Add generic entities

README.md

@@ -30,7 +30,7 @@
 
 ## News (and Coming Soon)
 - [ ] Support for IDF weightening of entities
-- [ ] Support for generic entities and concepts
+- [x] Support for generic entities and concepts (initial commit)
 - [x] [2024.12.02] Benchmarks comparing Fast GraphRAG to LightRAG, GraphRAG and VectorDBs released [here](https://github.com/circlemind-ai/fast-graphrag/blob/main/benchmarks/README.md)
 
 ## Features

fast_graphrag/_models.py

@@ -53,8 +53,8 @@ def dump_to_csv(
             chain(
                 separator.join(
                     chain(
-                        (str(getattr(d, field)).replace("\t", "    ") for field in fields),
-                        (str(v).replace("\t", "    ") for v in vs),
+                        (str(getattr(d, field)).replace("\n", "  ").replace("\t", " ") for field in fields),
+                        (str(v).replace("\n", "  ").replace("\t", " ") for v in vs),
                     )
                 )
                 for d, *vs in zip(data, *values.values())
@@ -97,13 +97,21 @@ class TEntityDescription(BaseModel):
 
 
 class TQueryEntities(BaseModel):
-    entities: List[str] = Field(
+    named: List[str] = Field(
         ...,
-        description=("List of entities extracted from the query"),
+        description=("List of named entities extracted from the query"),
+    )
+    generic: List[str] = Field(
+        ...,
+        description=("List of generic entities extracted from the query"),
     )
-    n: int = Field(..., description="Number of named entities found")  # So that the LLM can answer 0.
 
-    @field_validator("entities", mode="before")
+    @field_validator("named", mode="before")
     @classmethod
-    def uppercase_source(cls, value: List[str]):
+    def uppercase_named(cls, value: List[str]):
         return [e.upper() for e in value] if value else value
+
+    # @field_validator("generic", mode="before")
+    # @classmethod
+    # def uppercase_generic(cls, value: List[str]):
+    #     return [e.upper() for e in value] if value else value

fast_graphrag/_prompt.py

@@ -47,12 +47,12 @@
 	{{"source": "Radio City", "target": "English", "desc": "Radio City broadcasts English songs."}},
 	{{"source": "Radio City", "target": "Hindi", "desc": "Radio City broadcasts songs in the Hindi language."}},
 	{{"source": "Radio City", "target": "PlanetRadiocity.com", "desc": "Radio City launched PlanetRadiocity.com in May 2008."}},
-	{{"source": "PlanetRadiocity.com", "target": "music portal", "desc": "PlanetRadiocity.com is a music portal that offers music related news, videos and more."}}
+	{{"source": "PlanetRadiocity.com", "target": "music portal", "desc": "PlanetRadiocity.com is a music portal that offers music related news, videos and more."}},
+	{{"source": "PlanetRadiocity.com", "target": "video", "desc": "PlanetRadiocity.com offers music related videos."}}
 	],
 	"other_relationships": [
 	{{"source": "Radio City", "target": "New Media", "desc": "Radio City forayed into New Media in May 2008."}},
 	{{"source": "PlanetRadiocity.com", "target": "news", "desc": "PlanetRadiocity.com offers music related news."}},
-	{{"source": "PlanetRadiocity.com", "target": "video", "desc": "PlanetRadiocity.com offers music related videos."}},
 	{{"source": "PlanetRadiocity.com", "target": "song", "desc": "PlanetRadiocity.com offers songs."}}
 	]
 }}
@@ -68,22 +68,23 @@
 
 PROMPTS["entity_relationship_gleaning_done_extraction"] = "Retrospectively check if all entities have been correctly identified: answer done if so, or continue if there are still entities that need to be added."
 
-PROMPTS["entity_extraction_query"] = """You are a helpful assistant that helps a human analyst identify all the named entities present in the input query that are important for answering the query.
+PROMPTS["entity_extraction_query"] = """Given the query below, your task is to extract all entities relevant to perform information retrieval to produce an answer.
 
