SPLADE (Sparse Lexical and Expansion Model) is a learned sparse retrieval architecture that uses a masked language model (typically BERT) to predict term importance scores across the entire vocabulary, producing high-dimensional sparse vectors where each dimension corresponds to a specific token weight. Unlike dense retrieval models that map text into uninterpretable latent spaces, SPLADE generates explicit, lexically grounded representations that support both exact keyword matching and semantic term expansion—automatically assigning weight to synonyms and related terms not present in the original text.
Glossary
Splade (Sparse Lexical and Expansion Model)
A learned sparse retrieval model that uses a masked language model to predict term importance, generating high-dimensional but sparse query and document expansions.
LEARNED SPARSE RETRIEVAL
What is Splade (Sparse Lexical and Expansion Model)?
SPLADE is a neural retrieval model that learns to predict term importance weights for both queries and documents, generating high-dimensional but sparse bag-of-words representations for efficient lexical matching with semantic expansion.
The model is trained using contrastive learning with in-batch negatives and a sparsity-inducing regularization penalty (often FLOPS regularization) to prevent the representations from becoming dense. During inference, SPLADE encodes queries and documents into sparse vectors compatible with inverted index structures, enabling retrieval via efficient dot-product scoring using established sparse retrieval infrastructure like Lucene. This hybrid nature bridges the interpretability and speed of traditional sparse methods like BM25 with the semantic understanding of neural models, making it a compelling alternative to both dense bi-encoders and late-interaction architectures like ColBERT.
SPARSE LEXICAL AND EXPANSION MODEL
Key Features of SPLADE
SPLADE bridges the gap between traditional keyword search and neural vector search by using a Masked Language Model to predict term importance, generating high-dimensional but sparse query and document expansions.
01
Learned Sparse Representations
Unlike static methods like BM25, SPLADE uses a deep neural network to learn which terms are important. It performs document expansion by predicting relevant terms that may not literally appear in the text, and query expansion by enriching the search query with related vocabulary. This results in a sparse vector where each dimension corresponds to a token in the vocabulary, weighted by its learned importance.
02
Masked Language Model Backbone
SPLADE leverages a pre-trained Masked Language Model (MLM) like BERT. For each token in the input, the model predicts the importance of every term in the vocabulary. The key mechanism is the log-saturation activation function:
wij = log(1 + ReLU(wij))This prevents a few terms from dominating and ensures the sparsity profile is well-behaved for inverted index storage.
03
FLOPS Regularization for Efficiency
A core innovation in SPLADE is the use of FLOPS (Floating Point Operations) regularization during training. This loss term penalizes the model for activating too many non-zero dimensions. The result is a model that learns to be highly selective, producing expansions that are both effective and computationally efficient for retrieval. This directly controls the trade-off between recall and latency.
04
Inverted Index Compatibility
Unlike dense vector embeddings that require specialized Approximate Nearest Neighbor (ANN) indexes like HNSW, SPLADE's output is a sparse vector. This means it can be plugged directly into existing, battle-tested inverted index infrastructure (like Lucene) with no architectural changes. This is a massive operational advantage for teams already managing keyword search systems.
05
First-Stage Ranker Performance
SPLADE models are powerful enough to serve as a first-stage ranker, eliminating the need for a separate multi-stage retrieval pipeline in many applications. On the MS MARCO passage ranking benchmark, SPLADE variants achieve competitive MRR@10 scores against dense bi-encoders, while maintaining the interpretability and infrastructure simplicity of a sparse lexical system.
06
Interpretable Term Expansion
A major advantage over dense embeddings is interpretability. You can inspect the exact terms that SPLADE expanded for a query or document. For the query 'what is CPU', SPLADE might expand it with terms like 'processor', 'central processing unit', and 'microprocessor'. This transparency is critical for debugging retrieval failures and building trust in high-stakes enterprise search.
