NLP graph parity (#1888)

* Update stopwords config

* Minor edits

* Update PMI

* Format

* Perf improvements

* Semver

* Remove edge collection apply

* Remove source/target apply

* Add edge weight to graph snapshot

* Revert breaking optimizations

* Add perf fixes back in

* Format/types

* Update defaults

* Fix source/target ordering

* Fix test

.semversioner/next-release/patch-20250422210800599071.json

@@ -0,0 +1,4 @@
+{
+  "type": "patch",
+  "description": "Brings parity with our latest NLP extraction approaches."
+}

docs/config/yaml.md

@@ -251,7 +251,7 @@ Parameters for manual graph pruning. This can be used to optimize the modularity
 - max_node_freq_std **float | None** - The maximum standard deviation of node frequency to allow.
 - min_node_degree **int** - The minimum node degree to allow.
 - max_node_degree_std **float | None** - The maximum standard deviation of node degree to allow.
-- min_edge_weight_pct **int** - The minimum edge weight percentile to allow.
+- min_edge_weight_pct **float** - The minimum edge weight percentile to allow.
 - remove_ego_nodes **bool** - Remove ego nodes.
 - lcc_only **bool** - Only use largest connected component.
 

graphrag/config/defaults.py

@@ -19,6 +19,9 @@ from graphrag.config.enums import (
     ReportingType,
     TextEmbeddingTarget,
 )
+from graphrag.index.operations.build_noun_graph.np_extractors.stop_words import (
+    EN_STOP_WORDS,
+)
 from graphrag.vector_stores.factory import VectorStoreType
 
 DEFAULT_OUTPUT_BASE_DIR = "output"
@@ -186,7 +189,7 @@ class TextAnalyzerDefaults:
     max_word_length: int = 15
     word_delimiter: str = " "
     include_named_entities: bool = True
-    exclude_nouns: None = None
+    exclude_nouns: list[str] = field(default_factory=lambda: EN_STOP_WORDS)
     exclude_entity_tags: list[str] = field(default_factory=lambda: ["DATE"])
     exclude_pos_tags: list[str] = field(
         default_factory=lambda: ["DET", "PRON", "INTJ", "X"]
@@ -317,8 +320,8 @@ class PruneGraphDefaults:
     max_node_freq_std: None = None
     min_node_degree: int = 1
     max_node_degree_std: None = None
-    min_edge_weight_pct: int = 40
-    remove_ego_nodes: bool = False
+    min_edge_weight_pct: float = 40.0
+    remove_ego_nodes: bool = True
     lcc_only: bool = False
 
 

graphrag/index/operations/build_noun_graph/build_noun_graph.py

@@ -3,8 +3,9 @@
 
 """Graph extraction using NLP."""
 
-import math
+from itertools import combinations
 
+import numpy as np
 import pandas as pd
 
 from graphrag.cache.noop_pipeline_cache import NoopPipelineCache
@@ -30,7 +31,6 @@ async def build_noun_graph(
         text_units, text_analyzer, num_threads=num_threads, cache=cache
     )
     edges_df = _extract_edges(nodes_df, normalize_edge_weights=normalize_edge_weights)
-
     return (nodes_df, edges_df)
 
 
@@ -69,7 +69,7 @@ async def _extract_nodes(
     noun_node_df = text_unit_df.explode("noun_phrases")
     noun_node_df = noun_node_df.rename(
         columns={"noun_phrases": "title", "id": "text_unit_id"}
-    ).drop_duplicates()
+    )
 
     # group by title and count the number of text units
     grouped_node_df = (
@@ -94,45 +94,37 @@ def _extract_edges(
     """
     text_units_df = nodes_df.explode("text_unit_ids")
     text_units_df = text_units_df.rename(columns={"text_unit_ids": "text_unit_id"})
-    text_units_df = (
-        text_units_df.groupby("text_unit_id").agg({"title": list}).reset_index()
-    )
-    text_units_df["edges"] = text_units_df["title"].apply(
-        lambda x: _create_relationships(x)
-    )
-    edge_df = text_units_df.explode("edges").loc[:, ["edges", "text_unit_id"]]
 
