Tighten API contract and simplify documentation

README.md

@@ -0,0 +1,33 @@
+# Поиск товара по сообщению
+
+## 1. Запуск
+
+Нужен Python 3.12. Из корня проекта:
+
+```bash
+python -m pip install -r requirements.txt && python -m app
+```
+
+Сервис поднимется на `http://localhost:8000`, метод — `POST /match`.
+
+## 2. Решения и компромиссы
+
+Каталог небольшой — 466 позиций, поэтому он один раз загружается в память. Запрос нормализуется, из него извлекаются тип товара, бренд, модель и характеристики. Несовместимые позиции отсекаются точными правилами, оставшиеся ранжируются с помощью BM25, символьных триграмм и Damerau–Levenshtein.
+
+Чистое нечёткое ранжирование я отбросил: оно помогает с опечатками, но не должно решать, совместимы ли `125` и `150`, `T30` и `T50` или `3×2.5` и `4×2.5`. Поэтому текстовая похожесть отвечает за поиск кандидатов, а числа, модели и характеристики — за жёсткую фильтрацию. Elasticsearch позволил бы собрать похожую схему, но для 466 строк добавил бы отдельный сервис и лишнюю настройку без практической пользы.
+
+Алиасы хранятся как «канонический тип → несколько вариантов»: например, одно понятие может находиться по полному названию, сокращению или разговорному слову. Регулярные выражения используются для нормализации и разбора размеров, резьбы, мощности и других форматов; пустые сообщения отсекаются отдельно.
+
+Трансформерные модели запрещены условиями задачи, поэтому я их не использовал. При сомнении сервис возвращает `ambiguous`, а не угадывает. `confidence` — оценка для сортировки, не вероятность.
+
+До продакшена я бы вынес алиасы и правила из кода, проверил качество на размеченных реальных запросах, откалибровал пороги, добавил метрики, журнал причин отказа и нагрузочные тесты.
+
+## 3. Масштаб
+
+При 2 млн позиций и 50 магазинах каталог нельзя загружать и перебирать в каждом процессе. В PostgreSQL я бы хранил `store_id`, `sku`, нормализованное название, тип, бренд, модель и часто используемые характеристики в отдельных типизированных колонках; редкие атрибуты — в `JSONB`.
+
+Нужны как минимум уникальный B-tree по `(store_id, sku)` и составные B-tree-индексы, начинающиеся со `store_id`, для точных фильтров — например `(store_id, product_type, brand)`. Для размеров и других чисел индексы подбирались бы по реальным запросам. По названию B-tree не поможет: там нужны GIN-индексы для PostgreSQL FTS и/или `pg_trgm`.
+
+Поиск останется двухэтапным: PostgreSQL по магазину, типу и точным характеристикам выберет небольшой набор кандидатов, приложение проверит совместимость и сформирует ответ. `store_id` должен приходить из контекста запроса и участвовать во всех запросах к БД; при необходимости таблицу можно секционировать по нему.
+
+Если снять ограничение на трансформеры, я бы добавил семантический поиск по эмбеддингам BERT-подобной модели и хранил векторы в `pgvector`. Большую модель вроде Gemma можно использовать для разбора запроса или переранжирования кандидатов. Точные фильтры всё равно останутся: векторы ненадёжны для размеров, моделей и артикулов. Загрузку каталогов я сделал бы версионной: импорт во временную версию, проверка и атомарное переключение без простоя.

app/attributes.py

@@ -30,7 +30,11 @@ _VOLTAGE_RE = re.compile(
     rf"(?<![a-zа-я0-9-])({_NUMBER})\s*(?:v|в|вольт\w*)\b", re.IGNORECASE
 )
 _POWER_RE = re.compile(rf"\b({_NUMBER})\s*(?:w|вт|ватт\w*)\b", re.IGNORECASE)
-_TEETH_RE = re.compile(r"\b(\d+)\s*(?:зуб(?:ьев|а|ов)?|z)\b", re.IGNORECASE)
+_TEETH_RE = re.compile(r"\b(\d+)\s*(?:зуб(?:ьев|а|ов)?|[zt])\b", re.IGNORECASE)
+_SAW_COMPACT_RE = re.compile(
+    rf"(?<![a-zа-я0-9.])({_NUMBER})(?:x|\s+)(\d+)\s*[zt]\b",
+    re.IGNORECASE,
+)
 _PACK_SIZE_RE = re.compile(
     r"(?:пачк\w*|упаковк\w*|\bуп\.?)(?:\s+(?:по|на))?\s*(\d+)\s*(?:шт\w*)?",
     re.IGNORECASE,
@@ -40,6 +44,7 @@ _MODEL_RE = re.compile(
     r"(?<![a-zа-я0-9])([a-zа-я]{2,6}-\d+[a-zа-я]?|[a-zа-я]{2,6}\d+[a-zа-я]?)(?![a-zа-я0-9])",
     re.IGNORECASE,
 )
+_NORMALIZED_MODEL_RE = re.compile(r"^([a-zа-я]{2,6})(\d+)([a-zа-я]?)$", re.IGNORECASE)
 
 _PAIR_DIMENSION_TYPES = frozenset(
     {
@@ -225,6 +230,31 @@ def _extract_model_codes(normalized: str) -> frozenset[str]:
     return frozenset(codes)
 
 
+def model_code_matches(query_code: str, candidate_code: str) -> bool:
+    """Match a complete model or a model with only its final letter omitted.
+
+    Numeric parts remain exact. Thus ``арс12`` may complete to
+    ``арс12л``, but it must not match the different model ``арс125``.
+    """
+
+    if query_code == candidate_code:
+        return True
+
+    query_match = _NORMALIZED_MODEL_RE.fullmatch(query_code)
+    candidate_match = _NORMALIZED_MODEL_RE.fullmatch(candidate_code)
+    if query_match is None or candidate_match is None:
+        return False
+
+    query_prefix, query_number, query_suffix = query_match.groups()
+    candidate_prefix, candidate_number, candidate_suffix = candidate_match.groups()
+    return (
+        query_prefix == candidate_prefix
+        and query_number == candidate_number
+        and not query_suffix
+        and bool(candidate_suffix)
+    )
+
+
 def _detect_sds(normalized: str) -> str | None:
     flattened = normalized.replace("-", " ")
     if "sds max" in flattened or "сдс макс" in flattened:
@@ -377,8 +407,15 @@ def parse_attributes(
     normalized = normalize_text(text)
     dimensions = _extract_dimensions(normalized)
 
+    compact_saw_match = _SAW_COMPACT_RE.search(normalized)
     teeth_match = _TEETH_RE.search(normalized)
-    teeth = int(teeth_match.group(1)) if teeth_match else None
+    teeth = (
+        int(compact_saw_match.group(2))
+        if compact_saw_match is not None
+        else int(teeth_match.group(1))
+        if teeth_match is not None
+        else None
+    )
 
     if not dimensions and product_type in _PAIR_DIMENSION_TYPES:
         pair_match = _PAIR_WITH_NA_RE.search(normalized)
@@ -395,6 +432,8 @@ def parse_attributes(
 
     mm_values = [_decimal(match.group(1)) for match in _MM_RE.finditer(normalized)]
     meter_values = [_decimal(match.group(1)) for match in _METER_RE.finditer(normalized)]
+    if compact_saw_match is not None:
+        mm_values.append(_decimal(compact_saw_match.group(1)))
 
