ragflow/graphrag/entity_resolution.py
Yongteng Lei b705ff08fe
Refa: improve GraphRAG similarity sensitivity to numeric differences (#8479)
### What problem does this PR solve?

Improve GraphRAG similarity sensitivity to numeric differences. #8444.

### Type of change

- [x] Refactoring
2025-06-25 16:20:59 +08:00

244 lines
11 KiB
Python

#
# Copyright 2024 The InfiniFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
import logging
import itertools
import re
from dataclasses import dataclass
from typing import Any, Callable
import networkx as nx
import trio
from graphrag.general.extractor import Extractor
from rag.nlp import is_english
import editdistance
from graphrag.entity_resolution_prompt import ENTITY_RESOLUTION_PROMPT
from rag.llm.chat_model import Base as CompletionLLM
from graphrag.utils import perform_variable_replacements, chat_limiter, GraphChange
DEFAULT_RECORD_DELIMITER = "##"
DEFAULT_ENTITY_INDEX_DELIMITER = "<|>"
DEFAULT_RESOLUTION_RESULT_DELIMITER = "&&"
@dataclass
class EntityResolutionResult:
"""Entity resolution result class definition."""
graph: nx.Graph
change: GraphChange
class EntityResolution(Extractor):
"""Entity resolution class definition."""
_resolution_prompt: str
_output_formatter_prompt: str
_record_delimiter_key: str
_entity_index_delimiter_key: str
_resolution_result_delimiter_key: str
def __init__(
self,
llm_invoker: CompletionLLM,
):
super().__init__(llm_invoker)
"""Init method definition."""
self._llm = llm_invoker
self._resolution_prompt = ENTITY_RESOLUTION_PROMPT
self._record_delimiter_key = "record_delimiter"
self._entity_index_dilimiter_key = "entity_index_delimiter"
self._resolution_result_delimiter_key = "resolution_result_delimiter"
self._input_text_key = "input_text"
async def __call__(self, graph: nx.Graph,
subgraph_nodes: set[str],
prompt_variables: dict[str, Any] | None = None,
callback: Callable | None = None) -> EntityResolutionResult:
"""Call method definition."""
if prompt_variables is None:
prompt_variables = {}
# Wire defaults into the prompt variables
self.prompt_variables = {
**prompt_variables,
self._record_delimiter_key: prompt_variables.get(self._record_delimiter_key)
or DEFAULT_RECORD_DELIMITER,
self._entity_index_dilimiter_key: prompt_variables.get(self._entity_index_dilimiter_key)
or DEFAULT_ENTITY_INDEX_DELIMITER,
self._resolution_result_delimiter_key: prompt_variables.get(self._resolution_result_delimiter_key)
or DEFAULT_RESOLUTION_RESULT_DELIMITER,
}
nodes = sorted(graph.nodes())
entity_types = sorted(set(graph.nodes[node].get('entity_type', '-') for node in nodes))
node_clusters = {entity_type: [] for entity_type in entity_types}
for node in nodes:
node_clusters[graph.nodes[node].get('entity_type', '-')].append(node)
candidate_resolution = {entity_type: [] for entity_type in entity_types}
for k, v in node_clusters.items():
candidate_resolution[k] = [(a, b) for a, b in itertools.combinations(v, 2) if (a in subgraph_nodes or b in subgraph_nodes) and self.is_similarity(a, b)]
num_candidates = sum([len(candidates) for _, candidates in candidate_resolution.items()])
callback(msg=f"Identified {num_candidates} candidate pairs")
remain_candidates_to_resolve = num_candidates
resolution_result = set()
resolution_result_lock = trio.Lock()
resolution_batch_size = 100
max_concurrent_tasks = 5
semaphore = trio.Semaphore(max_concurrent_tasks)
async def limited_resolve_candidate(candidate_batch, result_set, result_lock):
nonlocal remain_candidates_to_resolve, callback
async with semaphore:
try:
with trio.move_on_after(180) as cancel_scope:
await self._resolve_candidate(candidate_batch, result_set, result_lock)
remain_candidates_to_resolve = remain_candidates_to_resolve - len(candidate_batch[1])
callback(msg=f"Resolved {len(candidate_batch[1])} pairs, {remain_candidates_to_resolve} are remained to resolve. ")
if cancel_scope.cancelled_caught:
logging.warning(f"Timeout resolving {candidate_batch}, skipping...")
remain_candidates_to_resolve = remain_candidates_to_resolve - len(candidate_batch[1])
callback(msg=f"Fail to resolved {len(candidate_batch[1])} pairs due to timeout reason, skipped. {remain_candidates_to_resolve} are remained to resolve. ")
except Exception as e:
logging.error(f"Error resolving candidate batch: {e}")
async with trio.open_nursery() as nursery:
for candidate_resolution_i in candidate_resolution.items():
if not candidate_resolution_i[1]:
continue
for i in range(0, len(candidate_resolution_i[1]), resolution_batch_size):
candidate_batch = candidate_resolution_i[0], candidate_resolution_i[1][i:i + resolution_batch_size]
