"""Подбор порога нормализации: выборка позиций → каскад → покрытие при порогах."""

import random
import sys
from pathlib import Path

import numpy as np

REPO = Path(__file__).resolve().parents[2]
sys.path.insert(0, str(REPO))

from rapidfuzz import fuzz, process

from etl.dictionary import load_dictionary, normalize_name
from etl.extractors import extract
from etl.normalize.embedding import GeminiEmbedder
from etl.normalize.matcher import _CODE_CORE_RE

rows = []
for path in sorted((REPO / "data/raw").glob("*")):
    for row in extract(str(path)).rows:
        rows.append((row.service_name_raw, row.service_code_source))
random.seed(0)
sample = random.sample(rows, min(2000, len(rows)))
print(f"всего {len(rows)}, выборка {len(sample)}")

services = load_dictionary(REPO / "data/reference/dictionary.xlsx")
by_code = {s.tarificator_code: s for s in services if s.tarificator_code}
by_norm: dict[str, object] = {}
for s in services:
    by_norm.setdefault(s.name_norm, s)
names_norm = [s.name_norm for s in services]

embedder = GeminiEmbedder()
print("эмбеддинг справочника (1281)…")
dict_emb = embedder.encode([s.name_ru for s in services], task_type="RETRIEVAL_DOCUMENT")

code_n = exact_n = 0
pending: list[tuple[str, str]] = []
for name, code in sample:
    m = _CODE_CORE_RE.search(code) if code else None
    if m and m.group(0) in by_code:
        code_n += 1
        continue
    norm = normalize_name(name)
    if norm in by_norm:
        exact_n += 1
        continue
    pending.append((name, norm))

print(f"код: {code_n} | точное: {exact_n} | дальше эмбеддинги/fuzzy: {len(pending)}")
print("эмбеддинг запросов…")
query_emb = embedder.encode([p[0] for p in pending], task_type="RETRIEVAL_QUERY")
sims = query_emb @ dict_emb.T
top_idx = sims.argmax(axis=1)
top_score = sims.max(axis=1)
fuzzy_score = np.array(
    [process.extractOne(nn, names_norm, scorer=fuzz.token_set_ratio)[1] for _, nn in pending]
)

base = code_n + exact_n
print(f"\nБаза (код+точное): {base}/{len(sample)} = {100 * base / len(sample):.0f}%")
for thr in (0.58, 0.60, 0.62, 0.64, 0.66, 0.70):
    emb_matched = int((top_score >= thr).sum())
    fuzzy_matched = int(((fuzzy_score >= 88) & (top_score < thr)).sum())
    coverage = base + emb_matched + fuzzy_matched
    print(
        f"  порог {thr}: эмб {emb_matched} + fuzzy {fuzzy_matched} → покрытие {100 * coverage / len(sample):.0f}%"
    )

print("\nпримеры эмбеддинг-совпадений (порог 0.62):")
shown = 0
for k, (name, _) in enumerate(pending):
    if top_score[k] >= 0.62 and shown < 7:
        print(f"  «{name[:34]}» → «{services[int(top_idx[k])].name_ru[:34]}» ({top_score[k]:.2f})")
        shown += 1
print("примеры unmatched (top<0.62 и fuzzy<88):")
shown = 0
for k, (name, _) in enumerate(pending):
    if top_score[k] < 0.62 and fuzzy_score[k] < 88 and shown < 6:
        print(
            f"  «{name[:48]}» (лучший {services[int(top_idx[k])].name_ru[:22]} {top_score[k]:.2f})"
        )
        shown += 1