-    edge_df["source"] = edge_df["edges"].apply(
-        lambda x: x[0] if isinstance(x, tuple) else None
+    text_units_df = (
+        text_units_df.groupby("text_unit_id")
+        .agg({"title": lambda x: list(x) if len(x) > 1 else np.nan})
+        .reset_index()
     )
-    edge_df["target"] = edge_df["edges"].apply(
-        lambda x: x[1] if isinstance(x, tuple) else None
+    text_units_df = text_units_df.dropna()
+    titles = text_units_df["title"].tolist()
+    all_edges: list[list[tuple[str, str]]] = [list(combinations(t, 2)) for t in titles]
+
+    text_units_df = text_units_df.assign(edges=all_edges)  # type: ignore
+    edge_df = text_units_df.explode("edges")[["edges", "text_unit_id"]]
+
+    edge_df[["source", "target"]] = edge_df.loc[:, "edges"].to_list()
+    edge_df["min_source"] = edge_df[["source", "target"]].min(axis=1)
+    edge_df["max_target"] = edge_df[["source", "target"]].max(axis=1)
+    edge_df = edge_df.drop(columns=["source", "target"]).rename(
+        columns={"min_source": "source", "max_target": "target"}  # type: ignore
     )
+
     edge_df = edge_df[(edge_df.source.notna()) & (edge_df.target.notna())]
     edge_df = edge_df.drop(columns=["edges"])
-
-    # make sure source is always smaller than target
-    edge_df["source"], edge_df["target"] = zip(
-        *edge_df.apply(
-            lambda x: (x["source"], x["target"])
-            if x["source"] < x["target"]
-            else (x["target"], x["source"]),
-            axis=1,
-        ),
-        strict=False,
-    )
-
-    # group by source and target, count the number of text units and collect their ids
+    # group by source and target, count the number of text units
     grouped_edge_df = (
         edge_df.groupby(["source", "target"]).agg({"text_unit_id": list}).reset_index()
     )
     grouped_edge_df = grouped_edge_df.rename(columns={"text_unit_id": "text_unit_ids"})
     grouped_edge_df["weight"] = grouped_edge_df["text_unit_ids"].apply(len)
-
     grouped_edge_df = grouped_edge_df.loc[
         :, ["source", "target", "weight", "text_unit_ids"]
     ]
-
     if normalize_edge_weights:
         # use PMI weight instead of raw weight
         grouped_edge_df = _calculate_pmi_edge_weights(nodes_df, grouped_edge_df)
@@ -140,18 +132,6 @@ def _extract_edges(
     return grouped_edge_df
 
 
-def _create_relationships(
-    noun_phrases: list[str],
-) -> list[tuple[str, str]]:
-    """Create a (source, target) tuple pairwise for all noun phrases in a list."""
-    relationships = []
-    if len(noun_phrases) >= 2:
-        for i in range(len(noun_phrases) - 1):
-            for j in range(i + 1, len(noun_phrases)):
-                relationships.extend([(noun_phrases[i], noun_phrases[j])])
-    return relationships
-
-
 def _calculate_pmi_edge_weights(
     nodes_df: pd.DataFrame,
     edges_df: pd.DataFrame,
@@ -192,8 +172,7 @@ def _calculate_pmi_edge_weights(
         .drop(columns=[node_name_col])
         .rename(columns={"prop_occurrence": "target_prop"})
     )
-    edges_df[edge_weight_col] = edges_df.apply(
-        lambda x: math.log2(x["prop_weight"] / (x["source_prop"] * x["target_prop"])),
-        axis=1,
+    edges_df[edge_weight_col] = edges_df["prop_weight"] * np.log2(
+        edges_df["prop_weight"] / (edges_df["source_prop"] * edges_df["target_prop"])
     )
     return edges_df.drop(columns=["prop_weight", "source_prop", "target_prop"])

graphrag/index/operations/graph_to_dataframes.py

@@ -21,6 +21,14 @@ def graph_to_dataframes(
 
     edges = nx.to_pandas_edgelist(graph)
 
+    # we don't deal in directed graphs, but we do need to ensure consistent ordering for df joins
+    # nx loses the initial ordering
+    edges["min_source"] = edges[["source", "target"]].min(axis=1)
+    edges["max_target"] = edges[["source", "target"]].max(axis=1)
+    edges = edges.drop(columns=["source", "target"]).rename(
+        columns={"min_source": "source", "max_target": "target"}  # type: ignore
+    )
+
     if node_columns:
         nodes = nodes.loc[:, node_columns]
 

graphrag/index/operations/prune_graph.py

@@ -21,7 +21,7 @@ def prune_graph(
     max_node_freq_std: float | None = None,
     min_node_degree: int = 1,
     max_node_degree_std: float | None = None,
-    min_edge_weight_pct: float = 0,
+    min_edge_weight_pct: float = 40,
     remove_ego_nodes: bool = False,
     lcc_only: bool = False,
 ) -> nx.Graph:

graphrag/index/workflows/finalize_graph.py

@@ -38,7 +38,8 @@ async def run_workflow(
 
     if config.snapshots.graphml:
         # todo: extract graphs at each level, and add in meta like descriptions
-        graph = create_graph(relationships)
+        graph = create_graph(final_relationships, edge_attr=["weight"])
+
         await snapshot_graphml(
             graph,
             name="graph",

tests/verbs/test_prune_graph.py

@@ -28,4 +28,4 @@ async def test_prune_graph():
 
     nodes_actual = await load_table_from_storage("entities", context.storage)
 
-    assert len(nodes_actual) == 21
+    assert len(nodes_actual) == 20