Pre-Filtering

Pre-filtering ensures that every document considered during the ANN traversal satisfies the metadata constraints. This guarantees that if the index contains matching documents, they will be found—unlike post-filtering, which can return empty result sets when semantic matches fail to meet filter criteria. The result count is bounded by the intersection of the filter and the semantic space.

The filter is applied to construct a bitmap or Roaring bitmap of eligible document IDs before the vector search begins. The ANN algorithm then restricts its graph traversal to only those nodes present in the bitmap. This is typically implemented through:

• Filtered HNSW: Only visiting nodes whose IDs are in the allowed set
• Pre-filtered IVF: Restricting the inverted file scan to clusters containing eligible documents

Pre-filtering introduces a filter evaluation cost before the vector search, but eliminates wasted computation on documents that would be discarded later. The latency profile depends on:

• Filter selectivity: Highly selective filters (e.g., a single user's documents) dramatically reduce the search space and speed up ANN traversal
• Filter complexity: Simple equality checks are fast; complex Boolean expressions with many OR clauses can increase pre-processing overhead
• Index structure: Some vector databases maintain separate indices per filter value for zero-cost filtering

Pre-filtering is the foundational mechanism behind faceted search in e-commerce and enterprise search. Users combine free-text queries with structured facets like:

• Price ranges: price:[100 TO 500]
• Categories: category:electronics
• Availability: in_stock:true The vector search then finds semantically similar items within the filtered subset, ensuring results are both relevant and contextually constrained.

Pre-filtering is the primary mechanism for document-level security in retrieval-augmented generation (RAG) systems. Each document is tagged with access control lists (ACLs) or user/group identifiers at indexing time. At query time, the user's credentials are converted into a filter that restricts the ANN search to only documents they are authorized to see, preventing data leakage across permission boundaries.

Pre-filtering applies constraints first, then searches semantically within the allowed set. Post-filtering searches semantically first, then discards results that don't match filters. The critical difference:

• Pre-filtering: Guaranteed recall within filter bounds, but may miss semantically strong matches just outside the filter
• Post-filtering: Finds the globally best semantic matches, but risks returning fewer than k results if top matches are filtered out Hybrid approaches use adaptive filtering to choose the strategy based on filter selectivity.

The direct sibling of pre-filtering, this mechanism restricts the candidate document set based on structured attributes like date ranges, document types, or access permissions. While pre-filtering applies constraints before the vector search, metadata filtering can also be applied post-retrieval. The choice between the two is a critical architectural decision that balances semantic recall against strict constraint enforcement.

The inverse strategy where an Approximate Nearest Neighbor (ANN) search is executed first to find the top semantic matches, and only then are metadata constraints applied to prune the results. This approach preserves semantic recall but carries the risk of empty result sets if none of the initial k-nearest neighbors satisfy the filter conditions, making it unsuitable for highly restrictive or sparse metadata criteria.

The foundational data structure that makes pre-filtering efficient. An inverted index maps each unique structured attribute value to a postings list of document IDs. During pre-filtering, the system intersects these lists to construct a bitmap or allowed-document set before the vector index is traversed. This allows the ANN algorithm to skip entire regions of the vector space that don't match the filter criteria.

A graph-based ANN algorithm that can natively integrate pre-filtering during its greedy traversal. When a filter is active, the search only considers edges that connect to nodes in the allowed set. This is more efficient than post-filtering because the algorithm never wastes distance computations on documents that would be discarded. Many vector databases, such as Weaviate and Qdrant, implement this as a core optimization.

An algorithm that merges multiple ranked result lists—such as a pre-filtered vector search and a separate sparse keyword search—into a single ranking. RRF assigns a score of 1 / (k + rank) to each document, where k is a constant (typically 60). This effectively balances the contributions of different retrieval systems without requiring score calibration, making it a robust choice for hybrid ensembles that include pre-filtered result sets.

The process of scaling the uncalibrated relevance scores from disparate retrieval sources onto a common, comparable scale before they can be meaningfully merged. When combining a pre-filtered dense search with a BM25 result set, methods like Min-Max or Z-Score normalization are essential. Without this step, the raw cosine similarity scores from the vector search would dominate the BM25 scores, negating the benefit of the hybrid approach.