Spaces:
Build error
Build error
File size: 7,937 Bytes
0bd1b0f 37f0321 d02622b 37f0321 d02622b 8c66169 f5f93a3 37f0321 0b9d8c7 37f0321 0b9d8c7 37f0321 0b9d8c7 37f0321 ad9c5d9 37f0321 9e67c3b 37f0321 0bd1b0f 37f0321 ad9c5d9 37f0321 f5f93a3 37f0321 0bd1b0f 37f0321 ad9c5d9 37f0321 9e67c3b 37f0321 6978916 37f0321 ad9c5d9 f5f93a3 ad9c5d9 9e67c3b 37f0321 f5f93a3 37f0321 ad9c5d9 37f0321 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 | import json
import re
from collections import Counter, defaultdict
from typing import Dict, List, Tuple
from graphgen.bases import BaseGraphStorage, BaseKGBuilder, BaseLLMWrapper, Chunk
from graphgen.templates import KG_EXTRACTION_PROMPT, KG_SUMMARIZATION_PROMPT
from graphgen.utils import (
detect_main_language,
handle_single_entity_extraction,
handle_single_relationship_extraction,
logger,
pack_history_conversations,
split_string_by_multi_markers,
)
class LightRAGKGBuilder(BaseKGBuilder):
def __init__(self, llm_client: BaseLLMWrapper, max_loop: int = 3):
super().__init__(llm_client)
self.max_loop = max_loop
self.tokenizer = llm_client.tokenizer
async def extract(
self, chunk: Chunk
) -> Tuple[Dict[str, List[dict]], Dict[Tuple[str, str], List[dict]]]:
"""
Extract entities and relationships from a single chunk using the LLM client.
:param chunk
:return: (nodes_data, edges_data)
"""
chunk_id = chunk.id
content = chunk.content
# step 1: language_detection
language = detect_main_language(content)
hint_prompt = KG_EXTRACTION_PROMPT[language]["TEMPLATE"].format(
**KG_EXTRACTION_PROMPT["FORMAT"], input_text=content
)
# step 2: initial glean
final_result = await self.llm_client.generate_answer(hint_prompt)
logger.debug("First extraction result: %s", final_result)
# step3: iterative refinement
history = pack_history_conversations(hint_prompt, final_result)
for loop_idx in range(self.max_loop):
if_loop_result = await self.llm_client.generate_answer(
text=KG_EXTRACTION_PROMPT[language]["IF_LOOP"], history=history
)
if_loop_result = if_loop_result.strip().strip('"').strip("'").lower()
if if_loop_result != "yes":
break
glean_result = await self.llm_client.generate_answer(
text=KG_EXTRACTION_PROMPT[language]["CONTINUE"], history=history
)
logger.debug("Loop %s glean: %s", loop_idx + 1, glean_result)
history += pack_history_conversations(
KG_EXTRACTION_PROMPT[language]["CONTINUE"], glean_result
)
final_result += glean_result
# step 4: parse the final result
records = split_string_by_multi_markers(
final_result,
[
KG_EXTRACTION_PROMPT["FORMAT"]["record_delimiter"],
KG_EXTRACTION_PROMPT["FORMAT"]["completion_delimiter"],
],
)
nodes = defaultdict(list)
edges = defaultdict(list)
for record in records:
match = re.search(r"\((.*)\)", record)
if not match:
continue
inner = match.group(1)
attributes = split_string_by_multi_markers(
inner, [KG_EXTRACTION_PROMPT["FORMAT"]["tuple_delimiter"]]
)
entity = await handle_single_entity_extraction(attributes, chunk_id)
if entity is not None:
nodes[entity["entity_name"]].append(entity)
continue
relation = await handle_single_relationship_extraction(attributes, chunk_id)
if relation is not None:
key = (relation["src_id"], relation["tgt_id"])
edges[key].append(relation)
return dict(nodes), dict(edges)
async def merge_nodes(
self,
node_data: tuple[str, List[dict]],
kg_instance: BaseGraphStorage,
) -> dict:
entity_name, node_data = node_data
entity_types = []
source_ids = []
descriptions = []
node = kg_instance.get_node(entity_name)
if node is not None:
entity_types.append(node["entity_type"])
source_ids.extend(
split_string_by_multi_markers(node["source_id"], ["<SEP>"])
)
descriptions.append(node["description"])
# take the most frequent entity_type
entity_type = sorted(
Counter([dp["entity_type"] for dp in node_data] + entity_types).items(),
key=lambda x: x[1],
reverse=True,
)[0][0]
description = "<SEP>".join(
sorted(set([dp["description"] for dp in node_data] + descriptions))
)
description = await self._handle_kg_summary(entity_name, description)
source_id = "<SEP>".join(
set([dp["source_id"] for dp in node_data] + source_ids)
)
node_data_dict = {
"entity_type": entity_type,
"entity_name": entity_name,
"description": description,
"source_id": source_id,
"length": self.tokenizer.count_tokens(description),
}
if entity_type in ("IMAGE", "TABLE", "FORMULA"):
metadata = next(
(dp["metadata"] for dp in node_data if dp.get("metadata")), None
)
if metadata:
node_data_dict["metadata"] = json.dumps(
metadata, ensure_ascii=False, default=str
)
kg_instance.upsert_node(entity_name, node_data=node_data_dict)
return node_data_dict
async def merge_edges(
self,
edges_data: tuple[Tuple[str, str], List[dict]],
kg_instance: BaseGraphStorage,
) -> dict:
(src_id, tgt_id), edge_data = edges_data
source_ids = []
descriptions = []
edge = kg_instance.get_edge(src_id, tgt_id)
if edge is not None:
source_ids.extend(
split_string_by_multi_markers(edge["source_id"], ["<SEP>"])
)
descriptions.append(edge["description"])
description = "<SEP>".join(
sorted(set([dp["description"] for dp in edge_data] + descriptions))
)
source_id = "<SEP>".join(
set([dp["source_id"] for dp in edge_data] + source_ids)
)
if not kg_instance.has_node(src_id) or not kg_instance.has_node(tgt_id):
logger.warning("Edge (%s, %s) has missing nodes.", src_id, tgt_id)
return {}
description = await self._handle_kg_summary(
f"({src_id}, {tgt_id})", description
)
edge_data = {
"src_id": src_id,
"tgt_id": tgt_id,
"description": description,
"source_id": source_id, # for traceability
"length": self.tokenizer.count_tokens(description),
}
kg_instance.upsert_edge(
src_id,
tgt_id,
edge_data=edge_data,
)
return edge_data
async def _handle_kg_summary(
self,
entity_or_relation_name: str,
description: str,
max_summary_tokens: int = 200,
) -> str:
"""
Handle knowledge graph summary
:param entity_or_relation_name
:param description
:param max_summary_tokens
:return summary
"""
tokenizer_instance = self.llm_client.tokenizer
language = detect_main_language(description)
tokens = tokenizer_instance.encode(description)
if len(tokens) < max_summary_tokens:
return description
use_description = tokenizer_instance.decode(tokens[:max_summary_tokens])
prompt = KG_SUMMARIZATION_PROMPT[language]["TEMPLATE"].format(
entity_name=entity_or_relation_name,
description_list=use_description.split("<SEP>"),
**KG_SUMMARIZATION_PROMPT["FORMAT"],
)
new_description = await self.llm_client.generate_answer(prompt)
logger.info(
"Entity or relation %s summary: %s",
entity_or_relation_name,
new_description,
)
return new_description
|