archivdms/internal/storage/document_date.go

package storage

import (
	"regexp"
	"strconv"
	"strings"
	"time"
)

// documentDateMinYear mirrors dateMinYear in internal/api/date_extraction.go.
// The whole scoring logic below is duplicated (rather than shared) because the
// storage package must not import internal/api (that would create an import
// cycle: api already depends on storage). Kept in sync with the api heuristic —
// the same duplication pattern as heuristicTitle vs. titleFromOCRText.
const (
	documentDateMinYear             = 1990
	documentDateFutureToleranceDays = 2
	documentDateWindowRadius        = 40
	documentDateScoreNoContext      = 0.4
)

// documentDateKeyword mirrors dateKeyword in internal/api/date_extraction.go.
type documentDateKeyword struct {
	word  string
	score float64
}

var documentDateKeywords = []documentDateKeyword{
	{"rechnungsdatum", 0.9},
	{"belegdatum", 0.9},
	{"ausstellungsdatum", 0.9},
	{"rechnung vom", 0.9},
	{"beleg vom", 0.9},
	{"datum", 0.75},
	{"vom", 0.55},
}

// documentDateGermanMonths mirrors dateGermanMonths in the api package.
var documentDateGermanMonths = map[string]int{
	"januar": 1, "jan": 1,
	"februar": 2, "feb": 2,
	"märz": 3, "maerz": 3, "mär": 3, "mrz": 3,
	"april": 4, "apr": 4,
	"mai": 5,
	"juni": 6, "jun": 6,
	"juli": 7, "jul": 7,
	"august": 8, "aug": 8,
	"september": 9, "sep": 9, "sept": 9,
	"oktober": 10, "okt": 10,
	"november": 11, "nov": 11,
	"dezember": 12, "dez": 12,
}

// documentDateCandidateRe mirrors dateCandidateRe in
// internal/api/date_extraction.go. Recognised: DD.MM.YYYY, DD.MM.YY,
// DD/MM/YYYY, YYYY-MM-DD and spelled-out German month names ("15. März 2026").
var documentDateCandidateRe = regexp.MustCompile(
	`(?i)(?:\b(?P<gd>\d{1,2})\.(?P<gm>\d{1,2})\.(?P<gy>\d{4}|\d{2})\b)` +
		`|(?:\b(?P<sd>\d{1,2})/(?P<sm>\d{1,2})/(?P<sy>\d{4}|\d{2})\b)` +
		`|(?:\b(?P<iy>\d{4})-(?P<im>\d{1,2})-(?P<id>\d{1,2})\b)` +
		`|(?:\b(?P<td>\d{1,2})\.?\s+(?P<tmon>[A-Za-zäöüÄÖÜ]+)\.?\s+(?P<ty>\d{4})\b)`,
)

// documentDateScoreForPosition mirrors scoreForDatePosition in the api package.
func documentDateScoreForPosition(lowerText string, start int) float64 {
	lo := start - documentDateWindowRadius
	if lo < 0 {
		lo = 0
	}
	hi := start + documentDateWindowRadius
	if hi > len(lowerText) {
		hi = len(lowerText)
	}
	window := lowerText[lo:hi]
	for _, kw := range documentDateKeywords {
		if strings.Contains(window, kw.word) {
			return kw.score
		}
	}
	return documentDateScoreNoContext
}

// documentDateParseMatch mirrors parseDateMatch in the api package.
func documentDateParseMatch(names, m []string) (time.Time, bool) {
	now := time.Now()
	maxDate := now.AddDate(0, 0, documentDateFutureToleranceDays)
	var day, month, year int
	var monthName string
	for i, name := range names {
		if m[i] == "" {
			continue
		}
		switch name {
		case "gd", "sd", "id", "td":
			day, _ = strconv.Atoi(m[i])
		case "gm", "sm", "im":
			month, _ = strconv.Atoi(m[i])
		case "gy", "sy":
			y, _ := strconv.Atoi(m[i])
			if len(m[i]) == 2 {
				y += 2000
			}
			year = y
		case "iy", "ty":
			year, _ = strconv.Atoi(m[i])
		case "tmon":
			monthName = m[i]
		}
	}
	if monthName != "" {
		mn, ok := documentDateGermanMonths[strings.ToLower(monthName)]
		if !ok {
			return time.Time{}, false
		}
		month = mn
	}
	if year < documentDateMinYear {
		return time.Time{}, false
	}
	if month < 1 || month > 12 {
		return time.Time{}, false
	}
	if day < 1 || day > 31 {
		return time.Time{}, false
	}
	d := time.Date(year, time.Month(month), day, 0, 0, 0, 0, time.UTC)
	if d.Day() != day || int(d.Month()) != month || d.Year() != year {
		return time.Time{}, false
	}
	if d.After(maxDate) {
		return time.Time{}, false
	}
	return d, true
}

// documentDateFromTextWithScore mirrors extractDocumentDateWithScore in the api
// package. Returns the best belegdatum candidate and its confidence.
func documentDateFromTextWithScore(ocrText string) (best time.Time, score float64, found bool) {
	if ocrText == "" {
		return time.Time{}, 0, false
	}
	lower := strings.ToLower(ocrText)
	idxMatches := documentDateCandidateRe.FindAllStringSubmatchIndex(ocrText, -1)
	names := documentDateCandidateRe.SubexpNames()
	for _, loc := range idxMatches {
		m := make([]string, len(names))
		for g := range names {
			s, e := loc[2*g], loc[2*g+1]
			if s >= 0 {
				m[g] = ocrText[s:e]
			}
		}
		d, ok := documentDateParseMatch(names, m)
		if !ok {
			continue
		}
		sc := documentDateScoreForPosition(lower, loc[0])
		if !found || sc > score {
			best, score, found = d, sc, true
		}
	}
	return best, score, found
}