Spaces:
Sleeping
Sleeping
| """Load extraction-stage table JSON using existing WaterScope media resolver. | |
| Not a validated cell grid when structure_extracted=False. Preserve PDF word | |
| coordinates, original strings and extraction quality flags; never silently | |
| interpret position-only data as exact, verified spreadsheet cells. | |
| """ | |
| from __future__ import annotations | |
| import json | |
| from pathlib import Path | |
| from typing import Callable, Any | |
| def load_table_evidence(folder: str, label: str, resolver: Callable[[str,str], Any]) -> dict | None: | |
| manifest_path = resolver(folder, "manifest.json") | |
| if manifest_path is None: | |
| return None | |
| manifest = json.loads(Path(manifest_path).read_text(encoding="utf-8-sig")) | |
| matches = [x for x in manifest.get("items", []) if x.get("label") == label and x.get("kind") == "table"] | |
| if len(matches) != 1: | |
| return None | |
| filename = matches[0].get("data_file") | |
| if not filename or Path(filename).name != filename or not filename.lower().endswith(".json"): | |
| return None | |
| path = resolver(folder, filename) | |
| if path is None: | |
| return None | |
| raw = Path(path).read_bytes() | |
| if len(raw) > 3_000_000: | |
| raise ValueError("Table JSON exceeds 3MB limit") | |
| table = json.loads(raw.decode("utf-8-sig")) | |
| if table.get("kind") != "table" or table.get("label") != label: | |
| raise ValueError("Table JSON does not match manifest label") | |
| rows = table.get("text_rows") or [] | |
| cells = table.get("cells") | |
| if not isinstance(rows, list) or len(rows) > 500: | |
| raise ValueError("Invalid/oversized table text rows") | |
| # Preserve every source row, but replace redundant per-word bounding boxes | |
| # with a compact word@horizontal-position representation. A text row is NOT | |
| # a verified table cell row, especially for merged/multi-line groups. | |
| compact_rows = [] | |
| for row in rows: | |
| words = [] | |
| for word in (row.get("words") or [])[:100]: | |
| bbox = word.get("bbox_points") or [] | |
| token = str(word.get("text", ""))[:120] | |
| if len(bbox) == 4: | |
| words.append([token, round(float(bbox[0]), 1)]) | |
| else: | |
| words.append([token, None]) | |
| compact_rows.append({"text": str(row.get("text", ""))[:1600], "positioned_words": words}) | |
| # Include all nonempty rows, including continuation rows and classifications. | |
| # The original JSON remains in the dataset; this is only an inference view. | |
| return { | |
| "label": label, "data_file": filename, | |
| "structure_extracted": bool(table.get("structure_extracted")), | |
| "row_count": table.get("row_count"), "column_count": table.get("column_count"), | |
| "cells": cells if table.get("structure_extracted") and isinstance(cells, list) else None, | |
| "text_rows": compact_rows, | |
| "source_quality": "structured_cells" if table.get("structure_extracted") and isinstance(cells,list) | |
| else "positioned_text_unverified", | |
| } | |