Event-driven UK payslip ingest: Paperless-ngx webhook -> claude-agent-service extraction -> Postgres -> Grafana
974181674d
## Context
v1 shipped a Claude Haiku-based extractor that validated only 10/71
backfilled rows. Haiku fumbles the arithmetic on pension salary-sacrifice,
conflates RSU vest with regular earnings, and occasionally misreads YTD
vs this-period columns — so 86% of rows land with validated=false and the
downstream dashboards under-report take-home.
Meta UK uses a stable two-variant template (pre/post 2022-01-31 boundary),
so a regex parser is both faster (ms vs. 30-90s + $0.01-0.05/call) and
more accurate. v2 introduces that parser as the primary path, keeps
Claude as the fallback for non-Meta payslips, and surfaces new fields
the dashboard needs to attribute PAYE between cash salary and RSU vests
correctly.
## This change
### Parser (new)
`payslip_ingest/parsers/meta_uk.py` detects the layout variant by header
presence:
- **Variant A** (pre-2022): vertical Description/This Period/This Year.
`AE Pension EE` is a positive deduction against a pre-sacrifice gross —
maps to `pension_employee` for the existing validation formula to hold.
- **Variant B** (post-2022): side-by-side Payments | Deductions | Year to
Date. `AE Pension EE` is NEGATIVE in Payments (salary sacrifice) — maps
to `pension_sacrifice` and is already netted into Total Payment.
`rsu_vest = RSU Tax Offset + RSU Excs Refund` (Meta's template inflates
Taxable Pay without using a matching offset deduction).
Column boundaries come from the header row's anchor positions; each data
row slices into 3 cells and the last numeric token per cell is the amount.
Anchor misses raise ParserError so the caller falls back to Claude rather
than silently returning bad data.
### New fields
Schema + DB + Claude prompt gain:
- `salary`, `bonus`, `pension_sacrifice` — earnings decomposition for the
dashboard's bonus-sacrifice visibility and earnings-breakdown chart
- `taxable_pay`, `ytd_tax_paid`, `ytd_taxable_pay`, `ytd_gross` — powers
the YTD-effective-rate method of attributing cash tax vs RSU tax, which
is the only method that's accurate month-to-month
All new columns default to 0 / null so v1 rows continue to round-trip.
### Orchestration
processor.py tries `parse_meta_uk(pdftotext(pdf))` first. On success the
result goes straight to the DB — zero Claude tokens spent, extraction in
milliseconds. On ParserError it falls through to ClaudeExtractor as before.
ProcessResult gains an `extractor` field ("meta_uk_regex" | "claude") so
backfill logs show the hit rate.
## Tests
- `test_meta_uk_parser.py` — 11 tests covering variant A, variant B
(standard + bonus month + bonus-sacrificed month), malformed inputs,
and end-to-end totals validation for all 4 golden fixtures.
- `test_processor.py` — 2 new tests proving the regex-first short-circuit
and the Claude fallback on non-Meta inputs.
Fixtures under `tests/fixtures/` are hand-crafted `pdftotext -layout`
emulations — real Meta numbers from the plan's sample payslips for
variant B, synthesized realistic variant A and bonus-sacrificed samples.
0001_initial.py reformat is yapf cleanup touched during the session's
format pass; not a behavior change.
## Test Plan
### Automated
```
$ poetry run pytest
============================= test session starts ==============================
collected 53 items
tests/test_extractor.py ..... [ 9%]
tests/test_meta_uk_parser.py ........... [ 30%]
tests/test_paperless.py ...... [ 41%]
tests/test_processor.py .............. [ 67%]
tests/test_schema.py .... [ 75%]
tests/test_tax_year.py ........ [ 90%]
tests/test_webhook.py ..... [100%]
============================== 53 passed in 1.66s ==============================
$ poetry run ruff check .
All checks passed!
$ poetry run mypy .
