Memory-efficient document retrieval system designed for edge devices and resource-constrained environments
LiteSearch is a lightweight document retrieval system built for scenarios where resources are limited. Unlike traditional info retrieval systems requiring 4-8GB RAM, LiteSearch delivers 74% recall@15 using only 512MB heap space through optimized BM25 lexical search with custom scoring enhancements.
Perfect for Raspberry Pi, old laptops, embedded systems, personal knowledge bases, and cost-optimized cloud deployments.
Current Support : Since this is the alpha release, the focus is on quantifying performance for a single dataset, and hence we only support SQuAD2.0 data for indexing and search. This is a known issue and support to other data types will be extended in further releases.
wget https://github.com/pri1712/LiteSearch/releases/download/v0.1.0-alpha/liteRAG-0.1.0-alpha.jar
java -Xmx512M -jar liteRAG-0.1.0-alpha.jar --mode=write --data={path to squad dataset}
java -Xmx512M -jar liteRAG-0.1.0-alpha.jar --mode=read --data={path to squad dataset}
Modern info retrieval systems are resource-hungry and complex, and involve complex set up pipelines.
LiteSearch uses pure BM25 lexical search with intelligent optimizations:
- 512MB heap space for 100K documents
- ~2GB total for 1M documents (linear scaling)
- Memory-mapped files for efficient I/O
- Compressed inverted index with delta encoding
- Streaming parsers (no full documents in RAM)
- Just Java 17+ - no Python, no databases, no containers
- Single JAR distribution (~15MB)
- No native libraries or GPU requirements
- Works on any platform with JVM
- 30 seconds from download to first search
- No configuration files required (sensible defaults)
- Auto-format detection for common file types
- Single command to index and search
Evaluated on SQuAD 2.0 (11,873 questions):
| TOP_K | Recall@K | Search Time | Memory |
|---|---|---|---|
| 1 | 55.13% | <50ms | 512MB |
| 5 | 70.43% | <70ms | 512MB |
| 10 | 73.05% | <100ms | 512MB |
| 15 | 74.05% | <120ms | 512MB |
Key Highlights:
- 74% recall@15 - competitive with heavier systems
- 70% of answers in top 5 - excellent for most use cases
- MRR of 0.62 - answers typically in positions 1-2
- Sub-100ms latency - fast enough for interactive use
- Consistent memory - no RAM spikes during queries
Beyond vanilla BM25, LiteSearch adds:
- Coverage Boost (1.0-1.3×): Rewards chunks matching more query terms
- Density Boost (1.0-1.2×): Prioritizes higher term concentration
- Proximity Boost (1.0-1.5×): Values terms appearing close together
- Overlap Deduplication: Prevents near-duplicate results
Result: Better ranking quality while maintaining speed.
Evaluation Dataset Details:
- 11,873 test questions
- 100k documents ingested
Detailed Results:
=== EVALUATION COMPARISON TABLE ===
TOP_K | Recall@K | Precision@K | MRR | Avg Rank
-----------|------------|-------------|------------|----------
1 | 55.13% | 55.13% | 0.5513 | 1.00
5 | 70.43% | 70.43% | 0.6134 | 1.37
10 | 73.05% | 73.05% | 0.6170 | 1.60
15 | 74.05% | 74.05% | 0.6178 | 1.74
What This Means:
- 55% of answers found in the #1 result (first chunk)
- 70% of answers found within top 5 results
- 74% of answers found within top 15 results
- Average rank when found: 1.6 (typically in position 1 or 2)
- MRR (Mean Reciprocal Rank): 0.62 - industry-standard metric showing high precision
Comparison to Other Systems on SQuAD:
- Dense Passage Retrieval (DPR): ~78-82% recall@10 (but needs 6GB+ RAM)
- BM25 (Elasticsearch): ~75-80% recall@10 (needs 2-3GB RAM)
- LiteSearch: ~73% recall@10 (only 512MB RAM)
| System | Time to Index 100K Docs | Time to First Query |
|---|---|---|
| LiteSearch | 10 minutes | 2 seconds |
| Elasticsearch | 15-20 minutes | 30 seconds (cluster startup) |
| Custom FAISS | 45-60 minutes | 5-10 seconds (model loading) |
LiteSearch advantages:
- No Docker/container setup
- No configuration files needed
- No model downloads
- Instant cold start
- Single command to index and search
- Raspberry Pi / Edge Devices - Runs on 512MB-1GB RAM
- Personal Knowledge Bases - Search your notes, docs, research papers
- Academic Research - Low-resource NLP experiments
- Cost-Optimized Cloud - Small instances = lower costs
- Small Team Deployments - <1M documents
- Privacy-First Applications - Self-hosted, no external APIs
- Proof-of-Concepts - Fast prototyping and iteration
-
Streaming Parsers
- Process XML/JSON without loading entire file
- Constant memory usage regardless of file size
-
Delta Encoding
- Compress inverted index by storing differences
- 40-60% size reduction vs. raw storage
-
Memory-Mapped Files
- Chunk data accessed via OS page cache
- Let OS handle memory management
-
Lazy Loading
- Metadata loaded on-demand
- Only active chunks in RAM
-
Efficient Data Structures
- Primitive arrays over objects (lower overhead)
- Min-heaps for top-K selection (O(n log k) vs O(n log n))
-
Inverted Index
- O(1) token lookup
- Pre-computed document frequencies
-
Early Termination
- Stop after K best results
- Skip low-scoring candidates
-
Batch Processing
- Group I/O operations
- Reduce disk seeks