     # ``190 на 48 зубьев`` describes diameter and tooth count, not 190x48.
     saw_pair = re.search(
@@ -632,7 +671,16 @@ def compare_attributes(
             return conflict
 
     if query.model_codes:
-        conflict = require(query.model_codes.issubset(candidate.model_codes), "model")
+        conflict = require(
+            all(
+                any(
+                    model_code_matches(query_code, candidate_code)
+                    for candidate_code in candidate.model_codes
+                )
+                for query_code in query.model_codes
+            ),
+            "model",
+        )
         if conflict:
             return conflict
 

app/dev_routes.py

@@ -0,0 +1,39 @@
+from __future__ import annotations
+
+from decimal import Decimal
+
+from fastapi import FastAPI, HTTPException
+from pydantic import BaseModel, ConfigDict
+
+from app.matcher import CatalogMatcher
+
+
+class CatalogItemInfo(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+    sku: str
+    name: str
+    unit: str
+    price: Decimal
+
+
+def register_test_info_route(
+    application: FastAPI, matcher: CatalogMatcher
+) -> None:
+    items_by_sku = {item.sku: item for item in matcher.items}
+
+    @application.get(
+        "/get_info_by/{sku}",
+        response_model=CatalogItemInfo,
+        include_in_schema=False,
+    )
+    def get_info_by(sku: str) -> CatalogItemInfo:
+        item = items_by_sku.get(sku.strip().upper())
+        if item is None:
+            raise HTTPException(status_code=404, detail="SKU not found")
+        return CatalogItemInfo(
+            sku=item.sku,
+            name=item.name,
+            unit=item.unit,
+            price=item.price,
+        )

app/main.py

@@ -4,6 +4,7 @@ import os
 from pathlib import Path
 
 from fastapi import FastAPI
+from fastapi.middleware.cors import CORSMiddleware
 
 from app.matcher import CatalogMatcher
 from app.models import MatchRequest, MatchResponse
@@ -11,18 +12,48 @@ from app.models import MatchRequest, MatchResponse
 BASE_DIR = Path(__file__).resolve().parent.parent
 CATALOG_PATH = Path(os.environ.get("CATALOG_PATH", BASE_DIR / "catalog_excel.csv"))
 MATCHER = CatalogMatcher(CATALOG_PATH)
+TEST_INFO_ROUTE_ENV = "ENABLE_TEST_INFO_ROUTE"
 
-# The task asks for one application endpoint. OpenAPI/Swagger routes are disabled
+
-# so the running service exposes only POST /match.
+def match(request: MatchRequest) -> MatchResponse:
-app = FastAPI(
+    return MatchResponse(results=[MATCHER.match(message) for message in request.messages])
+
+
+def create_app(*, enable_test_info_route: bool = False) -> FastAPI:
+    # OpenAPI/Swagger routes are disabled so the production application exposes
+    # exactly the single endpoint required by the task.
+    application = FastAPI(
         title="Offline catalog matcher",
         version="0.2.0",
         docs_url=None,
         redoc_url=None,
         openapi_url=None,
     )
+    application.add_api_route(
+        "/match", match, methods=["POST"], response_model=MatchResponse
+    )
+
+    if enable_test_info_route:
+        from app.dev_routes import register_test_info_route
+
+        application.add_middleware(
+            CORSMiddleware,
+            allow_origins=["*"],
+            allow_methods=["GET", "POST"],
+            allow_headers=["Content-Type"],
+        )
+        register_test_info_route(application, MATCHER)
+
+    return application
 
 
-@app.post("/match", response_model=MatchResponse)
+def _test_info_route_enabled() -> bool:
-def match(request: MatchRequest) -> MatchResponse:
+    return os.environ.get(TEST_INFO_ROUTE_ENV, "").strip().lower() in {
-    return MatchResponse(results=[MATCHER.match(message) for message in request.messages])
+        "1",
+        "true",
+        "yes",
+        "on",
+    }
+
+
+app = create_app(enable_test_info_route=_test_info_route_enabled())

app/matcher.py

@@ -5,10 +5,18 @@ from dataclasses import dataclass
 from pathlib import Path
 
 from app.aliases import ProductAliasIndex, ProductTypeMatch
-from app.attributes import BrandIndex, Compatibility, compare_attributes, parse_attributes
+from app.attributes import (
+    BrandIndex,
+    Compatibility,
+    ParsedAttributes,
+    compare_attributes,
+    model_code_matches,
+    parse_attributes,
+)
 from app.catalog import CatalogItem, build_catalog, load_raw_catalog
 from app.intents import detect_service_intent
 from app.models import Candidate, MatchResult
+from app.normalization import ZERO_WEIGHT_TOKENS, normalize_text, tokenize
 from app.search_index import LexicalIndex, LexicalScore
 
 
@@ -20,6 +28,43 @@ class _RankedCandidate:
     ranking_score: float
 
 
+_STRUCTURED_QUERY_WORDS = frozenset(
+    {
+        "x",
+        "max",
+        "plus",
+        "sds",
+        "t",
+        "z",
+        "v",
+        "w",
+        "в",
+        "ватт",
+        "ватта",
+        "ваттов",
+        "вольт",
+        "вольта",
+        "вольтов",
+        "зуб",
+        "зуба",
+        "зубов",
+        "зубьев",
+        "м",
+        "метр",
+        "метра",
+        "метров",
+        "миллиметр",
+        "миллиметра",
+        "миллиметров",
+        "мм",
+        "размер",
+        "размера",
+        "размером",
+        "вт",
+    }
+)
+
+
 class CatalogMatcher:
     def __init__(self, catalog_path: Path) -> None:
         self.alias_index = ProductAliasIndex()
@@ -31,8 +76,11 @@ class CatalogMatcher:
 
         type_index: dict[str, list[int]] = defaultdict(list)
         model_index: dict[str, set[int]] = defaultdict(set)
+        brand_item_index: dict[str, list[int]] = defaultdict(list)
         for index, item in enumerate(self.items):
             type_index[item.product_type].append(index)
+            if item.attributes.brand is not None:
+                brand_item_index[item.attributes.brand].append(index)
             for model_code in item.attributes.model_codes:
                 model_index[model_code].add(index)
         self.type_index = {
@@ -41,6 +89,9 @@ class CatalogMatcher:
         self.model_index = {
             model: frozenset(indexes) for model, indexes in model_index.items()
         }
+        self.brand_item_index = {
+            brand: tuple(indexes) for brand, indexes in brand_item_index.items()
+        }
 
     def match(self, message: str) -> MatchResult:
         if not message or not message.strip():
@@ -59,19 +110,57 @@ class CatalogMatcher:
                 product_types = {self.items[index].product_type for index in model_candidates}
                 if len(product_types) == 1:
                     product_type = next(iter(product_types))
+                    exact_model = all(
+                        code in self.model_index
+                        for code in provisional_attributes.model_codes
+                    )
                     product_match = ProductTypeMatch(
                         canonical=product_type,
+                        score=1.0 if exact_model else 0.93,
+                        exact=exact_model,
+                        alias="model_code" if exact_model else "model_code_prefix",
+                        token_start=0,
+                    )
+            if product_match is None and provisional_attributes.teeth is not None:
+                # A tooth count is type-specific evidence in this domain: only
+                # saw blades carry it. This safely supports requests such as
+                # "210 мм 24 зуба" without requiring the word "диск".
+                product_match = ProductTypeMatch(
+                    canonical="диск",
                     score=1.0,
                     exact=True,
-                        alias="model_code",
+                    alias="tooth_count",
                     token_start=0,
                 )
             if product_match is None:
-                # Product evidence always has priority. Service intent detection
+                brand_result = self._match_brand_only(message, provisional_attributes)
-                # is only a diagnostic fallback; both unknown and service
+                if brand_result is not None:
-                # messages share the required not_found response.
+                    return brand_result
-                detect_service_intent(message)
+                structured_result = self._match_structured_only(
-                return MatchResult(message=message, status="not_found", candidates=[])
+                    message, provisional_attributes
+                )
+                if structured_result is not None:
+                    return structured_result
+                # Trigrams used to rank only inside a type selected by aliases.
+                # Let them infer that missing type as well, but only when the
+                # leading catalog rows agree and share strong word-level evidence.
+                # This handles "удар" -> "ударная дрель" while
+                # keeping numeric coincidences and arbitrary text out.
+                if detect_service_intent(message) is None:
+                    lexical_inference = self.lexical_index.infer_product_type(message)
+                    if lexical_inference is not None:
+                        product_match = ProductTypeMatch(
+                            canonical=lexical_inference.product_type,
+                            score=lexical_inference.score,
+                            exact=lexical_inference.query_word
+                            == lexical_inference.catalog_word,
+                            alias=lexical_inference.catalog_word,
+                            token_start=0,
+                        )
+                if product_match is None:
+                    return MatchResult(
+                        message=message, status="not_found", candidates=[]
+                    )
 
         product_type = product_match.canonical
         query_attributes = parse_attributes(
@@ -132,10 +221,197 @@ class CatalogMatcher:
             candidates=response_candidates,
         )
 