-# Example 1
-Query: Do the magazines Arthur's Magazine or First for Women have the same publisher?
-Ouput: {{"entities": ["Arthur's Magazine", "First for Women"], "n": 2}}
+-Example 1-
+Query: Who directed the film that was shot in or around Leland, North Carolina in 1986?
+Ouput: {{"named": ["Leland", "North Carolina", "1986"], "generic": ["film director"]}}
 
-# Example 2
-Query: Which film has the director who was born earlier, Avatar II: The Return or The Interstellar?
-Ouput: {{"entities": ["Avatar II: The Return", "The Interstellar"], "n": 2}}
+-Example 2-
+Query: What relationship does Fred Gehrke have to the 23rd overall pick in the 2010 Major League Baseball Draft?
+Ouput: {{"named": ["Fred Gehrke", "2010 Major League Baseball Draft"], "generic": ["23rd baseball draft pick"]}}
 
 # INPUT
 Query: {query}
 Output:
 """
 
 
+
 PROMPTS[
 	"summarize_entity_descriptions"
 ] = """You are a helpful assistant responsible for generating a comprehensive summary of the data provided below.

fast_graphrag/_services/_base.py

@@ -22,7 +22,6 @@
     GTNode,
     TContext,
     TDocument,
-    TEntity,
     TIndex,
 )
 
@@ -58,7 +57,7 @@ def extract(
 
     async def extract_entities_from_query(
         self, llm: BaseLLMService, query: str, prompt_kwargs: Dict[str, str]
-    ) -> Iterable[TEntity]:
+    ) -> Dict[str, List[str]]:
         """Extract entities from the given query."""
         raise NotImplementedError
 
@@ -128,7 +127,7 @@ async def upsert(
         raise NotImplementedError
 
     async def get_context(
-        self, query: str, entities: Iterable[TEntity]
+        self, query: str, entities: Dict[str, List[str]]
     ) -> Optional[TContext[GTNode, GTEdge, GTHash, GTChunk]]:
         """Retrieve relevant state from the storage."""
         raise NotImplementedError

fast_graphrag/_services/_information_extraction.py

@@ -41,7 +41,7 @@ def extract(
 
     async def extract_entities_from_query(
         self, llm: BaseLLMService, query: str, prompt_kwargs: Dict[str, str]
-    ) -> Iterable[TEntity]:
+    ) -> Dict[str, List[str]]:
         """Extract entities from the given query."""
         prompt_kwargs["query"] = query
         entities, _ = await format_and_send_prompt(
@@ -51,7 +51,10 @@ async def extract_entities_from_query(
             response_model=TQueryEntities,
         )
 
-        return [TEntity(name=name, type="", description="") for name in entities.entities]
+        return {
+            "named": entities.named,
+            "generic": entities.generic
+        }
 
     async def _extract(
         self, llm: BaseLLMService, chunks: Iterable[TChunk], prompt_kwargs: Dict[str, str], entity_types: List[str]

fast_graphrag/_services/_state_manager.py

@@ -5,7 +5,7 @@
 
 import numpy as np
 import numpy.typing as npt
-from scipy.sparse import csr_matrix
+from scipy.sparse import csr_matrix, vstack
 from tqdm import tqdm
 
 from fast_graphrag._llm import BaseLLMService
@@ -35,7 +35,8 @@
 @dataclass
 class DefaultStateManagerService(BaseStateManagerService[TEntity, TRelation, THash, TChunk, TId, TEmbedding]):
     blob_storage_cls: Type[BaseBlobStorage[csr_matrix]] = field(default=PickleBlobStorage)
-    similarity_score_threshold: float = field(default=0.80)
+    insert_similarity_score_threshold: float = field(default=0.8)
+    query_similarity_score_threshold: Optional[float] = field(default=0.7)
 
     def __post_init__(self):
         assert self.workspace is not None, "Workspace must be provided."
@@ -77,7 +78,7 @@ async def upsert(
         llm: BaseLLMService,
         subgraphs: List[asyncio.Future[Optional[BaseGraphStorage[TEntity, TRelation, TId]]]],
         documents: Iterable[Iterable[TChunk]],
-        show_progress: bool = True
+        show_progress: bool = True,
     ) -> None:
         nodes: Iterable[List[TEntity]]
         edges: Iterable[List[TRelation]]
@@ -95,12 +96,16 @@ async def _get_graphs(
 
             return (nodes, edges)
 