RETRIEVAL PARADIGM COMPARISON
SPLADE vs. Dense Retrieval vs. BM25
A technical comparison of three distinct retrieval paradigms: learned sparse expansion (SPLADE), dense vector similarity, and traditional lexical matching (BM25).
| Feature | SPLADE | Dense Retrieval | BM25 |
|---|---|---|---|
Representation Type | Learned sparse high-dimensional vectors | Dense low-dimensional embeddings | Sparse high-dimensional vectors |
Vocabulary Dependence | Expands beyond exact terms via MLM | No explicit term mapping | Exact term matching only |
Query-Document Matching | Sparse dot product | Cosine similarity or dot product | TF-IDF weighted exact match |
Handles Lexical Gap | |||
Handles Exact Match | |||
Training Required | |||
Index Size | Large (sparse inverted index) | Compact (dense vector index) | Compact (sparse inverted index) |
Inference Latency | Moderate (sparse dot product) | Low (ANN search) | Very low (inverted index) |
Explainability | High (term weights visible) | Low (black-box embeddings) | High (exact term matches) |
Out-of-Vocabulary Robustness | High (subword expansion) | High (semantic generalization) | None (exact match only) |
Typical Recall@1000 (BEIR) | 0.48-0.52 | 0.45-0.50 | 0.30-0.35 |
Enabling Efficiency, Speed & Accuracy
Intelligent Analysis, Decision & Execution
We build AI systems for teams that need search across company data, workflow automation across tools, or AI features inside products and internal software.
Talk to Us
Search across company data
Give teams answers from docs, tickets, runbooks, and product data with sources and permissions.
Useful when people spend too long searching or get different answers from different systems.
Enterprise searchRAGPermissions
Read more
Automate internal workflows
Use AI to route work, draft outputs, trigger actions, and keep approvals and logs in place.
Useful when repetitive work moves across multiple tools and teams.
AI agentsWorkflow automationGovernance
Read more
Add AI to products and internal tools
Build assistants, guided actions, or decision support into the software your team or customers already use.
Useful when AI needs to be part of the product, not a separate tool.
AI integrationDecision supportModel routing
Read moreSPLADE DEEP DIVE
Frequently Asked Questions
Explore the mechanics and advantages of SPLADE, the sparse lexical and expansion model that bridges the gap between traditional keyword search and modern neural retrieval.
SPLADE (Sparse Lexical and Expansion Model) is a learned sparse retrieval architecture that uses a pre-trained language model, typically BERT, to predict term importance for both queries and documents. Unlike dense retrieval which maps text to uninterpretable vectors, SPLADE generates high-dimensional but sparse weight vectors in the vocabulary space. It works by feeding text through a masked language model head, applying a log-saturation activation to prevent term dominance, and using FLOPS regularization during training to enforce sparsity. The result is a model that performs automatic query and document expansion—adding semantically relevant terms like synonyms—while maintaining the efficiency of an inverted index for retrieval.
About the author
Prasad Kumkar
CEO & MD, Inference Systems
Prasad Kumkar is the CEO & MD of Inference Systems and writes about AI systems architecture, LLM infrastructure, model serving, evaluation, and production deployment. Over 5+ years, he has worked across computer vision models, L5 autonomous vehicle systems, and LLM research, with a focus on taking complex AI ideas into real-world engineering systems.
His work and writing cover AI systems, large language models, AI agents, multimodal systems, autonomous systems, inference optimization, RAG, evaluation, and production AI engineering.
LinkedIn
Limited slotsGet a Free AI Consultation
Partnered with leading AI, data, and software stack.
How We Work
Custom AI workflows for your Business
One-fit-all AI don't work for modern businesses. At Inferensys, we aim to understand your business & custom requirements; which we use to define most efficient agentic workflows, the data, and the tools for your business.
01
Review the use case
We understand the task, the users, and where AI can actually help.
Read more02
Pick the right approach
We define what needs search, automation, or product integration.
Read more03
Build the first useful version
We implement the part that proves the value first.
Read more04
Improve from there
We add the checks and visibility needed to keep it useful.
Read moreThe first call is a practical review of your use case and the right next step.
Talk to Us