Implement metadata filtering post retrieval

[rahulkrprajapati]

I meant is implement a metadata filter to filter based on this:

MetadataFilters(
                        filters=[
                            ExactMatchFilter(key="namespace", value=self.namespace)
                        ]
                    ),

    def process_streamlit_fetch_query_results(self, query="", verbose=False):
        try:
            query_bundle = QueryBundle(query_str=query)
            retrived_nodes = self.fusion_retriever.retrieve(query_bundle)
            recency_nodes = self.recency_postprocessor.postprocess_nodes(
                retrived_nodes, query_bundle=query_bundle
            )
            rerank_nodes = self.postprocessor.postprocess_nodes(
                nodes=recency_nodes, query_bundle=query_bundle
            )
            px.active_session().url
            return rerank_nodes
        except Exception as e:
            traceback.print_exc()
            raise Exception(f"An error occurred retrieving: {e}")

[Ultroncn]

At that point in your pipeline, filter the retrieved NodeWithScore objects directly, before recency processing and reranking:

query_bundle = QueryBundle(query_str=query)
retrieved_nodes = self.fusion_retriever.retrieve(query_bundle)

namespace = self.namespace
if not isinstance(namespace, str):
    raise ValueError("namespace must be a string")

filtered_nodes = [
    hit for hit in retrieved_nodes
    if hit.node.metadata.get("namespace") == namespace
]

recency_nodes = self.recency_postprocessor.postprocess_nodes(
    filtered_nodes, query_bundle=query_bundle
)
rerank_nodes = self.postprocessor.postprocess_nodes(
    nodes=recency_nodes, query_bundle=query_bundle
)
return rerank_nodes

This implements scalar, exact namespace equality and excludes nodes missing that metadata. It preserves the original nodes and retrieval scores. MetadataFilters(...) describes filters for a supporting retriever/vector-store integration; constructing one does not automatically filter a Python list already returned by fusion.

One tradeoff: if fusion returns 10 candidates and only 2 match, post-filtering leaves 2; it cannot recover matching nodes that never entered the candidate set. Prefer passing the namespace filter into each supporting underlying retriever when possible, and evaluate candidate limits if you must post-filter.

If namespace represents a tenant/security boundary, apply the restriction to every retrieval branch before retrieval, as well as any defensive final check. Post-filtering after an external reranker/LLM has seen unauthorized content would be too late.

Reference: NodeWithScore.node and metadata access.