src/mcr/data/base.py

"""Core data abstractions shared by every benchmark.

The whole framework is built around two ideas:

* :class:`Content` — one structure that represents a query OR a document for any
  modality (plain text, image+text, or interleaved image-text). Loaders normalise
  each benchmark's idiosyncratic schema into ``Content``; everything downstream
  (encoders, retrievers, metrics) only ever sees ``Content``.
* :class:`RetrievalTask` — a single evaluatable unit (a domain/schema of a
  benchmark). It yields normalised corpus/queries and BEIR-style qrels, plus an
  optional per-query "difficulty" key for multi-condition stratified reporting.
"""

from __future__ import annotations

from abc import ABC, abstractmethod
from dataclasses import dataclass, field
from typing import Any, Dict, Iterator, List, Optional, Tuple, Union

# A qrels mapping: {query_id: {doc_id: relevance}} — the BEIR convention.
QrelsType = Dict[str, Dict[str, int]]


@dataclass
class Segment:
    """One piece of content: a text span and/or an image.

    ``image`` is either a filesystem path (lazily opened by the encoder) or an
    already-decoded ``PIL.Image.Image`` (used by MERIT, whose images live inside
    parquet files and are decoded in memory).
    """

    text: Optional[str] = None
    image: Optional[Any] = None  # str path | PIL.Image.Image | None

    @property
    def has_image(self) -> bool:
        return self.image is not None and self.image != ""

    @property
    def has_text(self) -> bool:
        return bool(self.text and self.text.strip())


@dataclass
class Content:
    """A query or a document, modality-agnostic.

    The ``segments`` list preserves order, which is what makes interleaved
    image-text queries (MERIT) representable in the same structure as a plain
    string (a single text segment).
    """

    segments: List[Segment] = field(default_factory=list)
    meta: Dict[str, Any] = field(default_factory=dict)

    # -- constructors -----------------------------------------------------
    @classmethod
    def from_text(cls, text: str, **meta: Any) -> "Content":
        return cls(segments=[Segment(text=text)], meta=dict(meta))

    @classmethod
    def from_image_text(cls, image: Any, text: str, **meta: Any) -> "Content":
        return cls(segments=[Segment(image=image), Segment(text=text)], meta=dict(meta))

    # -- views ------------------------------------------------------------
    def as_text(self, sep: str = "\n") -> str:
        """Flatten to plain text (drops images). Used by text-only encoders."""
        return sep.join(s.text for s in self.segments if s.has_text).strip()

    @property
    def images(self) -> List[Any]:
        return [s.image for s in self.segments if s.has_image]

    @property
    def is_multimodal(self) -> bool:
        return any(s.has_image for s in self.segments)


class RetrievalTask(ABC):
    """One evaluatable unit (e.g. a single SSRB schema, or a MERIT split).

    Concrete loaders subclass this and implement the three ``load_*`` methods.
    A benchmark typically expands into many ``RetrievalTask`` instances (one per
    domain/schema); see :mod:`mcr.data.registry`.
    """

    #: short identifier, e.g. "ssrb/Academic/academic_collabo"
    name: str = "task"
    #: "text" | "image_text" | "interleaved"
    modality: str = "text"
    #: metric keys to highlight in the report so numbers line up with the paper
    official_metrics: Tuple[str, ...] = ("ndcg@10", "recall@10")

    @abstractmethod
    def load_corpus(self) -> Iterator[Tuple[str, Content]]:
        """Yield ``(doc_id, Content)`` for every document in the corpus."""

    @abstractmethod
    def load_queries(self) -> Iterator[Tuple[str, Content]]:
        """Yield ``(query_id, Content)`` for every query."""

    @abstractmethod
    def load_qrels(self) -> QrelsType:
        """Return relevance judgements as ``{qid: {doc_id: rel}}``."""

    def stratify_key(self, qid: str) -> Dict[str, Union[int, str]]:
        """Per-query difficulty descriptor for stratified reporting.

        Default: no stratification. Multi-condition tasks override this to expose
        e.g. ``{"n_cond": 3}`` or ``{"nf": 2, "ns": 1}``.
        """
        return {}