evidence.py

from __future__ import annotations

import re
from typing import Any

from normative_roles import build_normative_knowledge_unit, extract_normative_frame
from structural_parser import CLAUSE_MARKER_RES
from utils import normalize_for_search


EVIDENCE_SCHEMA_CLASS = "NormativeEvidenceSpan"


CONNECTIVE_STARTERS = (
    "ancak", "fakat", "ama", "bununla birlikte", "buna karşılık", "bu durumda",
    "şu kadar ki", "su kadar ki", "saklı kalmak", "saklidir", "saklıdır",
)

_STOPWORDS = {
    "madde", "kanun", "kanunda", "2547", "hangi", "neler", "nedir", "nasildir", "nasıl", "nasil",
    "gorevleri", "görevleri", "yetkileri", "sorumluluklari", "sorumlulukları", "duzenlenir", "düzenlenir",
    "var", "mi", "mı", "ne", "kac", "kaç", "kadar", "olarak", "icin", "için", "ilgili",
}


def build_evidence_spans(
    clause_id: str,
    article_id: str,
    label: str,
    clause_text: str,
    clause_source_span: dict[str, Any] | None = None,
    inherited_actors: list[Any] | None = None,
) -> list[dict[str, Any]]:
    """Split a clause into minimal evidence spans and attach deterministic role frames.

    v0.6 does not treat a full clause as the answer evidence unless the clause is already short.
    Long clauses are split by legal drafting boundaries: numbered items, sentence stops, semicolons,
    and exception/condition connectives. The resulting spans are small enough to highlight without
    painting an entire article or long paragraph.
    """
    text = " ".join(str(clause_text or "").split())
    if not text:
        return []

    local_units = _evidence_unit_spans(text)
    base_start = 0
    if clause_source_span and isinstance(clause_source_span.get("char_start"), int):
        base_start = int(clause_source_span.get("char_start", 0))

    spans: list[dict[str, Any]] = []
    for ordinal, (start, end) in enumerate(local_units, start=1):
        span_text = text[start:end].strip()
        if len(span_text) < 10:
            continue
        frame = extract_normative_frame(span_text, inherited_actors or [])
        roles = infer_evidence_roles(span_text, frame)
        evidence_id = f"{clause_id}__ev_{ordinal:03d}"
        source_ref = {
            "parent_clause_id": clause_id,
            "article_id": article_id,
            "label": label,
            "evidence_id": evidence_id,
        }
        spans.append({
            "evidence_id": evidence_id,
            "class": EVIDENCE_SCHEMA_CLASS,
            "parent_clause_id": clause_id,
            "article_id": article_id,
            "label": label,
            "norm_type": primary_evidence_norm_type(roles, frame),
            "semantic_roles": roles,
            "semantic_frame": frame,
            "source_span": {
                "article_id": article_id,
                "label": label,
                "clause_char_start": start,
                "clause_char_end": end,
                "char_start": base_start + start,
                "char_end": base_start + end,
            },
            "source_text": span_text,
            "normative_unit": build_normative_knowledge_unit(span_text, frame, source_ref),
        })
    return spans or [_fallback_evidence_span(clause_id, article_id, label, text, clause_source_span, inherited_actors)]


def select_best_evidence_spans(
    source_text: str,
    question: str = "",
    answer: str = "",
    max_spans: int = 4,
) -> list[tuple[int, int]]:
    """Return minimal source spans for UI highlighting.

