phase 2

detect/profiles/base.py

@@ -0,0 +1,71 @@
+"""
+ContentTypeProfile protocol and config dataclasses.
+
+The pipeline graph is fixed — what varies per content type is configuration
+and hooks. Each profile provides stage configs, a brand dictionary,
+VLM prompt templates, and an aggregation strategy.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import Protocol
+
+from detect.models import BrandDetection, DetectionReport
+
+
+@dataclass
+class FrameExtractionConfig:
+    fps: float = 2.0
+    max_frames: int = 500
+
+
+@dataclass
+class SceneFilterConfig:
+    hamming_threshold: int = 8
+    enabled: bool = True
+
+
+@dataclass
+class DetectionConfig:
+    model_name: str = "yolov8n.pt"
+    confidence_threshold: float = 0.3
+    target_classes: list[str] = field(default_factory=lambda: ["logo", "text"])
+
+
+@dataclass
+class OCRConfig:
+    languages: list[str] = field(default_factory=lambda: ["en"])
+    min_confidence: float = 0.5
+
+
+@dataclass
+class ResolverConfig:
+    fuzzy_threshold: int = 75
+
+
+@dataclass
+class BrandDictionary:
+    """Maps canonical brand name → list of known aliases/spellings."""
+    brands: dict[str, list[str]] = field(default_factory=dict)
+
+
+@dataclass
+class CropContext:
+    image: bytes
+    surrounding_text: str = ""
+    position_hint: str = ""
+
+
+class ContentTypeProfile(Protocol):
+    name: str
+
+    def frame_extraction_config(self) -> FrameExtractionConfig: ...
+    def scene_filter_config(self) -> SceneFilterConfig: ...
+    def detection_config(self) -> DetectionConfig: ...
+    def ocr_config(self) -> OCRConfig: ...
+    def brand_dictionary(self) -> BrandDictionary: ...
+    def resolver_config(self) -> ResolverConfig: ...
+    def vlm_prompt(self, crop_context: CropContext) -> str: ...
+    def aggregate(self, detections: list[BrandDetection]) -> DetectionReport: ...
+    def auxiliary_detections(self, source: str) -> list[BrandDetection]: ...