dosubot[bot]
commented
8 months ago š¤
Hey @mike4lyf, back at it again with the tech puzzles, I see! Hope you're doing well.
Thank you for reporting this issue. It seems like the EntityExtractor is returning a list of entities, which is causing a ValueError when the VectorStoreIndex tries to process the metadata. This is because VectorStoreIndex expects metadata values to be of type string or int.
The EntityExtractor class in the LlamaIndex codebase is used to extract entities from text and insert them into node metadata. It uses a default model tomaarsen/span-marker-mbert-base-multinerd and the SpanMarker library for entity extraction. The entities are then returned as a list of dictionaries. Here is the relevant code:
class EntityExtractor(BaseExtractor):
...
async def aextract(self, nodes: Sequence[BaseNode]) -> List[Dict]:
# Extract node-level entity metadata
metadata_list: List[Dict] = [{} for _ in nodes]
metadata_queue: Iterable[int] = get_tqdm_iterable(
range(len(nodes)), self.show_progress, "Extracting entities"
)
for i in metadata_queue:
metadata = metadata_list[i]
node_text = nodes[i].get_content(metadata_mode=self.metadata_mode)
words = self._tokenizer(node_text)
spans = self._model.predict(words)
for span in spans:
if span["score"] > self.prediction_threshold:
ent_label = self.entity_map.get(span["label"], span["label"])
metadata_label = ent_label if self.label_entities else "entities"
if metadata_label not in metadata:
metadata[metadata_label] = set()
metadata[metadata_label].add(self.span_joiner.join(span["span"]))
# convert metadata from set to list
for metadata in metadata_list:
for key, val in metadata.items():
metadata[key] = list(val)
return metadata_listTo resolve this issue, you could modify the EntityExtractor to return a string representation of the entities instead of a list. This can be done by joining the entities with a separator. Here is an example of how you could modify the EntityExtractor:
# convert metadata from set to list
for metadata in metadata_list:
for key, val in metadata.items():
metadata[key] = ', '.join(val) # join entities with a commaThis will ensure that the EntityExtractor returns a string, which can be processed by the VectorStoreIndex.
Please let me know if this resolves your issue or if you need further assistance.
Sources
- llama-index-integrations/vector_stores/llama-index-vector-stores-redis/llama_index/vector_stores/redis/base.py
- llama-index-integrations/extractors/llama-index-extractors-entity/llama_index/extractors/entity/base.py
erlebach
logan-markewich
Bug Description
VectorStoreIndex has metadata type issue when using EntityExtractor as a pipeline transformer. EntityExtractor returns a list, however, VectorStoreIndex expects metadata values to be string/int.
Other transformers (e.g. QuestionsAnsweredExtractor, KeywordExtractor) return strings.
Version
0.10.10
Steps to Reproduce
pipeline = IngestionPipeline(transformations=[ SemanticSplitterNodeParser(buffer_size=1, breakpoint_percentile_threshold=95, embed_model=OpenAIEmbedding()), TitleExtractor(nodes=5, llm=llm), QuestionsAnsweredExtractor(questions=3, llm=llm), EntityExtractor(prediction_threshold=0.5), KeywordExtractor(keywords=10, llm=llm) ]) nrma_home_nodes = pipeline.run(documents)
index = VectorStoreIndex(nrma_home_nodes, storage_context=storage_context)
Relevant Logs/Tracbacks