    This function is intentionally independent from the ontology file so legacy paths still get
    fine-grained highlighting. It prefers evidence units that match query terms and answer terms;
    it never returns the whole source if smaller candidates exist.
    """
    text = str(source_text or "")
    if not text:
        return []
    units = _evidence_unit_spans(text)
    if len(units) == 1 and units[0] == (0, len(text)):
        return units if len(text) <= 500 else [_trim_window_around_terms(text, 0, len(text), question, answer)]

    q_terms = _meaningful_terms(question)
    a_terms = _meaningful_terms(answer)
    scored: list[tuple[float, int, int]] = []
    for start, end in units:
        unit_text = text[start:end]
        unit_norm = normalize_for_search(unit_text)
        score = 0.0
        score += 2.0 * _term_overlap_ratio(unit_norm, q_terms)
        score += 1.0 * _term_overlap_ratio(unit_norm, a_terms)
        if any(marker in unit_norm for marker in ("halinde", "takdirde", "kaydiyla", "kaydiyla", "ancak", "haric", "sakli")):
            if any(term in normalize_for_search(question) for term in ("hangi", "sart", "şart", "kosul", "istisna", "haric", "ancak")):
                score += 0.75
        if any(term in unit_norm for term in ("yil", "ay", "gun", "sure", "suresi")):
            if any(term in normalize_for_search(question) for term in ("kac", "kaç", "sure", "suresi", "ne kadar")):
                score += 0.75
        # Prefer compact evidence over very long units.
        length = max(1, end - start)
        score -= min(0.45, max(0, length - 420) / 1200)
        if score > 0:
            scored.append((score, start, end))

    if not scored:
        return []
    scored.sort(key=lambda item: item[0], reverse=True)
    spans = [(start, end) for _, start, end in scored[:max_spans]]
    return _merge_spans(spans)


def infer_evidence_roles(text: str, frame: dict[str, Any]) -> list[str]:
    q = normalize_for_search(text)
    roles: set[str] = set()
    if frame.get("definition_term") or frame.get("modality") == "definition":
        roles.add("tanim")
    if frame.get("condition"):
        roles.add("sart")
    if frame.get("exception"):
        roles.add("istisna")
    if frame.get("temporal_constraint"):
        roles.add("sure")
    if frame.get("procedure_step"):
        roles.add("usul")
    if frame.get("eligibility_criteria"):
        roles.add("sart")
    if frame.get("sanction"):
        roles.add("yaptirim")
    if frame.get("remedy"):
        roles.add("itiraz")
    if frame.get("quantitative_rule"):
        roles.add("olcut")
    if frame.get("competent_authority"):
        roles.add("yetki")
    if frame.get("beneficiary"):
        roles.add("hak")
    if any(action in set(frame.get("action", []) or []) for action in ("atamak", "seçmek")):
        roles.add("atama")
    if any(action in set(frame.get("action", []) or []) for action in ("kurmak", "açmak", "kapatmak")):
        roles.add("kurulus")
    if any(term in q for term in ("gorev", "görev", "baskanlik", "başkanlık", "rapor", "denetim", "yapmak", "bildirmek", "sunmak", "saglamak", "sağlamak")):
        roles.add("gorev")
    if any(term in q for term in ("yetki", "yetkili", "karar", "belirler", "tespit", "onay")):
        roles.add("yetki")
    if any(term in q for term in ("sorumlu", "zorundadir", "zorundadır", "yukumludur", "gerekir")):
        roles.add("sorumluluk")
    if any(term in q for term in ("ceza", "yaptirim", "ilişiği kes", "ilisigi kes", "kapatilir")):
        roles.add("yaptirim")
    if any(term in q for term in ("olusur", "oluşur", "uyeden", "üyeden", "toplam")):
        roles.add("kurul_olusumu")
    return sorted(roles) or ["norm"]


def primary_evidence_norm_type(roles: list[str], frame: dict[str, Any]) -> str:
    priority = [
        "kurulus", "atama", "yaptirim", "itiraz", "tanim", "sart", "istisna", "sure",
        "olcut", "hak", "kurul_olusumu", "gorev", "yetki", "sorumluluk", "usul", "norm",
    ]
    role_set = set(roles)
    for role in priority:
        if role in role_set:
            return role
    modality = frame.get("modality", "")
    if modality == "definition":
        return "tanim"
    if modality == "prohibition":
        return "yaptirim"
    if modality in {"power", "permission"}:
        return "yetki"
    if modality == "obligation":
        return "sorumluluk"
    return "norm"


def _evidence_unit_spans(text: str) -> list[tuple[int, int]]:
    text = str(text or "")
    if not text:
        return []
    boundaries = {0, len(text)}