+    def _match_brand_only(
+        self, message: str, query_attributes: ParsedAttributes
+    ) -> MatchResult | None:
+        if (
+            query_attributes.brand is None
+            or not query_attributes.brand_required
+            or query_attributes.model_codes
+            or query_attributes.specificity != 1
+        ):
+            return None
+
+        candidate_indices = self.brand_item_index.get(query_attributes.brand, ())
+        brand_match = self.brand_index.detect(message)
+        if not candidate_indices or brand_match is None:
+            return None
+
+        compatible = {
+            index: Compatibility(
+                compatible=True,
+                matched_fields=1,
+                total_fields=1,
+            )
+            for index in candidate_indices
+        }
+        lexical_scores = self.lexical_index.rank(message, "", candidate_indices)
+        ranked = self._apply_preferences(
+            lexical_scores,
+            compatible,
+            cheap=query_attributes.cheap_preference,
+        )
+        if not ranked:
+            return None
+        if not query_attributes.cheap_preference:
+            # With no product type every item has the same brand evidence.
+            # Document length must not arbitrarily favor short catalog names.
+            ranked.sort(key=lambda candidate: candidate.item.sku)
+
+        status = "matched" if len(ranked) == 1 else "ambiguous"
+        brand_signal = ProductTypeMatch(
+            canonical="",
+            score=brand_match.score,
+            exact=brand_match.score == 1.0,
+            alias=brand_match.alias,
+            token_start=brand_match.token_start,
+        )
+        candidates = [
+            Candidate(
+                sku=candidate.item.sku,
+                confidence=self._confidence(
+                    candidate,
+                    product_match=brand_signal,
+                    candidate_count=len(ranked),
+                    specificity=query_attributes.specificity,
+                    status=status,
+                    rank=rank,
+                ),
+            )
+            for rank, candidate in enumerate(ranked[:3])
+        ]
+        return MatchResult(message=message, status=status, candidates=candidates)
+
+    def _match_structured_only(
+        self, message: str, query_attributes: ParsedAttributes
+    ) -> MatchResult | None:
+        if not self._is_structured_only_query(message, query_attributes):
+            return None
+
+        compatible: dict[int, Compatibility] = {}
+        for index, item in enumerate(self.items):
+            comparison = compare_attributes(
+                item.product_type,
+                query_attributes,
+                item.attributes,
+                candidate_unit=item.unit,
+            )
+            if comparison.compatible and comparison.matched_fields > 0:
+                compatible[index] = comparison
+
+        if not compatible:
+            return None
+
+        lexical_scores = self.lexical_index.rank(
+            message, "", tuple(compatible.keys())
+        )
+        ranked = self._apply_preferences(
+            lexical_scores,
+            compatible,
+            cheap=query_attributes.cheap_preference,
+        )
+        if not ranked:
+            return None
+
+        status = "matched" if len(ranked) == 1 else "ambiguous"
+        selected = self._select_diverse_product_types(ranked, limit=3)
+        structured_signal = ProductTypeMatch(
+            canonical="",
+            score=0.90,
+            exact=True,
+            alias="structured_attributes",
+            token_start=0,
+        )
+        candidates = [
+            Candidate(
+                sku=candidate.item.sku,
+                confidence=self._confidence(
+                    candidate,
+                    product_match=structured_signal,
+                    candidate_count=len(ranked),
+                    specificity=query_attributes.specificity,
+                    status=status,
+                    rank=rank,
+                ),
+            )
+            for rank, candidate in enumerate(selected)
+        ]
+        return MatchResult(message=message, status=status, candidates=candidates)
+
+    def _is_structured_only_query(
+        self, message: str, query_attributes: ParsedAttributes
+    ) -> bool:
+        has_strong_evidence = any(
+            (
+                bool(query_attributes.dimensions),
+                bool(query_attributes.mm_values),
+                bool(query_attributes.meter_values),
+                query_attributes.thread_size is not None,
+                query_attributes.profile is not None,
+                query_attributes.grit is not None,
+                query_attributes.voltage_v is not None,
+                query_attributes.power_w is not None,
+                query_attributes.teeth is not None,
+                query_attributes.brand is not None
+                and query_attributes.brand_required,
+                bool(query_attributes.model_codes),
+                query_attributes.sds_type is not None,
+            )
+        )
+        if not has_strong_evidence:
+            return False
+
+        allowed_words = set(ZERO_WEIGHT_TOKENS) | set(_STRUCTURED_QUERY_WORDS)
+        brand_match = self.brand_index.detect(message)
+        if brand_match is not None:
+            allowed_words.update(tokenize(brand_match.alias))
+
+        for token in tokenize(normalize_text(message), already_normalized=True):
+            if any(character.isdigit() for character in token):
+                continue
+            if token in allowed_words:
+                continue
+            if token.startswith(("бюджет", "дешев", "недорог")):
+                continue
+            return False
+        return True
+
+    @staticmethod
+    def _select_diverse_product_types(
+        ranked: list[_RankedCandidate], *, limit: int
+    ) -> list[_RankedCandidate]:
+        if len({candidate.item.product_type for candidate in ranked}) == 1:
+            return ranked[:limit]
+
+        selected: list[_RankedCandidate] = []
+        product_types: set[str] = set()
+
+        for candidate in ranked:
+            if candidate.item.product_type in product_types:
+                continue
+            selected.append(candidate)
+            product_types.add(candidate.item.product_type)
+            if len(selected) == limit:
+                break
+        return selected
+
     def _model_candidates(self, model_codes: frozenset[str]) -> frozenset[int]:
         if not model_codes:
             return frozenset()
-        sets = [self.model_index.get(code, frozenset()) for code in model_codes]
+
+        sets: list[frozenset[int]] = []
+        for query_code in model_codes:
+            exact = self.model_index.get(query_code)
+            if exact:
+                sets.append(exact)
+                continue
+
+            partial_indexes: set[int] = set()
+            for candidate_code, indexes in self.model_index.items():
+                if model_code_matches(query_code, candidate_code):
+                    partial_indexes.update(indexes)
+            sets.append(frozenset(partial_indexes))
+
         if not sets or any(not indexes for indexes in sets):
             return frozenset()
         result = set(sets[0])
@@ -167,7 +443,7 @@ class CatalogMatcher:
                     if price_range == 0
                     else float((maximum_price - item.price) / price_range)
                 )
-                ranking_score = lexical.score * 0.68 + cheap_score * 0.32
+                ranking_score = cheap_score
             else:
                 ranking_score = lexical.score
             ranked.append(
@@ -179,10 +455,19 @@ class CatalogMatcher:
                 )
             )
 
+        if cheap:
+            ranked.sort(
+                key=lambda candidate: (
+                    candidate.item.price,
+                    -candidate.lexical.score,
+                    candidate.item.sku,
+                )
+            )
+        else:
             ranked.sort(
                 key=lambda candidate: (
                     -candidate.ranking_score,
-                candidate.item.price if cheap else candidate.item.sku,
+                    candidate.item.sku,
                 )
             )
         return ranked

app/models.py

@@ -8,7 +8,7 @@ from pydantic import BaseModel, ConfigDict, Field
 class MatchRequest(BaseModel):
     model_config = ConfigDict(extra="forbid")
 
-    messages: list[str] = Field(default_factory=list)
+    messages: list[str]
 
 
 class Candidate(BaseModel):

app/search_index.py

@@ -6,7 +6,12 @@ from dataclasses import dataclass
 from typing import TYPE_CHECKING
 
 from app.normalization import meaningful_tokens, normalize_text, tokenize, word_tokens
-from app.typo import character_trigrams, trigram_dice, word_similarity
+from app.typo import (
+    character_trigrams,
+    damerau_levenshtein_distance,
+    trigram_dice,
+    word_similarity,
+)
 
 if TYPE_CHECKING:
     from app.catalog import CatalogItem
@@ -22,6 +27,14 @@ class LexicalScore:
     exact_coverage: float
 
 
+@dataclass(frozen=True, slots=True)
+class ProductTypeInference:
+    product_type: str
+    score: float
+    query_word: str
+    catalog_word: str
+
+
 class LexicalIndex:
     """Small in-memory lexical index: BM25 + char trigrams + typo alignment."""
 
