Preliminary Ranker Using Bit Vector Search Index

The patent replaces traditional posting list data structures with bit vector-based indexing. Bit vectors use bitwise operations (AND, OR, NOT) to perform set operations on document collections, substituting the mechanical list traversal and merging operations of posting lists with efficient bitwise computations. This enables parallel processing and significantly faster update operations, as bit vectors can be updated in-place without requiring batch processing or reindexing entire document collections.

The patent implements preliminary filtering using bit vector operations before documents reach the final ranker. The system performs bitwise AND operations on bit vectors corresponding to query terms to pre-identify candidate document sets. This preliminary action eliminates obviously irrelevant documents early in the process, allowing the final ranker to focus only on promising candidates. The bit vector intersection operations provide a fast, approximate filtering mechanism that preserves ranking accuracy while dramatically reducing the number of documents requiring expensive full-ranking operations.

The patent changes the fundamental parameters of the indexing data structure from variable-length posting lists to fixed-width bit vectors. Each bit vector has a predetermined size based on the total document collection, with each bit representing a document's presence or absence. This parameter change enables O(1) update time complexity, as adding or removing a document requires only a single bit flip. The fixed-width structure provides predictable memory allocation and enables efficient use of cache memory, offsetting the apparent storage overhead through better memory utilization patterns.