-        graphs = [r for graph in tqdm(
-            asyncio.as_completed([_get_graphs(fgraph) for fgraph in subgraphs]),
-            total=len(subgraphs),
-            desc="Extracting data",
-            disable=not show_progress,
-        ) if (r := await graph) is not None]
+        graphs = [
+            r
+            for graph in tqdm(
+                asyncio.as_completed([_get_graphs(fgraph) for fgraph in subgraphs]),
+                total=len(subgraphs),
+                desc="Extracting data",
+                disable=not show_progress,
+            )
+            if (r := await graph) is not None
+        ]
 
         if len(graphs) == 0:
             return
@@ -136,7 +141,7 @@ async def _get_graphs(
         # We use the fact that similarity scores are symmetric between entity pairs,
         # so we only select half of that by index order
         similar_indices[
-            (scores < self.similarity_score_threshold)
+            (scores < self.insert_similarity_score_threshold)
             | (similar_indices <= upserted_indices)  # remove indices smaller or equal the entity
         ] = 0  # 0 can be used here (not 100% sure, but 99% sure)
         progress_bar.update(1)
@@ -147,6 +152,7 @@ async def _get_graphs(
 
         # STEP: insert identity edges
         progress_bar.set_description("Building... [identity edges]")
+
         async def _insert_identiy_edges(
             source_index: TIndex, target_indices: npt.NDArray[np.int32]
         ) -> Iterable[Tuple[TIndex, TIndex]]:
@@ -177,30 +183,33 @@ async def _insert_identiy_edges(
         progress_bar.set_description("Building [done]")
 
     async def get_context(
-        self, query: str, entities: Iterable[TEntity]
+        self, query: str, entities: Dict[str, List[str]]
     ) -> Optional[TContext[TEntity, TRelation, THash, TChunk]]:
         if self.entity_storage.size == 0:
             return None
 
         try:
-            entity_names = [entity.name for entity in entities]
-
-            query_embeddings = await self.embedding_service.encode(entity_names + [query])
-
+            query_embeddings = await self.embedding_service.encode(
+                [f"[NAME] {n}" for n in entities["named"]] + [f"[NAME] {n}" for n in entities["generic"]] + [query]
+            )
+            entity_scores: List[csr_matrix] = []
             # Similarity-search over entities
-            if len(entity_names) > 0:
-                vdb_entity_scores_by_name = await self._score_entities_by_vectordb(
-                    query_embeddings=query_embeddings[:-1], top_k=1
+            if len(entities["named"]) > 0:
+                vdb_entity_scores_by_named_entity = await self._score_entities_by_vectordb(
+                    query_embeddings=query_embeddings[: len(entities["named"])],
+                    top_k=1,
+                    threshold=self.query_similarity_score_threshold,
                 )
-            else:
-                vdb_entity_scores_by_name = 0
-            vdb_entity_scores_by_query = await self._score_entities_by_vectordb(
-                query_embeddings=query_embeddings[-1:], top_k=8
+                entity_scores.append(vdb_entity_scores_by_named_entity)
+
+            vdb_entity_scores_by_generic_entity_and_query = await self._score_entities_by_vectordb(
+                query_embeddings=query_embeddings[len(entities["named"]) :], top_k=20, threshold=0.5
             )
+            entity_scores.append(vdb_entity_scores_by_generic_entity_and_query)
 
-            vdb_entity_scores = vdb_entity_scores_by_name + vdb_entity_scores_by_query
+            vdb_entity_scores = vstack(entity_scores).max(axis=0)
 
-            if vdb_entity_scores.nnz == 0:
+            if isinstance(vdb_entity_scores, int) or vdb_entity_scores.nnz == 0:
                 return None
         except Exception as e:
             logger.error(f"Error during information extraction and scoring for query entities {entities}.\n{e}")
@@ -254,24 +263,25 @@ async def get_context(
     async def _get_entities_to_num_docs(self) -> Any:
         raise NotImplementedError
 
-    async def _score_entities_by_vectordb(self, query_embeddings: Iterable[TEmbedding], top_k: int = 1) -> csr_matrix:
+    async def _score_entities_by_vectordb(
+        self, query_embeddings: Iterable[TEmbedding], top_k: int = 1, threshold: Optional[float] = None
+    ) -> csr_matrix:
         # TODO: check this
         # if top_k != 1:
         #     logger.warning(f"Top-k > 1 is not tested yet. Using top_k={top_k}.")
         if self.node_specificity:
             raise NotImplementedError("Node specificity is not supported yet.")
 
