mirror of
https://github.com/infiniflow/ragflow.git
synced 2025-12-08 20:42:30 +08:00
Optimize graphrag again (#6513)
### What problem does this PR solve? Removed set_entity and set_relation to avoid accessing doc engine during graph computation. Introduced GraphChange to avoid writing unchanged chunks. ### Type of change - [x] Performance Improvement
This commit is contained in:
Zhichang Yu
@ -15,11 +15,11 @@
|
||||
#
|
||||
import json
|
||||
import logging
|
||||
from functools import partial
|
||||
import networkx as nx
|
||||
import trio
|
||||
|
||||
from api import settings
|
||||
from api.utils import get_uuid
|
||||
from graphrag.light.graph_extractor import GraphExtractor as LightKGExt
|
||||
from graphrag.general.graph_extractor import GraphExtractor as GeneralKGExt
|
||||
from graphrag.general.community_reports_extractor import CommunityReportsExtractor
|
||||
@ -27,32 +27,15 @@ from graphrag.entity_resolution import EntityResolution
|
||||
from graphrag.general.extractor import Extractor
|
||||
from graphrag.utils import (
|
||||
graph_merge,
|
||||
set_entity,
|
||||
get_relation,
|
||||
set_relation,
|
||||
get_entity,
|
||||
get_graph,
|
||||
set_graph,
|
||||
chunk_id,
|
||||
update_nodes_pagerank_nhop_neighbour,
|
||||
does_graph_contains,
|
||||
get_graph_doc_ids,
|
||||
tidy_graph,
|
||||
GraphChange,
|
||||
)
|
||||
from rag.nlp import rag_tokenizer, search
|
||||
from rag.utils.redis_conn import REDIS_CONN
|
||||
|
||||
|
||||
def graphrag_task_set(tenant_id, kb_id, doc_id) -> bool:
|
||||
key = f"graphrag:{tenant_id}:{kb_id}"
|
||||
ok = REDIS_CONN.set(key, doc_id, exp=3600 * 24)
|
||||
if not ok:
|
||||
raise Exception(f"Faild to set the {key} to {doc_id}")
|
||||
|
||||
|
||||
def graphrag_task_get(tenant_id, kb_id) -> str | None:
|
||||
key = f"graphrag:{tenant_id}:{kb_id}"
|
||||
doc_id = REDIS_CONN.get(key)
|
||||
return doc_id
|
||||
from rag.utils.redis_conn import RedisDistributedLock
|
||||
|
||||
|
||||
async def run_graphrag(
|
||||
@ -72,7 +55,7 @@ async def run_graphrag(
|
||||
):
|
||||
chunks.append(d["content_with_weight"])
|
||||
|
||||
graph, doc_ids = await update_graph(
|
||||
subgraph = await generate_subgraph(
|
||||
LightKGExt
|
||||
if row["parser_config"]["graphrag"]["method"] != "general"
|
||||
else GeneralKGExt,
|
||||
@ -86,14 +69,26 @@ async def run_graphrag(
|
||||
embedding_model,
|
||||
callback,
|
||||
)
|
||||
if not graph:
|
||||
new_graph = None
|
||||
if subgraph:
|
||||
new_graph = await merge_subgraph(
|
||||
tenant_id,
|
||||
kb_id,
|
||||
doc_id,
|
||||
subgraph,
|
||||
embedding_model,
|
||||
callback,
|
||||
)
|
||||
|
||||
if not with_resolution or not with_community:
|
||||
return
|
||||
if with_resolution or with_community:
|
||||
graphrag_task_set(tenant_id, kb_id, doc_id)
|
||||
if with_resolution:
|
||||
|
||||
if new_graph is None:
|
||||
new_graph = await get_graph(tenant_id, kb_id)
|
||||
|
||||
if with_resolution and new_graph is not None:
|
||||
await resolve_entities(
|
||||
graph,
|
||||
doc_ids,
|
||||
new_graph,
|
||||
tenant_id,
|
||||
kb_id,
|
||||
doc_id,
|
||||
@ -101,10 +96,9 @@ async def run_graphrag(
|
||||
embedding_model,
|
||||