@@ -69,6 +82,87 @@ class LexicalIndex:
             counts.update(self._trigram_postings.get(trigram, ()))
         return tuple(index for index, _ in counts.most_common(limit))
 
+    def infer_product_type(self, query_text: str) -> ProductTypeInference | None:
+        """Infer a missing product type from stable lexical evidence.
+
+        Trigrams provide the shortlist, but are deliberately not sufficient on
+        their own: sizes and random strings can share a few character triples
+        with unrelated catalog rows. The leading rows must agree on one type
+        and share a catalog word that is an exact match, a morphological prefix,
+        or a tightly bounded typo of a query word.
+        """
+
+        candidate_indices = self.trigram_candidates(query_text, limit=60)
+        if len(candidate_indices) < 3:
+            return None
+
+        leaders = self.rank(query_text, "", candidate_indices)[:3]
+        if len(leaders) < 3:
+            return None
+
+        product_types = {
+            self._items[leader.item_index].product_type for leader in leaders
+        }
+        if len(product_types) != 1:
+            return None
+
+        common_words = set(self._document_words[leaders[0].item_index])
+        for leader in leaders[1:]:
+            common_words.intersection_update(self._document_words[leader.item_index])
+        if not common_words:
+            return None
+
+        normalized = normalize_text(query_text)
+        query_words = word_tokens(
+            meaningful_tokens(tokenize(normalized, already_normalized=True))
+        )
+        best: tuple[float, str, str] | None = None
+        for query_word in query_words:
+            if len(query_word) < 4:
+                continue
+            for catalog_word in common_words:
+                score = self._type_inference_word_score(query_word, catalog_word)
+                if score <= 0:
+                    continue
+                evidence = (score, query_word, catalog_word)
+                if best is None or evidence > best:
+                    best = evidence
+
+        if best is None:
+            return None
+
+        score, query_word, catalog_word = best
+        return ProductTypeInference(
+            product_type=next(iter(product_types)),
+            score=score,
+            query_word=query_word,
+            catalog_word=catalog_word,
+        )
+
+    @staticmethod
+    def _type_inference_word_score(query_word: str, catalog_word: str) -> float:
+        if query_word == catalog_word:
+            return 1.0
+
+        shorter_length = min(len(query_word), len(catalog_word))
+        longer_length = max(len(query_word), len(catalog_word))
+        if shorter_length >= 4 and (
+            query_word.startswith(catalog_word)
+            or catalog_word.startswith(query_word)
+        ):
+            length_ratio = shorter_length / longer_length
+            if length_ratio >= 0.55:
+                return 0.72 + length_ratio * 0.18
+
+        distance = damerau_levenshtein_distance(query_word, catalog_word)
+        if shorter_length >= 5 and distance <= 1:
+            return 0.90
+        if longer_length >= 8 and distance <= 2:
+            edit_similarity = 1.0 - distance / longer_length
+            if edit_similarity >= 0.75:
+                return 0.82
+        return 0.0
+
     def rank(
         self,
         query_text: str,

more_messages.txt

@@ -0,0 +1,122 @@
+добрый день, саморезы по дереву 3.5х25 пачка 200 штук
+саморез гкл 3,5 на 25 кг
+нужны черные саморезы 4.2x75 упаковка 1000
+самарезы па дереву 3.5х45 200шт
+саморезы 99х999
+дюбель гвоздь 6x40
+дюбеля 8 на 80 есть?
+анкерный болт м10
+гайка м8 оцинкованная
+шайбы м12 сколько стоят
+болт м6х30 50 штук
+шпилька м10 один метр
+крепеж какой-нибудь
+кабель ввг нг ls 3*2,5
+провод пвс 2х1.5 нужен
+ШВВП 2 X 0,75, цена за метр?
+кабель кг 3х2,5
+кабель для чайника
+3х1.5
+ввгнг 20х20
+труба профильная 40x20x2
+профтруба 60 40 3
+труба квадратная двадцать на двадцать
+труба 35х35х2
+нужна труба длиной до луны
+гипсокартон 12,5 влагостойкий
+гкл обычный 9.5 мм
+лист гипса зеленый
+гипсокортан 12.5
+фанера 18 мм
+дрель prowerk pw-750
+аккумуляторная дрель 18в
+шуруповёрт 12 вольт самый дешёвый
+перфоратор sds plus 800 вт
+болгарка 125
+ушм 230 мм 2200 вт
+лобзик посоветуйте
+makita df333 есть?
+инструмент чтобы сверлить стену
+дрэл 9999 ват
+бур sds-plus 8x160
+бур 10 на 210 sds+
+бур SDS MAX 20х400
+сверло по металлу 6 мм
+сверло по дереву 10x120
+бита ph2 25
+набор бит torx
+бита t30 50 мм
+бита t99
+насадка на шуруповерт
+диск отрезной 125х1,0 металл
+круг по металлу 230 на 2.5
+пильный диск 190x48t
+алмазный диск 125 бетон
+диск на болгарку
+круг квадратный 12х34
+наждачная бумага p80
+шкурка нулёвка
+круг лепестковый 125 р40
+лента фум 19 мм
+изолента синяя 15 мм
+хомут 4,8x300 черный
+стяжки пластиковые
+перчатки нитриловые размер l
+перчатки хб с пвх
+мешки для мусора 120 литров
+пена монтажная зимняя
+герметик силиконовый прозрачный
+клей жидкие гвозди
+WD-40 400 мл
+здравствуйте
+вы сегодня открыты?
+какой адрес
+есть доставка в химки?
+можно вернуть товар без чека
+примите заказ 12345
+где мой заказ 000001
+оплатил, но денег нет
+спасибо большое!
+позовите человека
+1234567890
+000000000000000000
+42
+qwerty asdf zxcv
+йцукен фывапр олдж
+абракадабра крокодил вертолёт
+!@#$%^&*()
+..........
+купить купить купить
+null
+None
+SELECT * FROM catalog;
+<script>alert(1)</script>
+../../etc/passwd
+🤖🔩🔨
+САМОБРЕЗЫ ГКЛ 3.5Х25???
+ну эээ диск этот круглый короче
+мне два метра того красного
+хочу всё и сразу за сто рублей
+статус заказа 4512 и бур 8х160
+диск
+24 зуба
+210 мм 24 зуба
+125 мм
+12 в
+900 вт
+M10
+P120
+SDS-plus
+PH2
+20x20x2
+ToolKraft 125 мм
+99999 мм
+аквариум 125 мм
+удар
+ударная
+удар мощный
+удар мощно 900 хромированый анус долбить
+дрель удар мощно 900 хромированый анус долбить
+анус
+долбить
+Арс-12

tests/test_api.py

@@ -2,7 +2,7 @@ from __future__ import annotations
 
 from fastapi.testclient import TestClient
 
-from app.main import app
+from app.main import app, create_app
 
 client = TestClient(app)
 
@@ -38,12 +38,81 @@ def test_match_contract_and_message_order() -> None:
             assert 0.0 <= candidate["confidence"] <= 1.0
 
 
+def test_task_example_uses_required_request_and_response_shape() -> None:
+    message = "саморезы по дереву 4.2 на 75, пачку"
+    response = client.post("/match", json={"messages": [message]})
+
+    assert response.status_code == 200
+    body = response.json()
+    assert set(body) == {"results"}
+    assert len(body["results"]) == 1
+
+    result = body["results"][0]
+    assert set(result) == {"message", "status", "candidates"}
+    assert result["message"] == message
+    assert result["status"] in {"matched", "ambiguous", "not_found"}
+    for candidate in result["candidates"]:
+        assert set(candidate) == {"sku", "confidence"}
+        assert isinstance(candidate["sku"], str)
+        assert isinstance(candidate["confidence"], float)
+
+
 def test_empty_message_list_is_valid() -> None:
     response = client.post("/match", json={"messages": []})
     assert response.status_code == 200
     assert response.json() == {"results": []}
 
 
+def test_messages_field_is_required() -> None:
+    response = client.post("/match", json={})
+
+    assert response.status_code == 422
+
+
 def test_unknown_request_fields_are_rejected() -> None:
     response = client.post("/match", json={"messages": [], "debug": True})
     assert response.status_code == 422
+
+
+def test_opt_in_info_route_returns_catalog_item() -> None:
+    test_app = create_app(enable_test_info_route=True)
+    test_client = TestClient(test_app)
+
+    assert [route.path for route in test_app.routes] == [
+        "/match",
+        "/get_info_by/{sku}",
+    ]
+
+    response = test_client.get("/get_info_by/bit-0073")
+    assert response.status_code == 200
+    assert set(response.json()) == {"sku", "name", "unit", "price"}
+    assert response.json()["sku"] == "BIT-0073"
+    assert response.json()["name"]
+    assert response.json()["unit"]
+    assert float(response.json()["price"]) > 0
+
+
+def test_opt_in_info_route_returns_404_for_unknown_sku() -> None:
+    test_client = TestClient(create_app(enable_test_info_route=True))
+
+    response = test_client.get("/get_info_by/UNKNOWN-0000")
+
+    assert response.status_code == 404
+    assert response.json() == {"detail": "SKU not found"}
+
+
+def test_opt_in_routes_allow_requests_from_local_html_page() -> None:
+    test_client = TestClient(create_app(enable_test_info_route=True))
+    response = test_client.options(
+        "/match",
+        headers={
+            "Origin": "null",
+            "Access-Control-Request-Method": "POST",
+            "Access-Control-Request-Headers": "content-type",
+        },
+    )
+
+    assert response.status_code == 200
+    assert response.headers["access-control-allow-origin"] == "*"
+    assert "POST" in response.headers["access-control-allow-methods"]
+    assert "GET" in response.headers["access-control-allow-methods"]