    # Legal texts often encode independent rules with semicolons.
    for match in re.finditer(r"[.;!?]\s+", text):
        if match.group(0).startswith(";") and _semicolon_continues_predicate(text, match.end()):
            continue
        boundaries.add(match.end())

    # New lines and list markers are hard structural signals.
    for match in re.finditer(r"\n+", text):
        boundaries.add(match.end())
    # Reuse the exact legal-marker definitions used by the structural parser.
    # In particular, ``(b) fıkrası`` is an inline cross-reference rather than
    # a new evidence unit.  Only the additional unparenthesized ``1)`` form is
    # evidence-specific here.
    marker_patterns = [
        *CLAUSE_MARKER_RES,
        re.compile(r"(?<![\w(])\d+\)\s"),
    ]
    for pattern in marker_patterns:
        for match in pattern.finditer(text):
            if not _looks_like_date_context(text, match.start()):
                boundaries.add(match.start())

    # Split exception/condition tails even when they remain inside one sentence.
    norm = normalize_for_search(text)
    for starter in CONNECTIVE_STARTERS:
        # Match on original text approximately by looking for Turkish variants.
        for variant in _starter_variants(starter):
            for match in re.finditer(r"\b" + re.escape(variant) + r"\b", text, flags=re.IGNORECASE):
                if match.start() > 0:
                    boundaries.add(match.start())

    ordered = sorted(b for b in boundaries if 0 <= b <= len(text))
    spans: list[tuple[int, int]] = []
    for start, end in zip(ordered, ordered[1:]):
        start, end = _trim(text, start, end)
        if end <= start:
            continue
        unit = text[start:end].strip()
        if len(unit) > 720:
            spans.extend(_split_long_unit(text, start, end))
        else:
            spans.append((start, end))
    return spans or [(0, len(text))]


def _semicolon_continues_predicate(text: str, boundary: int) -> bool:
    """Keep one normative proposition intact across a drafting semicolon.

    Turkish provisions commonly place the authority at the end of a sequence:
    ``Cumhurbaşkanı ... kurmaya; ... değiştirmeye yetkilidir.``  Splitting at
    that semicolon separates actor/action from modality and makes a true power
    look unsupported to the proof contract.
    """
    before = normalize_for_search(text[max(0, boundary - 180) : boundary]).rstrip()
    after = normalize_for_search(text[boundary : boundary + 240])
    incomplete_infinitive = bool(
        re.search(r"\b[a-z0-9]+(?:maya|meye|maga|mege)\s*$", before)
    )
    completing_modality = bool(
        re.search(r"\b(?:yetkilidir|yetkilidirler)\b", after)
    )
    return incomplete_infinitive and completing_modality


def _split_long_unit(text: str, start: int, end: int) -> list[tuple[int, int]]:
    unit = text[start:end]
    local_boundaries = {0, len(unit)}

    # Split long legal lists and embedded subitems. This is structural, not a question patch.
    split_patterns = [
        r",\s+(?=(?:ancak|bu|ilgili|yükseköğretim|öğrenci|öğretim|üniversite|fakülte|enstitü|rektör|dekan)\b)",
        r",\s+(?=\d{3,5}\s+sayılı)",
        r"(?<!\w)\d+\)\s",
    ]
    for pattern in split_patterns:
        for match in re.finditer(pattern, unit, flags=re.IGNORECASE):
            boundary = match.start() if pattern.startswith("(?<!") else match.end()
            if 0 < boundary < len(unit):
                local_boundaries.add(boundary)

