feat(etl): извлечение прайсов по форматам (xlsx/xls/docx/pdf) и распознавание сложных страниц через Gemini Vision

2026-06-26 15:10:00 +05:00
parent 3834e4811d
commit 26283d5432
8 changed files with 849 additions and 0 deletions
@@ -0,0 +1,84 @@
+"""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