callback,
|
||||
)
|
||||
if with_community:
|
||||
if with_community and new_graph is not None:
|
||||
await extract_community(
|
||||
graph,
|
||||
doc_ids,
|
||||
new_graph,
|
||||
tenant_id,
|
||||
kb_id,
|
||||
doc_id,
|
||||
@ -117,7 +111,7 @@ async def run_graphrag(
|
||||
return
|
||||
|
||||
|
||||
async def update_graph(
|
||||
async def generate_subgraph(
|
||||
extractor: Extractor,
|
||||
tenant_id: str,
|
||||
kb_id: str,
|
||||
@ -131,34 +125,41 @@ async def update_graph(
|
||||
):
|
||||
contains = await does_graph_contains(tenant_id, kb_id, doc_id)
|
||||
if contains:
|
||||
callback(msg=f"Graph already contains {doc_id}, cancel myself")
|
||||
return None, None
|
||||
callback(msg=f"Graph already contains {doc_id}")
|
||||
return None
|
||||
start = trio.current_time()
|
||||
ext = extractor(
|
||||
llm_bdl,
|
||||
language=language,
|
||||
entity_types=entity_types,
|
||||
get_entity=partial(get_entity, tenant_id, kb_id),
|
||||
set_entity=partial(set_entity, tenant_id, kb_id, embed_bdl),
|
||||
get_relation=partial(get_relation, tenant_id, kb_id),
|
||||
set_relation=partial(set_relation, tenant_id, kb_id, embed_bdl),
|
||||
)
|
||||
ents, rels = await ext(doc_id, chunks, callback)
|
||||
subgraph = nx.Graph()
|
||||
for en in ents:
|
||||
subgraph.add_node(en["entity_name"], entity_type=en["entity_type"])
|
||||
for ent in ents:
|
||||
assert "description" in ent, f"entity {ent} does not have description"
|
||||
ent["source_id"] = [doc_id]
|
||||
subgraph.add_node(ent["entity_name"], **ent)
|
||||
|
||||
ignored_rels = 0
|
||||
for rel in rels:
|
||||
assert "description" in rel, f"relation {rel} does not have description"
|
||||
if not subgraph.has_node(rel["src_id"]) or not subgraph.has_node(rel["tgt_id"]):
|
||||
ignored_rels += 1
|
||||
continue
|
||||
rel["source_id"] = [doc_id]
|
||||
subgraph.add_edge(
|
||||
rel["src_id"],
|
||||
rel["tgt_id"],
|
||||
weight=rel["weight"],
|
||||
# description=rel["description"]
|
||||
**rel,
|
||||
)
|
||||
# TODO: infinity doesn't support array search
|
||||
if ignored_rels:
|
||||
callback(msg=f"ignored {ignored_rels} relations due to missing entities.")
|
||||
tidy_graph(subgraph, callback)
|
||||
|
||||
subgraph.graph["source_id"] = [doc_id]
|
||||
chunk = {
|
||||
"content_with_weight": json.dumps(
|
||||
nx.node_link_data(subgraph, edges="edges"), ensure_ascii=False, indent=2
|
||||
nx.node_link_data(subgraph, edges="edges"), ensure_ascii=False
|
||||
),
|
||||
"knowledge_graph_kwd": "subgraph",
|
||||
"kb_id": kb_id,
|
||||
@ -167,6 +168,11 @@ async def update_graph(
|
||||
"removed_kwd": "N",
|
||||
}
|
||||
cid = chunk_id(chunk)
|
||||
await trio.to_thread.run_sync(
|
||||
lambda: settings.docStoreConn.delete(
|
||||
{"knowledge_graph_kwd": "subgraph", "source_id": doc_id}, search.index_name(tenant_id), kb_id
|
||||
)
|
||||
)
|
||||
await trio.to_thread.run_sync(
|
||||
lambda: settings.docStoreConn.insert(
|
||||
[{"id": cid, **chunk}], search.index_name(tenant_id), kb_id
|
||||
@ -174,39 +180,49 @@ async def update_graph(
|
||||
)
|
||||
now = trio.current_time()
|
||||
callback(msg=f"generated subgraph for doc {doc_id} in {now - start:.2f} seconds.")