tests/test_catalog_self_retrieval.py

@@ -0,0 +1,185 @@
+from __future__ import annotations
+
+from decimal import Decimal
+from pathlib import Path
+from typing import Callable
+
+import pytest
+
+from app.catalog import CatalogItem, RawCatalogItem, load_raw_catalog
+from app.matcher import CatalogMatcher
+
+
+CATALOG_ITEMS = load_raw_catalog(
+    Path(__file__).resolve().parents[1] / "catalog_excel.csv"
+)
+
+
+@pytest.mark.parametrize("item", CATALOG_ITEMS, ids=lambda item: item.sku)
+def test_full_catalog_name_retrieves_its_own_sku(
+    matcher: CatalogMatcher, item: RawCatalogItem
+) -> None:
+    result = matcher.match(item.name)
+
+    assert item.sku in {candidate.sku for candidate in result.candidates}
+
+
+def test_every_canonical_product_type_returns_only_its_own_items(
+    matcher: CatalogMatcher,
+) -> None:
+    items_by_sku = {item.sku: item for item in matcher.items}
+    product_types = {item.product_type for item in matcher.items}
+
+    for product_type in product_types:
+        result = matcher.match(product_type)
+
+        assert result.candidates, product_type
+        assert all(
+            items_by_sku[candidate.sku].product_type == product_type
+            for candidate in result.candidates
+        ), product_type
+
+
+def test_every_catalog_model_code_returns_only_matching_items(
+    matcher: CatalogMatcher,
+) -> None:
+    items_by_sku = {item.sku: item for item in matcher.items}
+    model_codes = {
+        model
+        for item in matcher.items
+        for model in item.attributes.model_codes
+    }
+
+    for model_code in model_codes:
+        result = matcher.match(model_code)
+
+        assert result.candidates, model_code
+        assert all(
+            model_code in items_by_sku[candidate.sku].attributes.model_codes
+            for candidate in result.candidates
+        ), model_code
+
+
+def test_every_catalog_brand_returns_only_its_own_items(
+    matcher: CatalogMatcher,
+) -> None:
+    items_by_sku = {item.sku: item for item in matcher.items}
+    brands = {
+        item.attributes.brand
+        for item in matcher.items
+        if item.attributes.brand is not None
+    }
+
+    for brand in brands:
+        result = matcher.match(brand)
+
+        assert result.candidates, brand
+        assert all(
+            items_by_sku[candidate.sku].attributes.brand == brand
+            for candidate in result.candidates
+        ), brand
+
+
+def _number_text(value: Decimal) -> str:
+    return format(value, "f")
+
+
+def _assert_query_returns_only(
+    matcher: CatalogMatcher,
+    query: str,
+    predicate: Callable[[CatalogItem], bool],
+) -> None:
+    items_by_sku = {item.sku: item for item in matcher.items}
+    result = matcher.match(query)
+
+    assert result.candidates, query
+    assert all(
+        predicate(items_by_sku[candidate.sku]) for candidate in result.candidates
+    ), query
+
+
+def test_every_catalog_millimetre_value_is_searchable_without_product_name(
+    matcher: CatalogMatcher,
+) -> None:
+    values = {
+        value
+        for item in matcher.items
+        for value in (
+            *item.attributes.mm_values,
+            *((item.attributes.dimensions[0],) if item.attributes.dimensions else ()),
+            *((item.attributes.thread_length_mm,) if item.attributes.thread_length_mm else ()),
+        )
+    }
+
+    for value in values:
+        _assert_query_returns_only(
+            matcher,
+            f"{_number_text(value)} мм",
+            lambda item, expected=value: (
+                expected in item.attributes.mm_values
+                or bool(
+                    item.attributes.dimensions
+                    and item.attributes.dimensions[0] == expected
+                )
+                or item.attributes.thread_length_mm == expected
+            ),
+        )
+
+
+def test_every_catalog_dimension_is_searchable_without_product_name(
+    matcher: CatalogMatcher,
+) -> None:
+    dimensions = {
+        item.attributes.dimensions
+        for item in matcher.items
+        if item.attributes.dimensions
+    }
+
+    for dimension in dimensions:
+        query = "x".join(_number_text(value) for value in dimension)
+        _assert_query_returns_only(
+            matcher,
+            query,
+            lambda item, expected=dimension: (
+                len(item.attributes.dimensions) >= len(expected)
+                and item.attributes.dimensions[: len(expected)] == expected
+            ),
+        )
+
+
+@pytest.mark.parametrize(
+    ("field", "suffix"),
+    [
+        ("voltage_v", " В"),
+        ("power_w", " Вт"),
+        ("thread_size", ""),
+        ("grit", ""),
+        ("profile", ""),
+        ("sds_type", ""),
+        ("teeth", " зубьев"),
+    ],
+)
+def test_every_scalar_catalog_attribute_is_searchable_without_product_name(
+    matcher: CatalogMatcher,
+    field: str,
+    suffix: str,
+) -> None:
+    values = {
+        getattr(item.attributes, field)
+        for item in matcher.items
+        if getattr(item.attributes, field) is not None
+    }
+
+    for value in values:
+        if field == "thread_size":
+            query = f"M{_number_text(value)}"
+        elif field == "grit":
+            query = f"P{value}"
+        else:
+            value_text = _number_text(value) if isinstance(value, Decimal) else str(value)
+            query = f"{value_text}{suffix}"
+        _assert_query_returns_only(
+            matcher,
+            query,
+            lambda item, expected=value: getattr(item.attributes, field) == expected,
+        )

tests/test_matcher.py

@@ -126,6 +126,105 @@ def test_product_evidence_has_priority_over_service_words(matcher: CatalogMatche
     assert result.candidates[0].sku == "BIT-0077"
 
 
+@pytest.mark.parametrize("brand", ["makita", "макита", "макиты", "макиту"])
+def test_makita_transliterations_share_the_same_brand_alias(
+    matcher: CatalogMatcher, brand: str
+) -> None:
+    result = matcher.match(f"шуруповерт как у {brand}, только дешевле")
+
+    assert result.status == "ambiguous"
+    assert [candidate.sku for candidate in result.candidates] == [
+        "INS-0017",
+        "INS-0010",
+        "INS-0003",
+    ]
+
+
+def test_cheap_preference_orders_candidates_by_ascending_price(
+    matcher: CatalogMatcher,
+) -> None:
+    result = matcher.match("шуруповерт как у макиты, только дешевле")
+    prices_by_sku = {item.sku: item.price for item in matcher.items}
+    candidate_prices = [
+        prices_by_sku[candidate.sku] for candidate in result.candidates
+    ]
+
+    assert candidate_prices == sorted(candidate_prices)
+
+
+@pytest.mark.parametrize(
+    "brand", ["ToolKraft", "toolkraft", "тулкрафт", "тул крафт"]
+)
+def test_brand_only_toolkraft_query_returns_catalog_items(
+    matcher: CatalogMatcher, brand: str
+) -> None:
+    result = matcher.match(brand)
+
+    assert result.status == "ambiguous"
+    assert [candidate.sku for candidate in result.candidates] == [
+        "INS-0022",
+        "INS-0023",
+        "INS-0024",
+    ]
+
+
+def test_unknown_toolkraft_model_does_not_fall_back_to_brand_results(
+    matcher: CatalogMatcher,
+) -> None:
+    result = matcher.match("ToolKraft TK-999")
+
+    assert result.status == "not_found"
+    assert result.candidates == []
+
+
+@pytest.mark.parametrize(
+    ("query", "sku"),
+    [
+        ("Арс-12", "INS-0017"),
+        ("АРС12", "INS-0017"),
+        ("АРС-18", "INS-0018"),
+        ("PW-12", "INS-0010"),
+        ("TK-12", "INS-0024"),
+        ("ТР-12", "INS-0003"),
+    ],
+)
+def test_model_code_can_omit_only_its_final_letter(
+    matcher: CatalogMatcher, query: str, sku: str
+) -> None:
+    result = matcher.match(query)
+
+    assert result.status == "matched"
+    assert [candidate.sku for candidate in result.candidates] == [sku]
+
+
+@pytest.mark.parametrize(
+    "query",
+    [
+        "АРС-1",
+        "АРС-999",
+        "PW-90",
+        "TK-999",
+        "УШМ АРС-12",
+    ],
+)
+def test_partial_model_does_not_guess_numeric_or_conflicting_models(
+    matcher: CatalogMatcher, query: str
+) -> None:
+    result = matcher.match(query)
+
+    assert result.status == "not_found"
+    assert result.candidates == []
+
+
+def test_complete_neighboring_model_keeps_exact_priority(
+    matcher: CatalogMatcher,
+) -> None:
+    result = matcher.match("АРС-125")
+
+    assert result.status == "matched"
+    assert [candidate.sku for candidate in result.candidates] == ["INS-0020"]
+
+
 @pytest.mark.parametrize("query", ["шурпуоверт на 12в", "шураыввавёрт на 12в"])
 def test_product_type_survives_typos(matcher: CatalogMatcher, query: str) -> None:
     result = matcher.match(query)
@@ -133,6 +232,55 @@ def test_product_type_survives_typos(matcher: CatalogMatcher, query: str) -> Non
     assert all(candidate.sku.startswith("INS-") for candidate in result.candidates)
 
