"""Compare a bank CSV statement with the books.

This tool takes an AMEX or First Republic CSV statement file and compares it

line-by-line with the Beancount books to make sure that everything matches. This

is designed for situations where transactions are entered into the books

directly, rather than being imported from a statement after the fact.

The reconciler will attempt to match transactions based on date, amount, check

number and payee, but is forgiving to differences in dates, the absensce of

check number and inexact matches on payee. Matches are ranked, so where there is

only one decent match for an amount/date this is accepted, but if there are

multiple similar candidates it will refuse to guess.

The reconciler will also attempt to identify where a single statement entry has

been split out into multiple Beancount postings, such as a single bank transfer

representing health insurance for multiple employees.

Run it like this:

  --beancount-file 2021.beancount \

  --account Liabilities:CreditCard:AMEX \

  --csv-statement ~/svn/2021-09-10_AMEX_activity.csv \

  --bank-statement ~/svn/2021-09-10_AMEX_activity.pdf

Background:

Regular Beancount users often write automated importers to create bookkeeping

entries direct from a bank statement or similar. That combines data entry and

reconciliation in one step. Conservancy uses a different approach; they manually

entering transactions and reconciling them later on. This workflow is helpful in

cases like writing checks (see below). This is the workflow implented by this

tool.

That said, this tool *is* still somewhat like an importer in that it needs to

extract transaction details from a third-party statement. Instead of creating

directives, it just checks to see that similar directives are already

present. This is a bit like diff-ing a statement with the books (though we're

only interested in the presence of lines, not so much their order).

Paper checks are entered into the books when written (a.k.a. "posted"), but may

not be cashed until months later sometimes causing reconciliation differences

that live beyond a month. It's worth noting that there are really two dates here

- the posting date and the cleared date. Beancount only allows us to model one,

which is why carrying these reconciliation differences between months feels a

bit awkward.

Problems in scope:

 - errors in the books take hours to find during reconciliation, requiring

   manually comparing statements and the books and are succeptible to mistakes,

   such as not noticing when there are two payments for the same amount on the

   statement, but not in the books (as Bradley likes to quote, "you're entering

   a world of pain")

 - adding statement/reconciliation metadata to books is/was manual and prone to

   mistakes

 - jumping to an individual transaction in a large ledger isn't trivial - Emacs

   grep mode is the current best option

 - not all staff use Emacs

 - auditors would prefer Bradley didn't perform reconciliation, ideally not

   Rosanne either

 - reconciliation reports are created by hand when there are mismatches

Other related problems we're not dealing with here:

 - after updates to the books files, beancount must be restarted to reflect

   updates

 - updates also invalidate the cache meaning restart takes several minutes

 - balance checks are manually updated in

   svn/Financial/Ledger/sanity-check-balances.yaml

 - transactions are entered manually and reconciled after the fact, but

   importing from statements may be useful in some cases

Current issue:

 - entry_point seems to swallow errors, meaning you get a fairly unhelpful

   message if there's an unhandled error

Future possibilities:

 - allow the reconciler to respect manually-applied metadata - not clear how

   this would work exactly

 - allow interactive matching where the user can specifiy a match

 - consider combining this with helper.py into one more complete tool that both

   reconciles and summarises the unreconciled transactions

"""

import argparse

import collections

import copy

import csv

import datetime

import decimal

import io

import itertools

import logging

import os

import re

import sys

from typing import Dict, List, Optional, Sequence, Tuple, TextIO

from beancount import loader

from beancount.query.query import run_query

from colorama import Fore, Style  # type: ignore

from .. import cliutil

from .. import config as configmod

if not sys.warnoptions:

    import warnings

    # Disable annoying warning from thefuzz prompting for a C extension. The

    # current pure-Python implementation isn't a bottleneck for us.

    warnings.filterwarnings('ignore', category=UserWarning, module='thefuzz.fuzz')

from thefuzz import fuzz  # type: ignore

PROGNAME = 'reconcile-statement'

logger = logging.getLogger(__name__)

# Get some interesting feedback on call to RT with this:

# logger.setLevel(logging.DEBUG)

# logger.addHandler(logging.StreamHandler())

JUNK_WORDS = [

    'software',

    'freedom',

    'conservancy',

    'conse',

    'payment',

    'echeck',

    'bill',

    'debit',

    'wire',

    'credit',

    "int'l",

    "in.l",

    'llc',

    'online',

    'donation',

    'usd',

    'inc',

]

JUNK_WORDS_RES = [re.compile(word, re.IGNORECASE) for word in JUNK_WORDS]