|
||||
start = now
|
||||
return subgraph
|
||||
|
||||
async def merge_subgraph(
|
||||
tenant_id: str,
|
||||
kb_id: str,
|
||||
doc_id: str,
|
||||
subgraph: nx.Graph,
|
||||
embedding_model,
|
||||
callback,
|
||||
):
|
||||
graphrag_task_lock = RedisDistributedLock("graphrag_task", lock_value=doc_id, timeout=600)
|
||||
while True:
|
||||
if graphrag_task_lock.acquire():
|
||||
break
|
||||
callback(msg=f"merge_subgraph {doc_id} is waiting graphrag_task_lock")
|
||||
await trio.sleep(10)
|
||||
|
||||
start = trio.current_time()
|
||||
change = GraphChange()
|
||||
old_graph = await get_graph(tenant_id, kb_id)
|
||||
if old_graph is not None:
|
||||
logging.info("Merge with an exiting graph...................")
|
||||
tidy_graph(old_graph, callback)
|
||||
new_graph = graph_merge(old_graph, subgraph, change)
|
||||
else:
|
||||
new_graph = subgraph
|
||||
now_docids = set([doc_id])
|
||||
old_graph, old_doc_ids = await get_graph(tenant_id, kb_id)
|
||||
if old_graph is not None:
|
||||
logging.info("Merge with an exiting graph...................")
|
||||
new_graph = graph_merge(old_graph, subgraph)
|
||||
await update_nodes_pagerank_nhop_neighbour(tenant_id, kb_id, new_graph, 2)
|
||||
if old_doc_ids:
|
||||
for old_doc_id in old_doc_ids:
|
||||
now_docids.add(old_doc_id)
|
||||
old_doc_ids2 = await get_graph_doc_ids(tenant_id, kb_id)
|
||||
delta_doc_ids = set(old_doc_ids2) - set(old_doc_ids)
|
||||
if delta_doc_ids:
|
||||
callback(
|
||||
msg="The global graph has changed during merging, try again"
|
||||
)
|
||||
await trio.sleep(1)
|
||||
continue
|
||||
break
|
||||
await set_graph(tenant_id, kb_id, new_graph, list(now_docids))
|
||||
change.added_updated_nodes = set(new_graph.nodes())
|
||||
change.added_updated_edges = set(new_graph.edges())
|
||||
pr = nx.pagerank(new_graph)
|
||||
for node_name, pagerank in pr.items():
|
||||
new_graph.nodes[node_name]["pagerank"] = pagerank
|
||||
|
||||
await set_graph(tenant_id, kb_id, embedding_model, new_graph, change, callback)
|
||||
graphrag_task_lock.release()
|
||||
now = trio.current_time()
|
||||
callback(
|
||||
msg=f"merging subgraph for doc {doc_id} into the global graph done in {now - start:.2f} seconds."
|
||||
)
|
||||
return new_graph, now_docids
|
||||
return new_graph
|
||||
|
||||
|
||||
async def resolve_entities(
|
||||
graph,
|
||||
doc_ids,
|
||||
tenant_id: str,
|
||||
kb_id: str,
|
||||
doc_id: str,
|
||||
@ -214,74 +230,30 @@ async def resolve_entities(
|
||||
embed_bdl,
|
||||
callback,
|
||||
):
|
||||
working_doc_id = graphrag_task_get(tenant_id, kb_id)
|
||||
if doc_id != working_doc_id:
|
||||
callback(
|
||||
msg=f"Another graphrag task of doc_id {working_doc_id} is working on this kb, cancel myself"
|
||||
)
|
||||
return
|
||||
graphrag_task_lock = RedisDistributedLock("graphrag_task", lock_value=doc_id, timeout=600)
|
||||
while True:
|
||||
if graphrag_task_lock.acquire():
|
||||
break
|
||||
await trio.sleep(10)
|
||||
|
||||
start = trio.current_time()
|
||||
er = EntityResolution(
|
||||
llm_bdl,
|
||||
get_entity=partial(get_entity, tenant_id, kb_id),
|
||||
set_entity=partial(set_entity, tenant_id, kb_id, embed_bdl),
|
||||
get_relation=partial(get_relation, tenant_id, kb_id),
|
||||
set_relation=partial(set_relation, tenant_id, kb_id, embed_bdl),
|
||||
)
|
||||
reso = await er(graph, callback=callback)
|
||||
graph = reso.graph
|
||||
callback(msg=f"Graph resolution removed {len(reso.removed_entities)} nodes.")