 
+@pytest.mark.parametrize(
+    ("query", "skus"),
+    [
+        ("удар", ("INS-0008", "INS-0009", "INS-0022")),
+        ("ударная", ("INS-0008", "INS-0009", "INS-0022")),
+        ("удар мощный", ("INS-0008", "INS-0009", "INS-0022")),
+        (
+            "удар мощно 900 хромированый анус долбить",
+            ("INS-0009", "INS-0016", "INS-0023"),
+        ),
+        (
+            "дрель удар мощно 900 хромированый анус долбить",
+            ("INS-0009", "INS-0023", "INS-0002"),
+        ),
+    ],
+)
+def test_trigram_fallback_finds_product_type_from_sparse_or_noisy_text(
+    matcher: CatalogMatcher, query: str, skus: tuple[str, ...]
+) -> None:
+    result = matcher.match(query)
+
+    assert result.status == "ambiguous"
+    assert tuple(candidate.sku for candidate in result.candidates) == skus
+    items_by_sku = {item.sku: item for item in matcher.items}
+    assert {
+        items_by_sku[candidate.sku].product_type for candidate in result.candidates
+    } == {"дрель"}
+
+
+@pytest.mark.parametrize(
+    "query",
+    [
+        "анус",
+        "долбить",
+        "какой адрес",
+        "аквариум 125 мм",
+        "<script>alert(1)</script>",
+        "qwerty asdf zxcv",
+    ],
+)
+def test_trigram_fallback_rejects_weak_or_accidental_similarity(
+    matcher: CatalogMatcher, query: str
+) -> None:
+    result = matcher.match(query)
+
+    assert result.status == "not_found"
+    assert result.candidates == []
+
+
 def test_bare_size_is_parsed_without_unit(matcher: CatalogMatcher) -> None:
     assert matcher.match("бита ph2 50").candidates[0].sku == "BIT-0005"
     assert matcher.match("диск пильный 190 48 зубьев").candidates[0].sku == "DSK-0034"
@@ -144,3 +292,110 @@ def test_unknown_size_does_not_fall_back_to_nearest_product(
     result = matcher.match("ушм 150")
     assert result.status == "not_found"
     assert result.candidates == []
+
+
+@pytest.mark.parametrize(
+    ("query", "sku"),
+    [
+        ("165 мм 24 зуба", "DSK-0029"),
+        ("165 мм 36 зубов", "DSK-0030"),
+        ("165 мм 48 зубьев", "DSK-0031"),
+        ("190 мм 24 зуба", "DSK-0032"),
+        ("190 мм 36 зубов", "DSK-0033"),
+        ("190 мм 48 зубьев", "DSK-0034"),
+        ("210 мм 24 зуба", "DSK-0035"),
+        ("210 мм 36 зубов", "DSK-0036"),
+        ("210 мм 48 зубьев", "DSK-0037"),
+    ],
+)
+def test_tooth_count_infers_saw_blade_without_product_name(
+    matcher: CatalogMatcher, query: str, sku: str
+) -> None:
+    result = matcher.match(query)
+
+    assert result.status == "matched"
+    assert [candidate.sku for candidate in result.candidates] == [sku]
+
+
+@pytest.mark.parametrize(
+    ("query", "skus"),
+    [
+        ("24 зуба", ("DSK-0029", "DSK-0032", "DSK-0035")),
+        ("36 зубов", ("DSK-0030", "DSK-0033", "DSK-0036")),
+        ("48 зубьев", ("DSK-0031", "DSK-0034", "DSK-0037")),
+    ],
+)
+def test_bare_tooth_count_returns_all_matching_saw_blade_sizes(
+    matcher: CatalogMatcher, query: str, skus: tuple[str, ...]
+) -> None:
+    result = matcher.match(query)
+
+    assert result.status == "ambiguous"
+    assert tuple(candidate.sku for candidate in result.candidates) == skus
+
+
+def test_unknown_tooth_count_does_not_return_nearest_saw_blade(
+    matcher: CatalogMatcher,
+) -> None:
+    result = matcher.match("210 мм 25 зубьев")
+
+    assert result.status == "not_found"
+    assert result.candidates == []
+
+
+@pytest.mark.parametrize(
+    "query",
+    [
+        "пильный диск 190x48t",
+        "190x48T",
+        "190x48z",
+        "190 48t",
+        "190/48T",
+    ],
+)
+def test_compact_saw_notation_parses_diameter_and_teeth(
+    matcher: CatalogMatcher, query: str
+) -> None:
+    result = matcher.match(query)
+
+    assert result.status == "matched"
+    assert [candidate.sku for candidate in result.candidates] == ["DSK-0034"]
+
+
+def test_bare_t_suffix_infers_saw_blades(matcher: CatalogMatcher) -> None:
+    result = matcher.match("24T")
+
+    assert result.status == "ambiguous"
+    assert [candidate.sku for candidate in result.candidates] == [
+        "DSK-0029",
+        "DSK-0032",
+        "DSK-0035",
+    ]
+
+
+@pytest.mark.parametrize(
+    ("query", "status", "skus"),
+    [
+        ("125 мм", "ambiguous", ("INS-0013", "DSK-0025", "RAS-0054")),
+        ("12 В", "ambiguous", ("INS-0003", "INS-0010", "INS-0024")),
+        ("900 Вт", "ambiguous", ("INS-0009", "INS-0013")),
+        ("M10", "ambiguous", ("KRP-0030", "KRP-0020", "KRP-0009")),
+        ("P120", "ambiguous", ("DSK-0024", "RAS-0050", "RAS-0057")),
+        ("SDS-plus", "ambiguous", ("BIT-0073", "INS-0012")),
+        ("PH2", "ambiguous", ("BIT-0004", "RIN-0010")),
+        ("20x20x2", "matched", ("TRB-0003",)),
+        ("ToolKraft 125 мм", "matched", ("INS-0027",)),
+        ("99999 мм", "not_found", ()),
+        ("аквариум 125 мм", "not_found", ()),
+    ],
+)
+def test_attribute_only_queries_use_exact_catalog_values(
+    matcher: CatalogMatcher,
+    query: str,
+    status: str,
+    skus: tuple[str, ...],
+) -> None:
+    result = matcher.match(query)
+
+    assert result.status == status
+    assert tuple(candidate.sku for candidate in result.candidates) == skus

