Upload iath_decoder.py with huggingface_hub
Browse files- iath_decoder.py +188 -0
iath_decoder.py
ADDED
|
@@ -0,0 +1,188 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import struct
|
| 2 |
+
import zstandard as zstd
|
| 3 |
+
import json
|
| 4 |
+
from typing import Dict # これを追加
|
| 5 |
+
# これを追加
|
| 6 |
+
from datetime import datetime
|
| 7 |
+
|
| 8 |
+
class IathDecoder:
|
| 9 |
+
"""
|
| 10 |
+
.iath互換の圧縮バイナリデータをKnowledge Tileオブジェクトにデコードします。
|
| 11 |
+
"""
|
| 12 |
+
|
| 13 |
+
def _decode_string_from_buffer(self, buffer, offset):
|
| 14 |
+
"""バッファからNULL終端文字列をデコードします。"""
|
| 15 |
+
end_offset = buffer.find(b'\0', offset)
|
| 16 |
+
if end_offset == -1:
|
| 17 |
+
raise ValueError("Invalid string format in buffer")
|
| 18 |
+
s = buffer[offset:end_offset].decode('utf-8')
|
| 19 |
+
return s, end_offset + 1
|
| 20 |
+
|
| 21 |
+
def _decode_metadata(self, buffer: bytes) -> dict:
|
| 22 |
+
"""メタデータセクションをデコードします。"""
|
| 23 |
+
offset = 0
|
| 24 |
+
kid, offset = self._decode_string_from_buffer(buffer, offset)
|
| 25 |
+
topic, offset = self._decode_string_from_buffer(buffer, offset)
|
| 26 |
+
created_at = buffer[offset:offset+27].decode('ascii').rstrip('\0')
|
| 27 |
+
|
| 28 |
+
return {"knowledge_id": kid, "topic": topic, "created_at": created_at}
|
| 29 |
+
|
| 30 |
+
def _decode_coordinates(self, buffer: bytes) -> dict:
|
| 31 |
+
"""座標セクションをデコードします。"""
|
| 32 |
+
coords = struct.unpack("<ffffff", buffer)
|
| 33 |
+
return {
|
| 34 |
+
"medical_space": (coords[0], coords[1], coords[2]),
|
| 35 |
+
"meta_space": (coords[3], coords[4], coords[5])
|
| 36 |
+
}
|
| 37 |
+
|
| 38 |
+
def _decode_content(self, buffer: bytes) -> dict:
|
| 39 |
+
"""コンテンツセクションをデコードします。"""
|
| 40 |
+
offset = 0
|
| 41 |
+
|
| 42 |
+
# thinking_process
|
| 43 |
+
think_len = struct.unpack("<I", buffer[offset:offset+4])[0]
|
| 44 |
+
offset += 4
|
| 45 |
+
thinking = buffer[offset:offset+think_len].decode('utf-8')
|
| 46 |
+
offset += think_len
|
| 47 |
+
|
| 48 |
+
# final_response
|
| 49 |
+
resp_len = struct.unpack("<I", buffer[offset:offset+4])[0]
|
| 50 |
+
offset += 4
|
| 51 |
+
response = buffer[offset:offset+resp_len].decode('utf-8')
|
| 52 |
+
|
| 53 |
+
return {"thinking_process": thinking, "final_response": response}
|
| 54 |
+
|
| 55 |
+
def _decode_verification(self, buffer: bytes) -> dict:
|
| 56 |
+
"""検証履歴セクションをデコードします。"""
|
| 57 |
+
status_map = {
|
| 58 |
+
0: "pending_review", 1: "partial_verified",
|
| 59 |
+
2: "verified", 3: "expert_confirmed"
|
| 60 |
+
}
|
| 61 |
+
|
| 62 |
+
status_code, initial_certainty, reviewer_count = struct.unpack("<BBI", buffer[:6])
|
| 63 |
+
status = status_map.get(status_code, "unknown")
|
| 64 |
+
|
| 65 |
+
# NOTE: レビュアーIDのデコードはエンコーダーに合わせて省略
|
| 66 |
+
|
| 67 |
+
return {
|
| 68 |
+
"status": status,
|
| 69 |
+
"initial_certainty": initial_certainty,
|
| 70 |
+
"reviewers": [] # ダミー
|
| 71 |
+
}
|
| 72 |
+
|
| 73 |
+
def decode_tile(self, compressed_binary: bytes) -> dict:
|
| 74 |
+
"""
|
| 75 |
+
単一の圧縮タイルデータをデコードしてKnowledge Tileオブジェクトを復元します。
|
| 76 |
+
|
| 77 |
+
Args:
|
| 78 |
+
compressed_binary (bytes): 圧縮されたバイナリデータ。
|
| 79 |
+
|
| 80 |
+
Returns:
|
| 81 |
+
dict: 復元されたKnowledge Tileオブジェクト。
|
| 82 |
+
"""
|
| 83 |
+
try:
|
| 84 |
+
dctx = zstd.ZstdDecompressor()
|
| 85 |
+
uncompressed = dctx.decompress(compressed_binary)
|
| 86 |
+
except zstd.ZstdError as e:
|
| 87 |
+
raise ValueError(f"Zstandard decompression failed: {e}")
|
| 88 |
+
|
| 89 |
+
offset = 0
|
| 90 |
+
decoded_sections = {}
|
| 91 |
+
|
| 92 |
+
try:
|
| 93 |
+
# Metadata
|
| 94 |
+
md_len = struct.unpack("<I", uncompressed[offset:offset+4])[0]
|
| 95 |
+
offset += 4
|
| 96 |
+
decoded_sections["metadata"] = self._decode_metadata(uncompressed[offset:offset+md_len])
|
| 97 |
+
offset += md_len
|
| 98 |
+
|
| 99 |
+
# Coordinates
|
| 100 |
+