|
||||
await update_nodes_pagerank_nhop_neighbour(tenant_id, kb_id, graph, 2)
|
||||
change = reso.change
|
||||
callback(msg=f"Graph resolution removed {len(change.removed_nodes)} nodes and {len(change.removed_edges)} edges.")
|
||||
callback(msg="Graph resolution updated pagerank.")
|
||||
|
||||
working_doc_id = graphrag_task_get(tenant_id, kb_id)
|
||||
if doc_id != working_doc_id:
|
||||
callback(
|
||||
msg=f"Another graphrag task of doc_id {working_doc_id} is working on this kb, cancel myself"
|
||||
)
|
||||
return
|
||||
await set_graph(tenant_id, kb_id, graph, doc_ids)
|
||||
|
||||
await trio.to_thread.run_sync(
|
||||
lambda: settings.docStoreConn.delete(
|
||||
{
|
||||
"knowledge_graph_kwd": "relation",
|
||||
"kb_id": kb_id,
|
||||
"from_entity_kwd": reso.removed_entities,
|
||||
},
|
||||
search.index_name(tenant_id),
|
||||
kb_id,
|
||||
)
|
||||
)
|
||||
await trio.to_thread.run_sync(
|
||||
lambda: settings.docStoreConn.delete(
|
||||
{
|
||||
"knowledge_graph_kwd": "relation",
|
||||
"kb_id": kb_id,
|
||||
"to_entity_kwd": reso.removed_entities,
|
||||
},
|
||||
search.index_name(tenant_id),
|
||||
kb_id,
|
||||
)
|
||||
)
|
||||
await trio.to_thread.run_sync(
|
||||
lambda: settings.docStoreConn.delete(
|
||||
{
|
||||
"knowledge_graph_kwd": "entity",
|
||||
"kb_id": kb_id,
|
||||
"entity_kwd": reso.removed_entities,
|
||||
},
|
||||
search.index_name(tenant_id),
|
||||
kb_id,
|
||||
)
|
||||
)
|
||||
await set_graph(tenant_id, kb_id, embed_bdl, graph, change, callback)
|
||||
graphrag_task_lock.release()
|
||||
now = trio.current_time()
|
||||
callback(msg=f"Graph resolution done in {now - start:.2f}s.")
