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

graphrag/general/community_report_prompt.py

@@ -2,7 +2,7 @@
 # Licensed under the MIT License
 """
 Reference:
- - [graphrag](https://github.com/microsoft/graphrag)
+ - [GraphRAG](https://github.com/microsoft/graphrag/blob/main/graphrag/prompts/index/community_report.py)
 """
 
 COMMUNITY_REPORT_PROMPT = """

graphrag/general/community_reports_extractor.py

@@ -40,13 +40,9 @@ class CommunityReportsExtractor(Extractor):
     def __init__(
             self,
             llm_invoker: CompletionLLM,
-            get_entity: Callable | None = None,
-            set_entity: Callable | None = None,
-            get_relation: Callable | None = None,
-            set_relation: Callable | None = None,
             max_report_length: int | None = None,
     ):
-        super().__init__(llm_invoker, get_entity=get_entity, set_entity=set_entity, get_relation=get_relation, set_relation=set_relation)
+        super().__init__(llm_invoker)
         """Init method definition."""
         self._llm = llm_invoker
         self._extraction_prompt = COMMUNITY_REPORT_PROMPT
@@ -63,21 +59,28 @@ class CommunityReportsExtractor(Extractor):
         over, token_count = 0, 0
         async def extract_community_report(community):
             nonlocal res_str, res_dict, over, token_count
-            cm_id, ents = community
-            weight = ents["weight"]
-            ents = ents["nodes"]
-            ent_df = pd.DataFrame(self._get_entity_(ents)).dropna()
-            if ent_df.empty or "entity_name" not in ent_df.columns:
+            cm_id, cm = community
+            weight = cm["weight"]
+            ents = cm["nodes"]
+            if len(ents) < 2:
                 return
-            ent_df["entity"] = ent_df["entity_name"]
-            del ent_df["entity_name"]
-            rela_df = pd.DataFrame(self._get_relation_(list(ent_df["entity"]), list(ent_df["entity"]), 10000))
-            if rela_df.empty:
-                return
-            rela_df["source"] = rela_df["src_id"]
-            rela_df["target"] = rela_df["tgt_id"]
-            del rela_df["src_id"]
-            del rela_df["tgt_id"]
+            ent_list = [{"entity": ent, "description": graph.nodes[ent]["description"]} for ent in ents]
+            ent_df = pd.DataFrame(ent_list)
+
+            rela_list = []
+            k = 0
+            for i in range(0, len(ents)):
+                if k >= 10000:
+                    break
+                for j in range(i + 1, len(ents)):
+                    if k >= 10000:
+                        break
+                    edge = graph.get_edge_data(ents[i], ents[j])
+                    if edge is None:
+                        continue
+                    rela_list.append({"source": ents[i], "target": ents[j], "description": edge["description"]})
+                    k += 1
+            rela_df = pd.DataFrame(rela_list)
 
             prompt_variables = {
                 "entity_df": ent_df.to_csv(index_label="id"),

graphrag/general/extractor.py

@@ -19,10 +19,11 @@ from collections import defaultdict, Counter
 from copy import deepcopy
 from typing import Callable
 import trio
+import networkx as nx
 
 from graphrag.general.graph_prompt import SUMMARIZE_DESCRIPTIONS_PROMPT
 from graphrag.utils import get_llm_cache, set_llm_cache, handle_single_entity_extraction, \
-    handle_single_relationship_extraction, split_string_by_multi_markers, flat_uniq_list, chat_limiter
+    handle_single_relationship_extraction, split_string_by_multi_markers, flat_uniq_list, chat_limiter, get_from_to, GraphChange
 from rag.llm.chat_model import Base as CompletionLLM
 from rag.prompts import message_fit_in
 from rag.utils import truncate
@@ -40,18 +41,10 @@ class Extractor:
         llm_invoker: CompletionLLM,
         language: str | None = "English",
         entity_types: list[str] | None = None,
-        get_entity: Callable | None = None,
-        set_entity: Callable | None = None,
-        get_relation: Callable | None = None,
-        set_relation: Callable | None = None,
     ):
         self._llm = llm_invoker
         self._language = language
         self._entity_types = entity_types or DEFAULT_ENTITY_TYPES
-        self._get_entity_ = get_entity
-        self._set_entity_ = set_entity
-        self._get_relation_ = get_relation
-        self._set_relation_ = set_relation
 