ZERO_RE = re.compile('^0+')

PAYEE_FULL_MATCH_THRESHOLD = 0.8

PAYEE_PARTIAL_MATCH_THRESHOLD = 0.4

OVERALL_EXCELLENT_MATCH_THRESHOLD = 0.8  # Clear winner

OVERALL_ACCEPTABLE_MATCH_THRESHOLD = 0.5  # Acceptable if only one match found

def remove_duplicate_words(text: str) -> str:

    unique_words = []

    known_words = set()

    for word in text.split():

        if word.lower() not in known_words:

            unique_words.append(word)

            known_words.add(word.lower())

    return ' '.join(unique_words)

def remove_payee_junk(payee: str) -> str:

    """Clean up payee field to improve quality of fuzzy matching.

    It turns out that bank statement "description" fields are

    difficult to fuzzy match on because they're long and

    noisey. Truncating them (see standardize_XXX_record fns) and

    removing the common junk helps significantly.

    """

    for r in JUNK_WORDS_RES:

        payee = r.sub('', payee)

    payee = ZERO_RE.sub('', payee)

    payee = payee.replace(' - ', ' ')

    payee = re.sub(r'\.0\.\d+', ' ', payee)

    payee = payee.replace('.0', ' ')

    payee = payee.replace('/', ' ')

    payee = re.sub(re.escape('.com'), ' ', payee, flags=re.IGNORECASE)

    payee = re.sub(re.escape('.net'), ' ', payee, flags=re.IGNORECASE)

    payee = payee.replace('*', ' ')

    payee = ' '.join([i for i in payee.split(' ') if len(i) > 2])

    payee = payee.replace('-', ' ')

    payee = remove_duplicate_words(payee)

    payee.strip()

    return payee

def parse_amount(amount: str) -> decimal.Decimal:

    """Parse amounts and handle comma separators as seen in some FR statements."""

    return decimal.Decimal(amount.replace('$', '').replace(',', ''))

def validate_amex_csv(sample: str) -> None:

    required_cols = {'Date', 'Amount', 'Description', 'Card Member'}

    reader = csv.DictReader(io.StringIO(sample))

    if reader.fieldnames and not required_cols.issubset(reader.fieldnames):

        sys.exit(

            f"This AMEX CSV doesn't seem to have the columns we're expecting, including: {', '.join(required_cols)}. Please use an unmodified statement direct from the institution."

        )

def standardize_amex_record(row: Dict, line: int) -> Dict:

    """Turn an AMEX CSV row into a standard dict format representing a transaction."""

    # NOTE: Statement doesn't seem to give us a running balance or a final total.

    return {

        'date': datetime.datetime.strptime(row['Date'], '%m/%d/%Y').date(),

        'amount': -1 * parse_amount(row['Amount']),

        # Descriptions have too much noise, so taking just the start

        # significantly assists the fuzzy matching.

        'payee': remove_payee_junk(row['Description'] or '')[:20],

        'check_id': '',

        'line': line,

    }

def read_amex_csv(f: TextIO) -> list:

    reader = csv.DictReader(f)

    # The reader.line_num is the source line number, not the spreadsheet row

    # number due to multi-line records.

    return sort_records(

        [standardize_amex_record(row, i) for i, row in enumerate(reader, 2)]

    )

def validate_fr_csv(sample: str) -> None:

    # No column headers in FR statements

    reader = csv.reader(io.StringIO(sample))

    next(reader)  # First row is previous statement ending balance

    row = next(reader)

    date = None

    try:

        date = datetime.datetime.strptime(row[1], '%m/%d/%Y')

    except ValueError:

        pass

    amount_found = '$' in row[4] and '$' in row[5]

    if len(row) != 6 or not date or not amount_found:

        sys.exit(

            "This First Republic CSV doesn't seem to have the 6 columns we're expecting, including a date in column 2 and an amount in columns 5 and 6. Please use an unmodified statement direct from the institution."

