Ride Receipts LLM

ride-receipts-llm

Run a reproducible 3-stage pipeline:

0. initialize/validate SQLite schema (fixed; do not edit)
1. fetch full receipt emails into JSONL
2. extract structured rides with LLM (one-shot + repair)
3. upsert into SQLite

Prerequisites and safety

• Require gog CLI installed and authenticated for the selected Gmail account.
• Prefer configured account: skills.entries.ride-receipts-llm.config.gmailAccount; if missing, ask user for account explicitly.
• Ask for date scope before fetch: all-time, after YYYY-MM-DD, or between dates.
• Treat receipt content as sensitive financial/location data.
• Before extraction, explicitly confirm user is okay sending raw email HTML to the active LLM.
• Extraction uses raw text_html from emails; do not claim local-only parsing.
• Never hallucinate fields; keep unknown values null.

Paths

• Schema (do not modify): skills/ride-receipts-llm/references/schema_rides.sql
• Emails JSONL: ./data/ride_emails.jsonl
• Extracted rides JSONL: ./data/rides_extracted.jsonl
• SQLite DB: ./data/rides.sqlite

0) Initialize DB

python3 skills/ride-receipts-llm/scripts/init_db.py \
  --db ./data/rides.sqlite \
  --schema skills/ride-receipts-llm/references/schema_rides.sql

1) Fetch Gmail receipts → JSONL

python3 skills/ride-receipts-llm/scripts/fetch_emails_jsonl.py \
  --account <gmail-account> \
  --after YYYY-MM-DD \
  --before YYYY-MM-DD \
  --max-per-provider 5000 \
  --out ./data/ride_emails.jsonl

• Omit --after / --before when not needed.
• Output rows include provider metadata, snippet, and raw text_html.

2) LLM extraction contract

Read ./data/ride_emails.jsonl; write one JSON object per line to ./data/rides_extracted.jsonl.

Per email:

1. Run one-shot extraction for all fields.
2. Quality-gate: amount,currency,pickup,dropoff,payment_method,distance_text,duration_text,start_time_text,end_time_text.
3. If any are missing, run repair pass(es) for missing fields only.
4. Merge additively; never replace existing non-null values with null.

Schema (one line per ride):

{
  "provider": "Uber|Bolt|Yandex|Lyft",
  "source": {"gmail_message_id": "...", "email_date": "YYYY-MM-DD HH:MM", "subject": "..."},
  "ride": {
    "start_time_text": "...",
    "end_time_text": "...",
    "total_text": "...",
    "currency": "EUR|PLN|USD|BYN|RUB|UAH|null",
    "amount": 12.34,
    "pickup": "...",
    "dropoff": "...",
    "pickup_city": "...",
    "pickup_country": "...",
    "dropoff_city": "...",
    "dropoff_country": "...",
    "payment_method": "...",
    "driver": "...",
    "distance_text": "...",
    "duration_text": "...",
    "notes": "..."
  }
}

Rules:

• Use text_html as primary source; fallback to snippet only if text_html is empty.
• Keep addresses/time strings verbatim.
• Keep amount numeric; if only textual total exists, set amount: null and preserve text in total_text.

3) Insert extracted rides → SQLite

python3 skills/ride-receipts-llm/scripts/insert_rides_sqlite_jsonl.py \
  --db ./data/rides.sqlite \
  --schema skills/ride-receipts-llm/references/schema_rides.sql \
  --rides-jsonl ./data/rides_extracted.jsonl

Schema is idempotent via UNIQUE(provider, gmail_message_id) ON CONFLICT REPLACE.