         all_entity_probs_by_query_entity = await self.entity_storage.score_all(
-            np.array(query_embeddings), top_k=top_k
+            np.array(query_embeddings), top_k=top_k, threshold=threshold
         )  # (#query_entities, #all_entities)
 
         # TODO: if top_k > 1, we need to aggregate the scores here
         if all_entity_probs_by_query_entity.shape[1] == 0:
             return all_entity_probs_by_query_entity
-        all_entity_weights: csr_matrix = all_entity_probs_by_query_entity.max(axis=0)  # (1, #all_entities)
-
         # Normalize the scores
-        all_entity_weights /= all_entity_weights.sum()
+        all_entity_probs_by_query_entity /= all_entity_probs_by_query_entity.sum(axis=1) + 1e-8
+        all_entity_weights: csr_matrix = all_entity_probs_by_query_entity.max(axis=0)  # (1, #all_entities)
 
         if self.node_specificity:
             all_entity_weights = all_entity_weights.multiply(1.0 / await self._get_entities_to_num_docs())

fast_graphrag/_storage/_base.py

@@ -184,7 +184,7 @@ async def upsert(
         raise NotImplementedError
 
     async def score_all(
-        self, embeddings: Iterable[GTEmbedding], top_k: int = 1, confidence_threshold: float = 0.0
+        self, embeddings: Iterable[GTEmbedding], top_k: int = 1, threshold: Optional[float] = None
     ) -> csr_matrix:
         """Score all embeddings against the given queries.
 

fast_graphrag/_storage/_vdb_hnswlib.py

@@ -1,6 +1,6 @@
 import pickle
 from dataclasses import dataclass, field
-from typing import Any, Dict, Iterable, List, Tuple, Union
+from typing import Any, Dict, Iterable, List, Optional, Tuple, Union
 
 import hnswlib
 import numpy as np
@@ -16,9 +16,9 @@
 
 @dataclass
 class HNSWVectorStorageConfig:
-    ef_construction: int = field(default=64)
-    M: int = field(default=48)
-    ef_search: int = field(default=64)
+    ef_construction: int = field(default=256)
+    M: int = field(default=64)
+    ef_search: int = field(default=96)
     num_threads: int = field(default=-1)
 
 
@@ -83,10 +83,8 @@ async def get_knn(
         return ids, 1.0 - np.array(distances, dtype=TScore) * 0.5
 
     async def score_all(
-        self, embeddings: Iterable[GTEmbedding], top_k: int = 1, confidence_threshold: float = 0.0
+        self, embeddings: Iterable[GTEmbedding], top_k: int = 1, threshold: Optional[float] = None
     ) -> csr_matrix:
-        if confidence_threshold > 0.0:
-            raise NotImplementedError("Confidence threshold is not supported yet.")
         if not isinstance(embeddings, np.ndarray):
             embeddings = np.array(list(embeddings), dtype=np.float32)
 
@@ -104,6 +102,9 @@ async def score_all(
         ids = np.array(ids)
         scores = 1.0 - np.array(distances, dtype=TScore) * 0.5
 
+        if threshold is not None:
+            scores[scores < threshold] = 0
+
         # Create sparse distance matrix with shape (#embeddings, #all_embeddings)
         flattened_ids = ids.ravel()
         flattened_scores = scores.ravel()

tests/_models_test.py

@@ -15,9 +15,9 @@
 
 class TestModels(unittest.TestCase):
     def test_tqueryentities(self):
-        query_entities = TQueryEntities(entities=["Entity1", "Entity2"], n=2)
-        self.assertEqual(query_entities.entities, ["ENTITY1", "ENTITY2"])
-        self.assertEqual(query_entities.n, 2)
+        query_entities = TQueryEntities(named=["Entity1", "Entity2"], generic=["Generic1", "Generic2"])
+        self.assertEqual(query_entities.named, ["ENTITY1", "ENTITY2"])
+        self.assertEqual(query_entities.generic, ["Generic1", "Generic2"])
 
         with self.assertRaises(ValidationError):
             TQueryEntities(entities=["Entity1", "Entity2"], n="two")

tests/_policies/_graph_upsert_test.py

@@ -75,7 +75,7 @@ async def test_call_same_node_no_summarize(self):
 
         self.assertEqual(
             upserted_nodes[0][1].description,
-            "This is a short random description 1.  This is a short random description 2.",
+            "This is a short random description 1.\nThis is a short random description 2.",
         )
 
         # Assertions