        )

def standardize_fr_record(line, row):

    record = {

        'date': datetime.datetime.strptime(row[1], '%m/%d/%Y').date(),

        'amount': parse_amount(row[4]),

        'payee': remove_payee_junk(row[3] or '')[:20],

        'check_id': row[2].replace('CHECK  ', '') if 'CHECK  ' in row[2] else '',

        'line': line,

    }

    return record

def read_fr_csv(f: TextIO) -> list:

    reader = csv.reader(f)

    # The reader.line_num is the source line number, not the spreadsheet row

    # number due to multi-line records.

    return sort_records(

        standardize_fr_record(i, row)

        for i, row in enumerate(reader, 1)

        if len(row) == 6 and row[2] not in {'LAST STATEMENT', 'THIS STATEMENT'}

    )

def standardize_beancount_record(row) -> Dict:  # type: ignore[no-untyped-def]

    """Turn a Beancount query result row into a standard dict representing a transaction."""

    return {

        'date': row.date,

        'amount': row.number_cost_position,

        'payee': remove_payee_junk(

            f'{row.payee or ""} {row.entity or ""} {row.narration or ""}'

        ),

        'check_id': str(row.check_id or ''),

        'filename': row.filename,

        'line': row.line,

        'bank_statement': row.bank_statement,

    }

def format_record(record: dict) -> str:

    """Generate output lines for a standard 1:1 match."""

    if record['payee'] and record['check_id']:

        output = f"{record['date'].isoformat()}: {record['amount']:11,.2f} {record['payee'][:25]} #{record['check_id']}".ljust(59)

    elif record['payee']:

        output = f"{record['date'].isoformat()}: {record['amount']:11,.2f} {record['payee'][:35]}".ljust(59)

    else:

        output = f"{record['date'].isoformat()}: {record['amount']:11,.2f} #{record['check_id']}".ljust(59)

    return output

def format_multirecord(r1s: List[dict], r2s: List[dict], note: str) -> List[list]:

    """Generates output lines for one statement:multiple books transaction match."""

    assert len(r1s) == 1

    assert len(r2s) > 1

    match_output = []

    match_output.append(

        [

            r1s[0]['date'],

            f'{format_record(r1s[0])}  →  {format_record(r2s[0])}  ✓ Matched{note}',

        ]

    )

    for r2 in r2s[1:]:

        match_output.append(

            [

                r1s[0]['date'],

                f'{r1s[0]["date"].isoformat()}:             ↳                                    →  {format_record(r2)}  ✓ Matched{note}',

            ]

        )

    return match_output

def _start_of_month(time, offset_months=0):

    if offset_months > 0:

        return _start_of_month(

            time.replace(day=28) + datetime.timedelta(days=4), offset_months - 1

        )

    else:

        return time.replace(day=1)

def round_to_month(begin_date, end_date):

    """Round a beginning and end date to beginning and end of months respectively."""

    return (

        _start_of_month(begin_date),

        _start_of_month(end_date, offset_months=1) - datetime.timedelta(days=1),

    )

def sort_records(records: List) -> List:

    return sorted(records, key=lambda x: (x['date'], x['amount']))

def first_word_exact_match(a: str, b: str) -> float:

    """Score a payee match based first word.

    We get a whole lot of good matches this way. Helps in the

    situation where the first word or two of a transaction description

    is useful and the rest is garbage.

    """

    if len(a) == 0 or len(b) == 0:

        return 0.0

    first_a = a.split()[0].strip()

    first_b = b.split()[0].strip()

    if first_a.casefold() == first_b.casefold():

        return min(1.0, 0.2 * len(first_a))

    else:

        return 0.0

def payee_match(a: str, b: str) -> float:

    """Score a match between two payees."""

    fuzzy_match = float(fuzz.token_set_ratio(a, b) / 100.00)

    first_word_match = first_word_exact_match(a, b)

    return max(fuzzy_match, first_word_match)

def records_match(r1: Dict, r2: Dict) -> Tuple[float, List[str]]:

    """Do these records represent the same transaction?"""

    date_score = date_proximity(r1['date'], r2['date'])

    if r1['date'] == r2['date']:

        date_message = ''

    elif date_score > 0.0:

        diff = abs((r1['date'] - r2['date']).days)

        date_message = f'+/- {diff} days'

    else:

        date_message = 'date mismatch'

    if r1['amount'] == r2['amount']:

        amount_score, amount_message = 2.0, ''

    else:

        amount_score, amount_message = 0.0, 'amount mismatch'

    # We never consider payee if there's a check_id in the books.

    check_message = ''

    payee_message = ''