tests/test_message_corpus.py

@@ -0,0 +1,290 @@
+from __future__ import annotations
+
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Literal
+
+import pytest
+
+from app.matcher import CatalogMatcher
+
+
+ROOT = Path(__file__).resolve().parents[1]
+MESSAGE_PATHS = (ROOT / "messages.txt", ROOT / "more_messages.txt")
+MatchStatus = Literal["matched", "ambiguous", "not_found"]
+ExpectedResult = tuple[MatchStatus, tuple[str, ...]]
+
+
+@dataclass(frozen=True, slots=True)
+class MessageCase:
+    id: str
+    source: str
+    line_number: int
+    message: str
+
+
+def _load_message_cases() -> tuple[MessageCase, ...]:
+    cases: list[MessageCase] = []
+    for path in MESSAGE_PATHS:
+        for line_number, raw_line in enumerate(
+            path.read_text(encoding="utf-8").splitlines(), start=1
+        ):
+            message = raw_line.strip()
+            if not message:
+                continue
+            cases.append(
+                MessageCase(
+                    id=f"{path.name}:{line_number}",
+                    source=path.name,
+                    line_number=line_number,
+                    message=message,
+                )
+            )
+    return tuple(cases)
+
+
+MESSAGE_CASES = _load_message_cases()
+VALIDATED_MESSAGE_CASES = tuple(
+    case for case in MESSAGE_CASES if case.source == "messages.txt"
+)
+MORE_MESSAGE_CASES = tuple(
+    case for case in MESSAGE_CASES if case.source == "more_messages.txt"
+)
+
+# Hand-reviewed acceptance results for every required messages.txt query.
+# Do not regenerate this table from CatalogMatcher output: it is the independent
+# oracle that should catch semantic regressions in the implementation.
+VALIDATED_MESSAGE_RESULTS: dict[str, ExpectedResult] = {
+    "здравствуйте, есть саморезы гкл 3.5х25?": ("ambiguous", ("SAM-0063", "SAM-0061", "SAM-0062")),
+    "дрель ударная prowerk pw-750 в наличии?": ("matched", ("INS-0008",)),
+    "кабель шввп 2х0.5 сколько за метр": ("matched", ("KAB-0017",)),
+    "диск пильный 190 на 48 зубьев": ("matched", ("DSK-0034",)),
+    "уровень 1000 мм есть?": ("matched", ("RIN-0024",)),
+    "гайки м10 нужны": ("matched", ("KRP-0020",)),
+    "лента фум 12 мм": ("matched", ("RAS-0060",)),
+    "перчатки нитриловые есть?": ("matched", ("PER-0002",)),
+    "проф труба 20х20 стенка полтора": ("matched", ("TRB-0002",)),
+    "шурик на 12в недорогой": ("ambiguous", ("INS-0017", "INS-0010", "INS-0003")),
+    "болгарка на 230 какая есть": ("ambiguous", ("INS-0014", "INS-0028", "INS-0007")),
+    "гкл 9.5 сколько лист": ("matched", ("GKL-0001",)),
+    "хомуты пластиковые 4.8х400": ("matched", ("RAS-0014",)),
+    "сдс бур 6 на 110": ("matched", ("BIT-0074",)),
+    "наждачка р120 листами": ("matched", ("RAS-0050",)),
+    "круг зачистной на 125": ("matched", ("DSK-0018",)),
+    "бита ph2 50 мм": ("matched", ("BIT-0005",)),
+    "саморезы по дереву 3.5х45 пачка 200": ("matched", ("SAM-0010",)),
+    "труба профильная 20х20х2": ("matched", ("TRB-0003",)),
+    "ввгнг лс 3х1.5": ("matched", ("KAB-0010",)),
+    "диск отрезной по металлу 115х1.2": ("matched", ("DSK-0002",)),
+    "бур sds 10х210": ("matched", ("BIT-0080",)),
+    "дюбель-гвоздь 6х60": ("matched", ("KRP-0048",)),
+    "саморезы по металлу 4.2х70": ("not_found", ()),
+    "труба профильная 35х35": ("not_found", ()),
+    "кабель ввгнг 4х2.5": ("not_found", ()),
+    "ушм на 150": ("not_found", ()),
+    "бита t50": ("not_found", ()),
+    "шпилька м16 метровая": ("not_found", ()),
+    "нужен кабель": ("ambiguous", ("KAB-0004", "KAB-0005", "KAB-0008")),
+    "дайте дюбелей": ("ambiguous", ("KRP-0050", "KRP-0047", "KRP-0048")),
+    "сверло нужно": ("ambiguous", ("BIT-0060", "BIT-0061", "BIT-0062")),
+    "какие есть диски": ("ambiguous", ("DSK-0011", "DSK-0012", "DSK-0013")),
+    "перфоратор посоветуйте": ("ambiguous", ("INS-0012", "INS-0026", "INS-0005")),
+    "шуруповерт как у макиты, только дешевле": ("ambiguous", ("INS-0017", "INS-0010", "INS-0003")),
+    "здравствуйте, вы до скольки работаете?": ("not_found", ()),
+    "можно оплатить картой при получении?": ("not_found", ()),
+    "где находится ваш магазин": ("not_found", ()),
+    "статус заказа 4512 подскажите": ("not_found", ()),
+    "спасибо, заказ получил, все отлично": ("not_found", ()),
+}
+
+# The larger exploratory corpus is useful as a regression snapshot, but it is
+# intentionally kept separate from the hand-reviewed acceptance set above.
+MORE_MESSAGE_BASELINE_RESULTS: dict[str, ExpectedResult] = {
+    "добрый день, саморезы по дереву 3.5х25 пачка 200 штук": ("matched", ("SAM-0001",)),
+    "саморез гкл 3,5 на 25 кг": ("matched", ("SAM-0063",)),
+    "нужны черные саморезы 4.2x75 упаковка 1000": ("matched", ("SAM-0023",)),
+    "самарезы па дереву 3.5х45 200шт": ("ambiguous", ("SAM-0010", "SAM-0012", "SAM-0011")),
+    "саморезы 99х999": ("not_found", ()),
+    "дюбель гвоздь 6x40": ("matched", ("KRP-0047",)),
+    "дюбеля 8 на 80 есть?": ("matched", ("KRP-0049",)),
+    "анкерный болт м10": ("ambiguous", ("KRP-0009", "KRP-0010", "KRP-0011")),
+    "гайка м8 оцинкованная": ("matched", ("KRP-0019",)),
+    "шайбы м12 сколько стоят": ("ambiguous", ("KRP-0032", "KRP-0031")),
+    "болт м6х30 50 штук": ("ambiguous", ("KRP-0002", "KRP-0001", "KRP-0003")),
+    "шпилька м10 один метр": ("matched", ("KRP-0034",)),
+    "крепеж какой-нибудь": ("not_found", ()),
+    "кабель ввг нг ls 3*2,5": ("matched", ("KAB-0011",)),
+    "провод пвс 2х1.5 нужен": ("not_found", ()),
+    "ШВВП 2 X 0,75, цена за метр?": ("matched", ("KAB-0018",)),
+    "кабель кг 3х2,5": ("not_found", ()),
+    "кабель для чайника": ("ambiguous", ("KAB-0004", "KAB-0005", "KAB-0008")),
+    "3х1.5": ("ambiguous", ("KAB-0022", "KAB-0002", "KAB-0010")),
+    "ввгнг 20х20": ("not_found", ()),
+    "труба профильная 40x20x2": ("not_found", ()),
+    "профтруба 60 40 3": ("ambiguous", ("TRB-0009", "TRB-0010", "TRB-0015")),
+    "труба квадратная двадцать на двадцать": ("ambiguous", ("TRB-0003", "TRB-0007", "TRB-0008")),
+    "труба 35х35х2": ("not_found", ()),
+    "нужна труба длиной до луны": ("ambiguous", ("TRB-0003", "TRB-0007", "TRB-0008")),
+    "гипсокартон 12,5 влагостойкий": ("matched", ("GKL-0003",)),
+    "гкл обычный 9.5 мм": ("matched", ("GKL-0001",)),
+    "лист гипса зеленый": ("ambiguous", ("GKL-0001", "GKL-0002", "GKL-0004")),
+    "гипсокортан 12.5": ("ambiguous", ("GKL-0002", "GKL-0004", "GKL-0003")),
+    "фанера 18 мм": ("not_found", ()),
+    "дрель prowerk pw-750": ("matched", ("INS-0008",)),
+    "аккумуляторная дрель 18в": ("not_found", ()),
+    "шуруповёрт 12 вольт самый дешёвый": ("ambiguous", ("INS-0017", "INS-0010", "INS-0003")),
+    "перфоратор sds plus 800 вт": ("not_found", ()),
+    "болгарка 125": ("ambiguous", ("INS-0013", "INS-0027", "INS-0006")),
+    "ушм 230 мм 2200 вт": ("not_found", ()),
+    "лобзик посоветуйте": ("ambiguous", ("INS-0031", "INS-0029")),
+    "makita df333 есть?": ("not_found", ()),
