ted index uses append-only immutable segments identified by UUID with an atomic meta.json tracking the segment list and schema. Tantivy and Lucene implement the term dictionary as a two-level finite-state transducer mapping term to term-ordinal to term-info. Postings lists are stored in blocks of 128 documents with delta-encoded bitpacked document ids and bitpacked term frequencies, with a variable-int fallback for the final block of 1 to 127 documents. Tantivy uses plain SIMD bitpacking for postings compression, not PForDelta or Roaring bitmaps, which remain alternatives to evaluate. The LSM-tree write-amplification advantage over B-plus-trees is not absolute; on modern storage hardware with transparent compression an optimized B-plus-tree matches or beats RocksDB write amplification. Choose the key-value storage engine LSM versus B-tree per target hardware rather than assuming LSM wins. DiskANN stores the Vamana graph on SSD, keeps Product Quantization codes in RAM for approximate distance, and keeps full-precision vectors on disk for precise re-ranking, enabling billion-scale vector search beyond RAM. IVF restricts approximate nearest neighbor search to nprobe of nlist k-means clusters; Product Quantization splits a vector into m sub-vectors, runs k-means per subspace, and stores m centroid-index codes per vector. Memory-mapped immutable segments with background merges give fast read-heavy retrieval and crash-safe atomic commits.