    # Sometimes we get unrelated numbers in the statement column with check-ids,

    # so we can't match based on the existence of a statement check-id.

    if r2['check_id']:

        payee_score = 0.0

        if r1['check_id'] and r2['check_id'] and r1['check_id'] == r2['check_id']:

            check_score = 1.0

        else:

            check_message = 'check-id mismatch'

            check_score = 0.0

    else:

        check_score = 0.0

        payee_score = payee_match(r1['payee'], r2['payee'])

        if payee_score > PAYEE_FULL_MATCH_THRESHOLD:

            payee_message = ''

        elif payee_score > PAYEE_PARTIAL_MATCH_THRESHOLD:

            payee_message = 'partial payee match'

        else:

            payee_message = 'payee mismatch'

    overall_score = (date_score + amount_score + check_score + payee_score) / 4

    overall_message = [

        m for m in [date_message, amount_message, check_message, payee_message] if m

    ]

    return overall_score, overall_message

def match_statement_and_books(

    statement_trans: List[Dict], books_trans: List[Dict]

) -> Tuple[List[Tuple[List, List, List]], List[Dict], List[Dict]]:

    """Match transactions between the statement and books.

    If matched, the books transaction is marked off so that it can

    only be matched once. Some transactions will be matched, some will

    be on the statement but not the books and some on the books but

    not the statement.

    Passes through any unmatched transactions.

    Currently we use the same matching logic for all types of

    statements. It's conceivable that you could have special cases to

    accurately match some types of statements, but that would be more

    work to maintain and test.

    """

    matches = []

    remaining_books_trans = []

    remaining_statement_trans = []

    for r1 in statement_trans:

        best_match_score = 0.0

        best_match_index = None

        best_match_note = []

        matches_found = 0

        for i, r2 in enumerate(books_trans):

            score, note = records_match(r1, r2)

            if score >= OVERALL_ACCEPTABLE_MATCH_THRESHOLD and score >= best_match_score:

                matches_found += 1

                best_match_score = score

                best_match_index = i

                best_match_note = note

        if (

            best_match_score > OVERALL_ACCEPTABLE_MATCH_THRESHOLD

            and matches_found == 1

            and 'check-id mismatch' not in best_match_note

            or best_match_score > OVERALL_EXCELLENT_MATCH_THRESHOLD

        ):

            matches.append(([r1], [books_trans[best_match_index]], best_match_note))

            # Don't try to make a second match against this books entry.

            if best_match_index is not None:

                del books_trans[best_match_index]

        else:

            remaining_statement_trans.append(r1)

    for r2 in books_trans:

        remaining_books_trans.append(r2)

    return matches, remaining_statement_trans, remaining_books_trans

def subset_match(

    statement_trans: List[dict], books_trans: List[dict]

) -> Tuple[List[Tuple[List, List, List]], List[Dict], List[Dict]]:

    """Match single statement transactions with multiple books transactions.

    Works similarly to match_statement_and_books in that it returns a

    list of matches and lists of remaining statement and books

    transactions.

    """

    matches = []

    remaining_books_trans = []

    remaining_statement_trans = []

    groups = itertools.groupby(books_trans, key=lambda x: (x['date'], x['payee']))

    for _, group in groups:

        best_match_score = 0.0

        best_match_index = None

        best_match_note = []

        matches_found = 0

        group_items = list(group)

        total = sum(x['amount'] for x in group_items)

        r2 = copy.copy(group_items[0])

        r2['amount'] = total

        for i, r1 in enumerate(statement_trans):

            score, note = records_match(r1, r2)

            if score >= OVERALL_ACCEPTABLE_MATCH_THRESHOLD and score >= best_match_score:

                matches_found += 1

                best_match_score = score

                best_match_index = i

                best_match_note = note

        if (

            best_match_score > OVERALL_ACCEPTABLE_MATCH_THRESHOLD

            and matches_found == 1

            and 'check-id mismatch' not in best_match_note

            or best_match_score > OVERALL_EXCELLENT_MATCH_THRESHOLD

        ):

            matches.append(

                ([statement_trans[best_match_index]], group_items, best_match_note)

            )

            if best_match_index is not None:

                del statement_trans[best_match_index]

        else:

            remaining_books_trans.extend(group_items)

    for r1 in statement_trans:

        remaining_statement_trans.append(r1)

    return matches, remaining_statement_trans, remaining_books_trans

# TODO: Return list of tuples (instead of list of lists).

def format_matches(

    matches: List, csv_statement: str, show_reconciled_matches: bool

) -> List[List]:

    """Produce a list of body output lines from the given matches.

    The first column is a date so we can re-sort the list to put the

    missing entries in the right place. The second column is the text

    output.

    """

    match_output = []

    for r1s, r2s, note in matches:

        note = ', '.join(note)

        note = ': ' + note if note else note

        if r1s and r2s:

            if show_reconciled_matches or not all(x['bank_statement'] for x in r2s):

                if len(r2s) == 1: