Skip to content

🧬 Learned Sparse Search (SPLADE)

Spector implements Learned Sparse Retrieval (LSR), supporting the SPLADE (Sparse Lexical and Expansion Model) architecture and the fast Li-LSR (Inference-Free Learned Sparse Retrieval) model.

Learned sparse retrieval combines the advantages of lexical search (exact-matching, fast inverted index lookups) with the advantages of dense vector search (semantic awareness, synonym matching).

🧠 How SPLADE Works

SPLADE encodes a document or query into a sparse vector that matches the size of the vocabulary (typically ~30,522 dimensions). However, instead of only marking terms that literally exist in the text (like BM25), the model performs term expansion:

Input: "Where is my flight status?"
        │
        ▼ (Neural Transformer Term Expansion)
Expanded Terms:
  - flight:    5.2
  - status:    4.8
  - airplane:  3.9  (expanded synonym)
  - departure: 3.5  (expanded synonym)
  - delay:     2.8  (expanded synonym)

By indexing these expanded terms with their relative weights into an inverted index, Spector captures synonyms and semantic relationships lexically, completely resolving the vocabulary mismatch problem.

🏗️ Architectural Concepts

Spector structures sparse retrieval through the following concepts:

1. Sparse Embeddings

When a text is ingested, it is passed to a sparse embedding model which outputs token IDs (or string terms) and their corresponding positive activation values (representing term importance weights).

2. Learned Sparse Indexing

Spector indexes sparse weights in a high-performance inverted index: - Inverted Posting Lists: The index maps each vocabulary token to a postings list of document IDs and activations. These lists are stored in contiguous memory structures for fast traversal. - Max-Score Pruning: At query time, Spector uses threshold-gating and max-score pruning algorithms to evaluate only the most promising postings lists, achieving sub-millisecond search latencies across millions of documents.

3. Li-LSR (Inference-Free Sparse)

For ultra-low latency requirements, Spector implements Li-LSR. This model replaces neural model query inference with static, high-quality term-expansion lookup tables. This eliminates the neural model execution penalty at query time, bringing learned sparse query latencies down to microseconds.

⚙️ Configuration & Execution

Sparse retrieval is controlled by global and query-level properties.

Global defaults in spector.yml

spector:
  memory:
    splade-enabled: true     # Master toggle for learned sparse search

Query-level settings

Configure sparse search via query parameters:

  • SPLADE: Runs learned sparse retrieval only, bypassing both dense HNSW and BM25 search.
  • SPLADE_HYBRID: Runs a hybrid query combining dense vector search and SPLADE, merging results via Reciprocal Rank Fusion.
  • LI_LSR: Forces query execution path to use the fast, inference-free Li-LSR lookup tables.
  • FULL_STACK: Runs BM25, SPLADE, and dense vector searches concurrently, then passes candidates to the late-interaction reranking layer.