+    "инструмент чтобы сверлить стену": ("ambiguous", ("BIT-0067", "BIT-0068", "BIT-0069")),
+    "дрэл 9999 ват": ("not_found", ()),
+    "бур sds-plus 8x160": ("matched", ("BIT-0077",)),
+    "бур 10 на 210 sds+": ("matched", ("BIT-0080",)),
+    "бур SDS MAX 20х400": ("not_found", ()),
+    "сверло по металлу 6 мм": ("matched", ("BIT-0045",)),
+    "сверло по дереву 10x120": ("not_found", ()),
+    "бита ph2 25": ("matched", ("BIT-0004",)),
+    "набор бит torx": ("not_found", ()),
+    "бита t30 50 мм": ("matched", ("BIT-0029",)),
+    "бита t99": ("not_found", ()),
+    "насадка на шуруповерт": ("ambiguous", ("INS-0010", "INS-0011", "INS-0024")),
+    "диск отрезной 125х1,0 металл": ("matched", ("DSK-0003",)),
+    "круг по металлу 230 на 2.5": ("not_found", ()),
+    "пильный диск 190x48t": ("matched", ("DSK-0034",)),
+    "алмазный диск 125 бетон": ("ambiguous", ("DSK-0025", "DSK-0026", "DSK-0027")),
+    "диск на болгарку": ("ambiguous", ("DSK-0011", "DSK-0012", "DSK-0013")),
+    "круг квадратный 12х34": ("not_found", ()),
+    "наждачная бумага p80": ("matched", ("RAS-0049",)),
+    "шкурка нулёвка": ("ambiguous", ("RAS-0049", "RAS-0048", "RAS-0052")),
+    "круг лепестковый 125 р40": ("matched", ("RAS-0054",)),
+    "лента фум 19 мм": ("not_found", ()),
+    "изолента синяя 15 мм": ("not_found", ()),
+    "хомут 4,8x300 черный": ("not_found", ()),
+    "стяжки пластиковые": ("ambiguous", ("RAS-0009", "RAS-0011", "RAS-0012")),
+    "перчатки нитриловые размер l": ("matched", ("PER-0002",)),
+    "перчатки хб с пвх": ("matched", ("PER-0001",)),
+    "мешки для мусора 120 литров": ("ambiguous", ("RAS-0035", "RAS-0034")),
+    "пена монтажная зимняя": ("ambiguous", ("RAS-0015", "RAS-0016")),
+    "герметик силиконовый прозрачный": ("ambiguous", ("RAS-0018", "RAS-0017")),
+    "клей жидкие гвозди": ("ambiguous", ("KRP-0043", "KRP-0044", "KRP-0045")),
+    "WD-40 400 мл": ("not_found", ()),
+    "здравствуйте": ("not_found", ()),
+    "вы сегодня открыты?": ("not_found", ()),
+    "какой адрес": ("not_found", ()),
+    "есть доставка в химки?": ("not_found", ()),
+    "можно вернуть товар без чека": ("not_found", ()),
+    "примите заказ 12345": ("not_found", ()),
+    "где мой заказ 000001": ("not_found", ()),
+    "оплатил, но денег нет": ("not_found", ()),
+    "спасибо большое!": ("not_found", ()),
+    "позовите человека": ("not_found", ()),
+    "1234567890": ("not_found", ()),
+    "000000000000000000": ("not_found", ()),
+    "42": ("not_found", ()),
+    "qwerty asdf zxcv": ("not_found", ()),
+    "йцукен фывапр олдж": ("not_found", ()),
+    "абракадабра крокодил вертолёт": ("not_found", ()),
+    "!@#$%^&*()": ("not_found", ()),
+    "..........": ("not_found", ()),
+    "купить купить купить": ("not_found", ()),
+    "null": ("not_found", ()),
+    "None": ("not_found", ()),
+    "SELECT * FROM catalog;": ("not_found", ()),
+    "<script>alert(1)</script>": ("not_found", ()),
+    "../../etc/passwd": ("not_found", ()),
+    "🤖🔩🔨": ("not_found", ()),
+    "САМОБРЕЗЫ ГКЛ 3.5Х25???": ("not_found", ()),
+    "ну эээ диск этот круглый короче": ("ambiguous", ("DSK-0011", "DSK-0012", "DSK-0013")),
+    "мне два метра того красного": ("not_found", ()),
+    "хочу всё и сразу за сто рублей": ("not_found", ()),
+    "статус заказа 4512 и бур 8х160": ("matched", ("BIT-0077",)),
+    "диск": ("ambiguous", ("DSK-0011", "DSK-0012", "DSK-0013")),
+    "24 зуба": ("ambiguous", ("DSK-0029", "DSK-0032", "DSK-0035")),
+    "210 мм 24 зуба": ("matched", ("DSK-0035",)),
+    "125 мм": ("ambiguous", ("INS-0013", "DSK-0025", "RAS-0054")),
+    "12 в": ("ambiguous", ("INS-0003", "INS-0010", "INS-0024")),
+    "900 вт": ("ambiguous", ("INS-0009", "INS-0013")),
+    "M10": ("ambiguous", ("KRP-0030", "KRP-0020", "KRP-0009")),
+    "P120": ("ambiguous", ("DSK-0024", "RAS-0050", "RAS-0057")),
+    "SDS-plus": ("ambiguous", ("BIT-0073", "INS-0012")),
+    "PH2": ("ambiguous", ("BIT-0004", "RIN-0010")),
+    "20x20x2": ("matched", ("TRB-0003",)),
+    "ToolKraft 125 мм": ("matched", ("INS-0027",)),
+    "99999 мм": ("not_found", ()),
+    "аквариум 125 мм": ("not_found", ()),
+    "удар": ("ambiguous", ("INS-0008", "INS-0009", "INS-0022")),
+    "ударная": ("ambiguous", ("INS-0008", "INS-0009", "INS-0022")),
+    "удар мощный": ("ambiguous", ("INS-0008", "INS-0009", "INS-0022")),
+    "удар мощно 900 хромированый анус долбить": ("ambiguous", ("INS-0009", "INS-0016", "INS-0023")),
+    "дрель удар мощно 900 хромированый анус долбить": ("ambiguous", ("INS-0009", "INS-0023", "INS-0002")),
+    "анус": ("not_found", ()),
+    "долбить": ("not_found", ()),
+    "Арс-12": ("matched", ("INS-0017",)),
+}
+
+EXPECTED_RESULTS = VALIDATED_MESSAGE_RESULTS | MORE_MESSAGE_BASELINE_RESULTS
+
+
+def test_message_files_are_present_and_have_expected_size() -> None:
+    counts = {
+        path.name: sum(
+            bool(line.strip())
+            for line in path.read_text(encoding="utf-8").splitlines()
+        )
+        for path in MESSAGE_PATHS
+    }
+    assert counts["messages.txt"] > 0
+    assert counts["more_messages.txt"] >= 100
+
+    messages = [case.message for case in MESSAGE_CASES]
+    assert len(messages) == len(set(messages)), "message texts must be unique"
+    assert set(EXPECTED_RESULTS) == set(messages)
+    assert set(VALIDATED_MESSAGE_RESULTS) == {
+        case.message for case in MESSAGE_CASES if case.source == "messages.txt"
+    }
+    assert set(MORE_MESSAGE_BASELINE_RESULTS) == {
+        case.message for case in MESSAGE_CASES if case.source == "more_messages.txt"
+    }
+
+
+def _assert_expected_result(
+    matcher: CatalogMatcher, case: MessageCase
+) -> None:
+    result = matcher.match(case.message)
+
+    assert result.message == case.message
+    assert result.status in {"matched", "ambiguous", "not_found"}
+    assert len(result.candidates) <= 3
+    assert len({candidate.sku for candidate in result.candidates}) == len(
+        result.candidates
+    )
+    assert all(0.0 <= candidate.confidence <= 1.0 for candidate in result.candidates)
+
+    if result.status == "matched":
+        assert len(result.candidates) == 1
+    elif result.status == "ambiguous":
+        assert 2 <= len(result.candidates) <= 3
+    else:
+        assert result.candidates == []
+
+    expected_status, expected_skus = EXPECTED_RESULTS[case.message]
+    assert result.status == expected_status
+    assert tuple(candidate.sku for candidate in result.candidates) == expected_skus
+
+
+@pytest.mark.parametrize(
+    "case", VALIDATED_MESSAGE_CASES, ids=lambda case: case.id
+)
+def test_validated_messages_txt_result(
+    matcher: CatalogMatcher, case: MessageCase
+) -> None:
+    _assert_expected_result(matcher, case)
+
+
+@pytest.mark.parametrize("case", MORE_MESSAGE_CASES, ids=lambda case: case.id)
+def test_more_messages_regression(
+    matcher: CatalogMatcher, case: MessageCase
+) -> None:
+    _assert_expected_result(matcher, case)