Generates personalised meeting recommendations from the EA Global London 2026 attendee sheet, then produces an optimised schedule based on mutual availability.

Everything runs locally in the terminal. Retrieval uses a local hybrid search (embeddings + BM25) — no data leaves the machine. The LLM reasoning uses your own API key.

npm install
pip install -r requirements.txt
cp .env.example .env          # fill in at minimum LLM_PROVIDER + the matching API key
npm run embed                  # one-time: build the local embedding cache (~3 min)

Create your context file:

cp requester_context.example.json requester_context.json
# edit requester_context.json — name, goals, extra_context

python recommend.py            # uses requester_context.json by default
python recommend.py my_ctx.json  # or pass a custom file

Two output files appear in output/:

• recommendations_<name>_<date>.md — detailed cards with why/talking-points/opener, sorted rating 5 → 1
• meeting_candidates_<name>_<date>.md — 5-column table for the scheduler (also symlinked to output/meeting_candidates_latest.md)

Review the detailed file, then edit the candidates table: adjust ratings, delete people you don't want to meet, add anyone you found manually (paste their Swapcard profile URL into the Link column; the Swapcard ID column is auto-extracted from the URL).

Export your Swapcard availability first (see below), then:

python schedule.py

Prints the optimal schedule to stdout. Does not book meetings — you send requests manually so you can write a personal message to each person.

All settings live in .env (copy .env.example). Key variables:

The scheduler needs your existing agenda to avoid double-booking:

1. Log into Swapcard and navigate to your schedule page.
2. Open DevTools (F12) → Network tab.
3. Refresh the page.
4. Find requests to api/graphql — look for responses containing "agenda" data.
5. Right-click → Copy response → paste into a file called my_availability.json as a JSON array: [<response1>, <response2>, ...].

Getting a fresh SWAPCARD_TOKEN:

1. In the same Network tab, click any request.
2. Headers → find Authorization: Bearer <token>.
3. Copy the token value (everything after Bearer ) into .env.

# edit SHEET_ID, SHEET_GID, EVENT_ID, DATE_RANGE_START, DATE_RANGE_END in .env
npm run embed -- --force

• SHEET_ID / SHEET_GID env vars point to a different Google Sheet
• If the sheet column order changes, update lib/sheet.ts

• No auto-booking. The scheduler outputs a plan; you send Swapcard meeting requests manually. This is intentional — a personal message explaining why you want to talk performs far better than an auto-generated one.
• Context window. ~190 full profiles × ~400 tokens ≈ 90k tokens total for the recommendation call. Fits in 200k (Sonnet) with room to spare.
• Bearer token TTL. The Swapcard token expires in ~24h. Refresh it in .env before each Stage 2 run.
• Embedding cache. Stored in cache/ (git-ignored). Rebuilt automatically when the sheet changes. First build takes ~3 minutes.