sketchir
Sketching primitives for IR: MinHash, SimHash, and LSH indexes for near-duplicate detection and approximate similarity search.
[]
= "0.3"
Best starting points
- Near-duplicate detection:
MinHashTextLSH+BlockingConfig - SimHash fingerprints:
SimHashFingerprint/SimHashLSH - Dense-vector LSH (batch):
LSHIndex-- multi-table random projection, add/build/search lifecycle - Dense-vector LSH (incremental):
DenseSimHashLSH-- SimHash + Hamming-1 probing, no build step
Tuning knobs (BlockingConfig)
| Param | Typical | Tradeoff |
|---|---|---|
ngram_size |
3-9 chars | Smaller = more sensitive to noise; Larger = stricter. |
num_bands * num_hashes_per_band |
100-256 | More = better Jaccard estimation, higher storage. |
num_bands |
20-50 | Controls recall/precision curve (S-curve). |
Example (MinHash blocking)
use ;
let cfg = default;
let mut index = new.unwrap;
index.insert_text;
index.insert_text;
let pairs = index.candidate_pairs;
assert!;
Examples
Runnable examples live in examples/:
dedup_documentsindexes a document set and returns candidate near-duplicate pairs via MinHash LSH blocking, the corpus-hygiene use (the engine behind a dedup tool like siftr).minhash_jaccard_validationchecks the MinHash Jaccard estimate against the exact Jaccard index, confirming the sketch is accurate before trusting it at scale.
Updatable index (store feature)
store::UpdatableIndex wraps the MinHash near-duplicate index in a durable,
segmented store (segstore): incremental
add/delete, a write-ahead log, checkpoint, compaction, and crash recovery. The
per-segment LSH blocks are cached and rebuilt only on mutation, not per query.
Opt-in; the default build does not depend on segstore.
License
MIT OR Apache-2.0