|
||||
|
||||
|
||||
async def extract_community(
|
||||
graph,
|
||||
doc_ids,
|
||||
tenant_id: str,
|
||||
kb_id: str,
|
||||
doc_id: str,
|
||||
@ -289,49 +261,34 @@ async def extract_community(
|
||||
embed_bdl,
|
||||
callback,
|
||||
):
|
||||
working_doc_id = graphrag_task_get(tenant_id, kb_id)
|
||||
if doc_id != working_doc_id:
|
||||
callback(
|
||||
msg=f"Another graphrag task of doc_id {working_doc_id} is working on this kb, cancel myself"
|
||||
)
|
||||
return
|
||||
graphrag_task_lock = RedisDistributedLock("graphrag_task", lock_value=doc_id, timeout=600)
|
||||
while True:
|
||||
if graphrag_task_lock.acquire():
|
||||
break
|
||||
await trio.sleep(10)
|
||||
|
||||
start = trio.current_time()
|
||||
ext = CommunityReportsExtractor(
|
||||
llm_bdl,
|
||||
get_entity=partial(get_entity, tenant_id, kb_id),
|
||||
set_entity=partial(set_entity, tenant_id, kb_id, embed_bdl),
|
||||
get_relation=partial(get_relation, tenant_id, kb_id),
|
||||
set_relation=partial(set_relation, tenant_id, kb_id, embed_bdl),
|
||||
)
|
||||
cr = await ext(graph, callback=callback)
|
||||
community_structure = cr.structured_output
|
||||
community_reports = cr.output
|
||||
working_doc_id = graphrag_task_get(tenant_id, kb_id)
|
||||
if doc_id != working_doc_id:
|
||||
callback(
|
||||
msg=f"Another graphrag task of doc_id {working_doc_id} is working on this kb, cancel myself"
|
||||
)
|
||||
return
|
||||
await set_graph(tenant_id, kb_id, graph, doc_ids)
|
||||
doc_ids = graph.graph["source_id"]
|
||||
|
||||
now = trio.current_time()
|
||||
callback(
|
||||
msg=f"Graph extracted {len(cr.structured_output)} communities in {now - start:.2f}s."
|
||||
)
|
||||
start = now
|
||||
await trio.to_thread.run_sync(
|
||||
lambda: settings.docStoreConn.delete(
|
||||
{"knowledge_graph_kwd": "community_report", "kb_id": kb_id},
|
||||
search.index_name(tenant_id),
|
||||
kb_id,
|
||||
)
|
||||
)
|
||||
chunks = []
|
||||
for stru, rep in zip(community_structure, community_reports):
|
||||
obj = {
|
||||
"report": rep,
|
||||
"evidences": "\n".join([f["explanation"] for f in stru["findings"]]),
|
||||
}
|
||||
chunk = {
|
||||
"id": get_uuid(),
|
||||
"docnm_kwd": stru["title"],
|
||||
"title_tks": rag_tokenizer.tokenize(stru["title"]),
|
||||
"content_with_weight": json.dumps(obj, ensure_ascii=False),
|
||||
@ -349,17 +306,23 @@ async def extract_community(
|
||||
chunk["content_sm_ltks"] = rag_tokenizer.fine_grained_tokenize(
|
||||
chunk["content_ltks"]
|
||||
)
|
||||
# try:
|
||||
# ebd, _ = embed_bdl.encode([", ".join(community["entities"])])
|
||||
# chunk["q_%d_vec" % len(ebd[0])] = ebd[0]
|
||||
# except Exception as e:
|
||||
# logging.exception(f"Fail to embed entity relation: {e}")
|
||||
await trio.to_thread.run_sync(
|
||||
lambda: settings.docStoreConn.insert(
|
||||
[{"id": chunk_id(chunk), **chunk}], search.index_name(tenant_id)
|
||||
)
|
||||
)
|
||||
chunks.append(chunk)
|
||||
|
||||
await trio.to_thread.run_sync(
|
||||
lambda: settings.docStoreConn.delete(
|
||||
{"knowledge_graph_kwd": "community_report", "kb_id": kb_id},
|
||||
search.index_name(tenant_id),
|
||||
kb_id,
|
||||
)
|
||||
)
|
||||
es_bulk_size = 4
|
||||
for b in range(0, len(chunks), es_bulk_size):
|
||||
doc_store_result = await trio.to_thread.run_sync(lambda: settings.docStoreConn.insert(chunks[b:b + es_bulk_size], search.index_name(tenant_id), kb_id))
|
||||
if doc_store_result:
|
||||
error_message = f"Insert chunk error: {doc_store_result}, please check log file and Elasticsearch/Infinity status!"
|
||||
raise Exception(error_message)
|
||||
|
||||
graphrag_task_lock.release()
|
||||
now = trio.current_time()
|
||||
callback(
|
||||
msg=f"Graph indexed {len(cr.structured_output)} communities in {now - start:.2f}s."
|
||||
|
||||
Reference in New Issue
Block a user