Success: no issues found in 24 source files
$ poetry run yapf --style pyproject.toml --diff --recursive payslip_ingest tests
(no output — all files are yapf-clean)
```
### Manual Verification
Smoke-test the parser against a real Meta payslip PDF on the deploy host:
```
# After 0003 migration applied to prod DB
$ poetry run python -c "
from payslip_ingest.parsers import parse_meta_uk
import subprocess
text = subprocess.check_output(['pdftotext', '-layout', '/path/to/real.pdf', '-']).decode()
p = parse_meta_uk(text)
print(p.model_dump_json(indent=2))
"
```
Expected: JSON with salary/bonus/rsu_vest/pension_sacrifice populated and
`validate_totals(p)` returning True.
## Reproduce locally
1. `cd payslip-ingest && poetry install`
2. `poetry run pytest tests/test_meta_uk_parser.py -v`
3. Expected: 11 tests pass, each fixture validates totals within 2p.
Closes: code-un1
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
||
|---|---|---|
| alembic | ||
| payslip_ingest | ||
| tests | ||
| .gitignore | ||
| .woodpecker.yml | ||
| alembic.ini | ||
| Dockerfile | ||
| poetry.lock | ||
| pyproject.toml | ||
| README.md | ||
payslip-ingest
Event-driven UK payslip ingest: Paperless-ngx fires a webhook when a document
is tagged payslip; this service fetches the PDF, calls claude-agent-service
to extract structured fields, and upserts into Postgres keyed by
paperless_doc_id (idempotent). A CLI backfill mode enumerates every
existing payslip in Paperless for initial population.
Local dev
poetry install
poetry run pytest -q
poetry run mypy .
poetry run ruff check .
# Smoke-test extraction against a real PDF (no DB writes):
export CLAUDE_AGENT_URL=http://claude-agent-service.claude-agent.svc.cluster.local:8080
export CLAUDE_AGENT_BEARER_TOKEN=...
poetry run python -m payslip_ingest extract-one /tmp/sample.pdf
Env vars
| Variable | Purpose |
|---|---|
PAPERLESS_URL |
Paperless-ngx base URL (e.g. https://paperless.viktorbarzin.me) |
PAPERLESS_API_TOKEN |
Paperless API token (User → My Profile → API Auth Token) |
CLAUDE_AGENT_URL |
claude-agent-service URL (http://claude-agent-service.claude-agent.svc.cluster.local:8080) |
CLAUDE_AGENT_BEARER_TOKEN |
Vault secret/claude-agent-service → api_bearer_token |
DB_CONNECTION_STRING |
SQLAlchemy async URL: postgresql+asyncpg://user:pass@host/db |
WEBHOOK_BEARER_TOKEN |
Shared secret Paperless sends in Authorization: Bearer ... |
Paperless workflow configuration
In Paperless-ngx, create a workflow:
- Name:
payslip-ingest - Trigger: Document Added, matching tag
payslip - Action type: Webhook
- URL:
http://payslip-ingest.payslip-ingest.svc.cluster.local:8080/webhook - Method:
POST - Headers:
Authorization: Bearer <WEBHOOK_BEARER_TOKEN>Content-Type: application/json
- Body (template):
{"document_id": {{ document_id }}}
Deployment
Ship to the payslip-ingest namespace. The service serializes incoming
webhooks onto an in-process queue so it never collides with
claude-agent-service's single-job lock.
Run the initial backfill once the deployment is live:
kubectl -n payslip-ingest create job \
--from=deployment/payslip-ingest \
payslip-backfill-$(date +%s) \
-- python -m payslip_ingest backfill --all
Architecture notes
extract-onenever touches the DB — safe for ad-hoc re-extraction on disk.- The
backfillcommand is idempotent (skips rows whosepaperless_doc_idalready exists) so it can be re-run freely. - Totals validation is a best-effort sanity check; mismatches are stored with
validated=falseandraw_extractionretained for manual review, rather than rejected. - The agent service is single-threaded. The webhook handler enqueues and returns 202; a single background worker drains the queue one at a time and absorbs 409-busy responses from the agent with retry-with-backoff.
- New agent prompt lives at
.claude/agents/payslip-extractorin theinfrarepo — this is a separate deliverable (see TODOs).