    ordered = sorted(local_boundaries)
    spans = []
    for a, b in zip(ordered, ordered[1:]):
        s, e = _trim(text, start + a, start + b)
        if e > s:
            spans.append((s, e))
    return spans or [(start, end)]


def _fallback_evidence_span(clause_id: str, article_id: str, label: str, text: str, clause_source_span: dict[str, Any] | None, inherited_actors: list[Any] | None) -> dict[str, Any]:
    frame = extract_normative_frame(text, inherited_actors or [])
    roles = infer_evidence_roles(text, frame)
    base_start = int((clause_source_span or {}).get("char_start", 0) or 0)
    return {
        "evidence_id": f"{clause_id}__ev_001",
        "class": EVIDENCE_SCHEMA_CLASS,
        "parent_clause_id": clause_id,
        "article_id": article_id,
        "label": label,
        "norm_type": primary_evidence_norm_type(roles, frame),
        "semantic_roles": roles,
        "semantic_frame": frame,
        "source_span": {"article_id": article_id, "label": label, "clause_char_start": 0, "clause_char_end": len(text), "char_start": base_start, "char_end": base_start + len(text)},
        "source_text": text,
        "normative_unit": build_normative_knowledge_unit(
            text,
            frame,
            {"parent_clause_id": clause_id, "article_id": article_id, "label": label, "evidence_id": f"{clause_id}__ev_001"},
        ),
    }


def _meaningful_terms(text: str) -> set[str]:
    normalized = normalize_for_search(text or "")
    terms = {term for term in normalized.split() if len(term) >= 4 and term not in _STOPWORDS}
    compact = normalized.replace(" ", "")
    if "onlisans" in compact:
        terms.update({"on", "lisans", "onlisans", "önlisans", "diploma"})
    return terms


def _term_overlap_ratio(unit_norm: str, terms: set[str]) -> float:
    if not terms:
        return 0.0
    hits = sum(1 for term in terms if normalize_for_search(term) in unit_norm)
    return hits / max(1, min(len(terms), 8))


def _trim_window_around_terms(text: str, start: int, end: int, question: str, answer: str) -> tuple[int, int]:
    terms = list(_meaningful_terms(question) or _meaningful_terms(answer))
    norm_text = normalize_for_search(text)
    idx = -1
    for term in terms:
        idx = norm_text.find(normalize_for_search(term))
        if idx >= 0:
            break
    if idx < 0:
        return (start, min(end, start + 480))
    return _expand_to_word_bounds(text, max(start, idx - 160), min(end, idx + 320))


def _expand_to_word_bounds(text: str, start: int, end: int) -> tuple[int, int]:
    while start > 0 and not text[start - 1].isspace():
        start -= 1
    while end < len(text) and not text[end - 1].isspace():
        end += 1
    return start, end


def _trim(text: str, start: int, end: int) -> tuple[int, int]:
    while start < end and text[start].isspace():
        start += 1
    while end > start and text[end - 1].isspace():
        end -= 1
    return start, end


def _merge_spans(spans: list[tuple[int, int]]) -> list[tuple[int, int]]:
    if not spans:
        return []
    ordered = sorted((max(0, s), max(0, e)) for s, e in spans if e > s)
    merged = [ordered[0]]
    for start, end in ordered[1:]:
        last_start, last_end = merged[-1]
        if start <= last_end:
            merged[-1] = (last_start, max(last_end, end))
        else:
            merged.append((start, end))
    return merged


def _looks_like_date_context(text: str, position: int) -> bool:
    window = text[max(0, position - 16):position + 16]
    return bool(re.search(r"\d{1,2}/\d{1,2}/\d{4}", window))


def _starter_variants(starter: str) -> list[str]:
    variants = {starter, starter.replace("ş", "s").replace("ı", "i").replace("ğ", "g").replace("ü", "u").replace("ö", "o").replace("ç", "c")}
    if starter == "ancak":
        variants.update({"Ancak"})
    return list(variants)