     def _chat(self, system, history, gen_conf):
         hist = deepcopy(history)
@@ -152,25 +145,15 @@ class Extractor:
     async def _merge_nodes(self, entity_name: str, entities: list[dict], all_relationships_data):
         if not entities:
             return
-        already_entity_types = []
-        already_source_ids = []
-        already_description = []
-
-        already_node = self._get_entity_(entity_name)
-        if already_node:
-            already_entity_types.append(already_node["entity_type"])
-            already_source_ids.extend(already_node["source_id"])
-            already_description.append(already_node["description"])
-
         entity_type = sorted(
             Counter(
-                [dp["entity_type"] for dp in entities] + already_entity_types
+                [dp["entity_type"] for dp in entities]
             ).items(),
             key=lambda x: x[1],
             reverse=True,
         )[0][0]
         description = GRAPH_FIELD_SEP.join(
-            sorted(set([dp["description"] for dp in entities] + already_description))
+            sorted(set([dp["description"] for dp in entities]))
         )
         already_source_ids = flat_uniq_list(entities, "source_id")
         description = await self._handle_entity_relation_summary(entity_name, description)
@@ -180,7 +163,6 @@ class Extractor:
             source_id=already_source_ids,
         )
         node_data["entity_name"] = entity_name
-        self._set_entity_(entity_name, node_data)
         all_relationships_data.append(node_data)
 
     async def _merge_edges(
@@ -192,36 +174,11 @@ class Extractor:
     ):
         if not edges_data:
             return
-        already_weights = []
-        already_source_ids = []
-        already_description = []
-        already_keywords = []
-
-        relation = self._get_relation_(src_id, tgt_id)
-        if relation:
-            already_weights = [relation["weight"]]
-            already_source_ids = relation["source_id"]
-            already_description = [relation["description"]]
-            already_keywords = relation["keywords"]
-
-        weight = sum([dp["weight"] for dp in edges_data] + already_weights)
-        description = GRAPH_FIELD_SEP.join(
-            sorted(set([dp["description"] for dp in edges_data] + already_description))
-        )
-        keywords = flat_uniq_list(edges_data, "keywords") + already_keywords
-        source_id = flat_uniq_list(edges_data, "source_id") + already_source_ids
-
-        for need_insert_id in [src_id, tgt_id]:
-            if self._get_entity_(need_insert_id):
-                continue
-            self._set_entity_(need_insert_id, {
-                        "source_id": source_id,
-                        "description": description,
-                        "entity_type": 'UNKNOWN'
-                    })
-        description = await self._handle_entity_relation_summary(
-            f"({src_id}, {tgt_id})", description
-        )
+        weight = sum([edge["weight"] for edge in edges_data])
+        description = GRAPH_FIELD_SEP.join(sorted(set([edge["description"] for edge in edges_data])))
+        description = await self._handle_entity_relation_summary(f"{src_id} -> {tgt_id}", description)
+        keywords = flat_uniq_list(edges_data, "keywords")
+        source_id = flat_uniq_list(edges_data, "source_id")
         edge_data = dict(
             src_id=src_id,
             tgt_id=tgt_id,
@@ -230,9 +187,41 @@ class Extractor:
             weight=weight,
             source_id=source_id
         )
-        self._set_relation_(src_id, tgt_id, edge_data)
-        if all_relationships_data is not None:
-            all_relationships_data.append(edge_data)
+        all_relationships_data.append(edge_data)
+
+    async def _merge_graph_nodes(self, graph: nx.Graph, nodes: list[str], change: GraphChange):
+        if len(nodes) <= 1:
+            return
+        change.added_updated_nodes.add(nodes[0])
+        change.removed_nodes.extend(nodes[1:])
+        nodes_set = set(nodes)
+        node0_attrs = graph.nodes[nodes[0]]
+        node0_neighbors = set(graph.neighbors(nodes[0]))
+        for node1 in nodes[1:]:
+            # Merge two nodes, keep "entity_name", "entity_type", "page_rank" unchanged.
+            node1_attrs = graph.nodes[node1]
+            node0_attrs["description"] += f"{GRAPH_FIELD_SEP}{node1_attrs['description']}"
+            for attr in ["keywords", "source_id"]:
+                node0_attrs[attr] = sorted(set(node0_attrs[attr].extend(node1_attrs[attr])))
+            for neighbor in graph.neighbors(node1):
+                change.removed_edges.add(get_from_to(node1, neighbor))
+                if neighbor not in nodes_set:
+                    edge1_attrs = graph.get_edge_data(node1, neighbor)
+                    if neighbor in node0_neighbors:
+                        # Merge two edges
+                        change.added_updated_edges.add(get_from_to(nodes[0], neighbor))
+                        edge0_attrs = graph.get_edge_data(nodes[0], neighbor)
+                        edge0_attrs["weight"] += edge1_attrs["weight"]
+                        edge0_attrs["description"] += f"{GRAPH_FIELD_SEP}{edge1_attrs['description']}"
+                        edge0_attrs["keywords"] = list(set(edge0_attrs["keywords"].extend(edge1_attrs["keywords"])))
+                        edge0_attrs["source_id"] = list(set(edge0_attrs["source_id"].extend(edge1_attrs["source_id"])))
+                        edge0_attrs["description"] = await self._handle_entity_relation_summary(f"({nodes[0]}, {neighbor})", edge0_attrs["description"])
+                        graph.add_edge(nodes[0], neighbor, **edge0_attrs)
+                    else:
+                        graph.add_edge(nodes[0], neighbor, **edge1_attrs)
+            graph.remove_node(node1)
+        node0_attrs["description"] = await self._handle_entity_relation_summary(nodes[0], node0_attrs["description"])
+        graph.nodes[nodes[0]].update(node0_attrs)
 
