Query Fusion Retriever with metadata filtering fails

bbeaverTango

I have a redis vector index store and a doc store and an upsatash vector store defined. When I define my retriever such that it filters the nodes by metadata it seems to fail.

Steps to Reproduce

defining a base retriever with metadata filter enabled

Query Fusion Retriever with metadata filtering fails

Plain Text

base_retriever = self.base_index.as_retriever(
                    similarity_top_k=self.similarity_top_k,
                    filters=MetadataFilters(
                        filters=[
                            ExactMatchFilter(key="namespace", value=self.namespace)
                        ]
                    ),
                )
retriever = AutoMergingRetriever(
                    base_retriever, self.storage_context, verbose=verbose
                )
                
bm24_retriever = BM25Retriever.from_defaults(
                docstore=self.docstore, similarity_top_k=self.similarity_top_k
            )

fusion_retriever = QueryFusionRetriever(
                [retriever, bm24_retriever],
                similarity_top_k=self.similarity_top_k,
                num_queries=1,  # set this to 1 to disable query generation
                mode="reciprocal_rerank",
                use_async=True,
                verbose=verbose,
            )

https://github.com/run-llama/llama_index/issues/11391#issue-2153660643

10 comments

bbeaverTango

Is this my upstash vector store that's is causing the issue ?

LLogan M

Is namespace actually a field you added to the metadata of each document?

bbeaverTango

yeah added it :

Plain Text

def create_document(self, text, filename):
        documents = []
        for idx, page in text.items():
            document = Document(text=page)
            current_date = datetime.now().strftime("%Y-%m-%d")
            document.metadata = {
                "filename": filename,
                "page_number": idx,
                "creation_date": current_date,
                "last_accessed_date": current_date,
                "last_modified_date": current_date,
                "namespace": self.namespace
            }
            documents.append(document)
        return documents

bbeaverTango

the current stack uses upstash vector store. Does namespace as a metadata field should also be present while defining the vector store or adding it to the doc is enough? Upstash as a vector store currently does not seem to have support for metadata fields

LLogan M

Looking at the upstash code, it should support storing metadata. But not filtering

bbeaverTango

It seemed to throw an error when I tried adding the metadata when I did this:

Plain Text

vector_store = UpstashVectorStore(
                    url=self.config.get("UPSTASH_VECTOR_URL"),
                    token=self.config.get("UPSTASH_VECTOR_TOKEN"),
                    metadata_fields=['namespace']
                )

LLogan M

metadata_fields=['namespace'] is not an arg in the source code 👀

LLogan M

Attachment