coord_len = struct.unpack("<I", uncompressed[offset:offset+4])[0]
|
| 101 |
+
offset += 4
|
| 102 |
+
decoded_sections["coordinates"] = self._decode_coordinates(uncompressed[offset:offset+coord_len])
|
| 103 |
+
offset += coord_len
|
| 104 |
+
|
| 105 |
+
# Content
|
| 106 |
+
content_len = struct.unpack("<I", uncompressed[offset:offset+4])[0]
|
| 107 |
+
offset += 4
|
| 108 |
+
decoded_sections["content"] = self._decode_content(uncompressed[offset:offset+content_len])
|
| 109 |
+
offset += content_len
|
| 110 |
+
|
| 111 |
+
# Verification
|
| 112 |
+
verif_len = struct.unpack("<I", uncompressed[offset:offset+4])[0]
|
| 113 |
+
offset += 4
|
| 114 |
+
decoded_sections["verification"] = self._decode_verification(uncompressed[offset:offset+verif_len])
|
| 115 |
+
offset += verif_len
|
| 116 |
+
|
| 117 |
+
except (struct.error, IndexError, UnicodeDecodeError) as e:
|
| 118 |
+
raise ValueError(f"Failed to parse tile structure at offset {offset}: {e}")
|
| 119 |
+
|
| 120 |
+
# スキーマに準拠するよう、デコードしたセクションを再構成
|
| 121 |
+
restored_tile = {
|
| 122 |
+
"metadata": decoded_sections.get("metadata"),
|
| 123 |
+
"content": decoded_sections.get("content"),
|
| 124 |
+
"coordinates": decoded_sections.get("coordinates", {}),
|
| 125 |
+
"verification": decoded_sections.get("verification"),
|
| 126 |
+
# 以下はエンコードしていないためデフォルト値
|
| 127 |
+
"source": {},
|
| 128 |
+
"history": []
|
| 129 |
+
}
|
| 130 |
+
# 不足しているキーを補完
|
| 131 |
+
if "coordinates" in restored_tile:
|
| 132 |
+
restored_tile["coordinates"].setdefault("reasoning_path", [])
|
| 133 |
+
|
| 134 |
+
return restored_tile
|
| 135 |
+
|
| 136 |
+
def decode_batch(self, full_db_content: bytes) -> Dict[str, Dict]:
|
| 137 |
+
"""
|
| 138 |
+
ヘッダー、インデックス、データセクションを含む完全な.iath DBファイルをデコードします。
|
| 139 |
+
|
| 140 |
+
Args:
|
| 141 |
+
full_db_content (bytes): .iathファイル全体のバイナリコンテンツ。
|
| 142 |
+
|
| 143 |
+
Returns:
|
| 144 |
+
Dict[str, Dict]: tile_idをキーとする、デコードされた知識タイルの辞書。
|
| 145 |
+
"""
|
| 146 |
+
print("--- .iathデータベースのバッチデコード開始 ---")
|
| 147 |
+
# 1. ヘッダーをパース
|
| 148 |
+
if len(full_db_content) < 64:
|
| 149 |
+
raise ValueError("Invalid .iath file: Header is too short.")
|
| 150 |
+
|
| 151 |
+
magic, version, domain_code, compression_type, checksum, index_offset, data_offset = \
|
| 152 |
+
struct.unpack("<4sIBB32sQQ6x", full_db_content[:64])
|
| 153 |
+
|
| 154 |
+
if magic != b'ILMA':
|
| 155 |
+
raise ValueError("Invalid .iath file: Magic number is incorrect.")
|
| 156 |
+
|
| 157 |
+
print(f" - Header OK: Version={version}, Domain={domain_code}, Index Offset={index_offset}, Data Offset={data_offset}")
|
| 158 |
+
|
| 159 |
+
# 2. インデックスセクションを読み込み
|
| 160 |
+
# データオフセットの開始位置までがインデックスセクション
|
| 161 |
+
index_data_binary = full_db_content[index_offset:data_offset]
|
| 162 |
+
index = json.loads(index_data_binary.decode('utf-8'))
|
| 163 |
+
print(f" - インデックス読み込み完了: {len(index)}件")
|
| 164 |
+
|
| 165 |
+
# 3. データセクションから各タイルをデコード
|
| 166 |
+
all_tiles = {}
|
| 167 |
+
for item in index:
|
| 168 |
+
tile_id, offset, length = item['id'], item['offset'], item['length']
|
| 169 |
+
|
| 170 |
+
# データセクション内でのタイルの範囲を特定
|
| 171 |
+
start = data_offset + offset
|
| 172 |
+
end = start + length
|
| 173 |
+
tile_compressed_data = full_db_content[start:end]
|
| 174 |
+
|
| 175 |
+
# 個別のタイルをデコード
|
| 176 |
+
try:
|
| 177 |
+
decoded_tile = self.decode_tile(tile_compressed_data)
|
| 178 |
+
# デコード結果にIDを付与(JSONにはIDがないため)
|
| 179 |
+
if "metadata" in decoded_tile and "knowledge_id" not in decoded_tile["metadata"]:
|
| 180 |
+
decoded_tile["metadata"]["knowledge_id"] = tile_id
|
| 181 |
+
all_tiles[tile_id] = decoded_tile
|
| 182 |
+
except Exception as e:
|
| 183 |
+
print(f"警告: タイルID {tile_id} のデコードに失敗しました。スキップします。エラー: {e}")
|
| 184 |
+
|
| 185 |
+
print(f" - 全タイルのデコード完了: {len(all_tiles)}件")
|
| 186 |
+
print("--- バッチデコード完了 ---")
|
| 187 |
+
return all_tiles
|
| 188 |
+
|