls use Standard mode: per Meta's docs, "your content is not used to train Meta models." PhilAI stores no audio. If every Spark id is down, a single call can fall back to OpenAI, but only when OPENAI_API_KEY is set.

The library

56 public-domain works by 44 authors (published 1930 or earlier, from Project Gutenberg), listed in data/sources.json. The pipeline in scripts/library/ downloads the texts and cuts them into 1–4 sentence passages. Muse Spark then tags each passage with emotions, themes, a plain-English summary and a quality score. Passages rated 4 or higher are embedded and loaded. A curated set of passages also has editorial notes on the situations and tensions they speak to.

Privacy

• Anonymous to other users. Your email is used only for login and is never returned by any API. Other people see a random handle, and a new alias in every forum thread.
• Posts aren't linked to entries. Sharing copies the passage and the line you chose. Nothing points back to your journal.
• Row-level security everywhere. Entries, matches and your saved passages can only be read by you. user_id is withheld from forum rows at the column level.
• Delete means delete. Deleting your account cascades to everything you've written.

Tech stack

App	Next.js 15 (App Router), React 19, TypeScript, Tailwind CSS 4
Data	Supabase: Postgres, pgvector, Auth, row-level security
AI	Meta Muse Spark (text), Muse voice transcription, all-MiniLM-L6-v2 via transformers.js (embeddings)
Hosting	Vercel

Run it locally

You'll need Node 20+, a Supabase project, and Muse API keys.

npm install
cp .env.example .env.local      # add Supabase and Muse keys

1. Apply the migrations in supabase/migrations/ in order, using the Supabase SQL editor or supabase db push.
2. Build the passage library (see docs/DEVELOPMENT.md), or leave NEXT_PUBLIC_MATCH_MODE=mock to try the UI with sample matches.
3. npm run dev and open http://localhost:3000.

Offline tests for the matcher (no credentials needed):

npm run test:matching

Repository map

src/app/            Pages (Write, Entries, Explore, Saved, Community) and API routes
src/lib/            Matching pipeline, Muse client, embeddings, moderation, Supabase clients
supabase/           Database schema and RPCs, as migrations
scripts/library/    Build the passage library from Project Gutenberg
scripts/            Evaluation, smoke tests and seed data
data/               Source list and the tagged passage library
shared/             Shared vocabularies (themes, emotions) and API mocks
docs/               Development notes, matching design, evaluation results

Team

Built in a weekend by Sebastian Le (sebgle), who built the app and community, and Emir Durakovic (emirdur), who built the passage library, matching and Muse integration.

License

MIT. The literary passages are in the public domain.