     async def _handle_entity_relation_summary(
             self,

graphrag/general/graph_extractor.py

@@ -6,7 +6,7 @@ Reference:
 """
 
 import re
-from typing import Any, Callable
+from typing import Any
 from dataclasses import dataclass
 import tiktoken
 import trio
@@ -53,10 +53,6 @@ class GraphExtractor(Extractor):
         llm_invoker: CompletionLLM,
         language: str | None = "English",
         entity_types: list[str] | None = None,
-        get_entity: Callable | None = None,
-        set_entity: Callable | None = None,
-        get_relation: Callable | None = None,
-        set_relation: Callable | None = None,
         tuple_delimiter_key: str | None = None,
         record_delimiter_key: str | None = None,
         input_text_key: str | None = None,
@@ -66,7 +62,7 @@ class GraphExtractor(Extractor):
         max_gleanings: int | None = None,
         on_error: ErrorHandlerFn | None = None,
     ):
-        super().__init__(llm_invoker, language, entity_types, get_entity, set_entity, get_relation, set_relation)
+        super().__init__(llm_invoker, language, entity_types)
         """Init method definition."""
         # TODO: streamline construction
         self._llm = llm_invoker

graphrag/general/graph_prompt.py

@@ -2,7 +2,7 @@
 # Licensed under the MIT License
 """
 Reference:
- - [graphrag](https://github.com/microsoft/graphrag)
+ - [GraphRAG](https://github.com/microsoft/graphrag/blob/main/graphrag/prompts/index/extract_graph.py)
 """
 
 GRAPH_EXTRACTION_PROMPT = """

graphrag/general/index.py

@@ -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."

graphrag/general/leiden.py

@@ -100,7 +100,8 @@ def run(graph: nx.Graph, args: dict[str, Any]) -> dict[int, dict[str, dict]]:
         logging.debug(
             "Running leiden with max_cluster_size=%s, lcc=%s", max_cluster_size, use_lcc
         )
-    if not graph.nodes():
+    nodes = set(graph.nodes())
+    if not nodes:
         return {}
 
     node_id_to_community_map = _compute_leiden_communities(
@@ -120,7 +121,7 @@ def run(graph: nx.Graph, args: dict[str, Any]) -> dict[int, dict[str, dict]]:
         result = {}
         results_by_level[level] = result
         for node_id, raw_community_id in node_id_to_community_map[level].items():
-            if node_id not in graph.nodes:
+            if node_id not in nodes:
                 logging.warning(f"Node {node_id} not found in the graph.")
                 continue
             community_id = str(raw_community_id)