nursery.start_soon(limited_resolve_candidate, candidate_batch, resolution_result, resolution_result_lock)
callback(msg=f"Resolved {num_candidates} candidate pairs, {len(resolution_result)} of them are selected to merge.")
change = GraphChange()
connect_graph = nx.Graph()
connect_graph.add_edges_from(resolution_result)
async def limited_merge_nodes(graph, nodes, change):
async with semaphore:
await self._merge_graph_nodes(graph, nodes, change)
async with trio.open_nursery() as nursery:
for sub_connect_graph in nx.connected_components(connect_graph):
merging_nodes = list(sub_connect_graph)
nursery.start_soon(limited_merge_nodes, graph, merging_nodes, change)
# Update pagerank
pr = nx.pagerank(graph)
for node_name, pagerank in pr.items():
graph.nodes[node_name]["pagerank"] = pagerank
return EntityResolutionResult(
graph=graph,
change=change,
)
async def _resolve_candidate(self, candidate_resolution_i: tuple[str, list[tuple[str, str]]], resolution_result: set[str], resolution_result_lock: trio.Lock):
gen_conf = {"temperature": 0.5}
pair_txt = [
f'When determining whether two {candidate_resolution_i[0]}s are the same, you should only focus on critical properties and overlook noisy factors.\n']
for index, candidate in enumerate(candidate_resolution_i[1]):
pair_txt.append(
f'Question {index + 1}: name of{candidate_resolution_i[0]} A is {candidate[0]} ,name of{candidate_resolution_i[0]} B is {candidate[1]}')
sent = 'question above' if len(pair_txt) == 1 else f'above {len(pair_txt)} questions'
pair_txt.append(
f'\nUse domain knowledge of {candidate_resolution_i[0]}s to help understand the text and answer the {sent} in the format: For Question i, Yes, {candidate_resolution_i[0]} A and {candidate_resolution_i[0]} B are the same {candidate_resolution_i[0]}./No, {candidate_resolution_i[0]} A and {candidate_resolution_i[0]} B are different {candidate_resolution_i[0]}s. For Question i+1, (repeat the above procedures)')
pair_prompt = '\n'.join(pair_txt)
variables = {
**self.prompt_variables,
self._input_text_key: pair_prompt
}
text = perform_variable_replacements(self._resolution_prompt, variables=variables)
logging.info(f"Created resolution prompt {len(text)} bytes for {len(candidate_resolution_i[1])} entity pairs of type {candidate_resolution_i[0]}")
async with chat_limiter:
try:
with trio.move_on_after(120) as cancel_scope:
response = await trio.to_thread.run_sync(self._chat, text, [{"role": "user", "content": "Output:"}], gen_conf)
if cancel_scope.cancelled_caught:
logging.warning("_resolve_candidate._chat timeout, skipping...")
return
except Exception as e:
logging.error(f"_resolve_candidate._chat failed: {e}")
return
logging.debug(f"_resolve_candidate chat prompt: {text}\nchat response: {response}")
result = self._process_results(len(candidate_resolution_i[1]), response,
self.prompt_variables.get(self._record_delimiter_key,
DEFAULT_RECORD_DELIMITER),
self.prompt_variables.get(self._entity_index_dilimiter_key,
DEFAULT_ENTITY_INDEX_DELIMITER),
self.prompt_variables.get(self._resolution_result_delimiter_key,
DEFAULT_RESOLUTION_RESULT_DELIMITER))
async with resolution_result_lock:
for result_i in result:
resolution_result.add(candidate_resolution_i[1][result_i[0] - 1])
def _process_results(
self,
records_length: int,
results: str,
record_delimiter: str,
entity_index_delimiter: str,
resolution_result_delimiter: str
) -> list:
ans_list = []
records = [r.strip() for r in results.split(record_delimiter)]
for record in records:
pattern_int = f"{re.escape(entity_index_delimiter)}(\d+){re.escape(entity_index_delimiter)}"
match_int = re.search(pattern_int, record)
res_int = int(str(match_int.group(1) if match_int else '0'))
if res_int > records_length:
continue
pattern_bool = f"{re.escape(resolution_result_delimiter)}([a-zA-Z]+){re.escape(resolution_result_delimiter)}"
match_bool = re.search(pattern_bool, record)
res_bool = str(match_bool.group(1) if match_bool else '')
if res_int and res_bool:
if res_bool.lower() == 'yes':
ans_list.append((res_int, "yes"))
return ans_list
def _has_digit_in_2gram_diff(self, a, b):
def to_2gram_set(s):
return {s[i:i+2] for i in range(len(s) - 1)}
set_a = to_2gram_set(a)
set_b = to_2gram_set(b)
diff = set_a ^ set_b
return any(any(c.isdigit() for c in pair) for pair in diff)
def is_similarity(self, a, b):
if self._has_digit_in_2gram_diff(a, b):
return False
if is_english(a) and is_english(b):
if editdistance.eval(a, b) <= min(len(a), len(b)) // 2:
return True
return False
if len(set(a) & set(b)) > 1:
return True
return False