Blocking Strategies

Blocking strategies reduce the quadratic comparison space by grouping records that are likely to match into blocks. Only pairs within the same block are compared, dramatically improving performance on large datasets.

All blocking is configured through the pipeline builder.

from reclink.pipeline import ReclinkPipeline

pipeline = (
    ReclinkPipeline.builder()
    .block_exact("last_name")
    .block_phonetic("first_name", algorithm="soundex")
    .compare_string("first_name", metric="jaro_winkler")
    .classify_threshold(0.85)
    .build()
)

---

Available Strategies

Strategy	Type	Best for	Parameters
Exact	Hash-based	Exact field matches	field
Phonetic	Hash-based	Name matching	field, algorithm
Sorted Neighborhood	Window	Ordered data	field, window_size
Q-gram	MinHash	Approximate matching	field, q, threshold
LSH	Locality-sensitive	High-dimensional	field, num_bands, band_size
Canopy	Clustering	Large datasets	field, metric, tight, loose
Trie	Prefix tree	Prefix matching	field, max_frequency
Numeric	Bucketing	Numeric fields	field, bucket_size
Date	Resolution	Date fields	field, resolution
Hybrid	Combination	Complex matching	strategies, mode
Custom	Plugin	Domain-specific	User-defined

---

Exact Blocking

Groups records by identical field values. Fast and precise but no tolerance for typos.

pipeline = ReclinkPipeline.builder().block_exact("last_name")

Phonetic Blocking

Groups records by phonetic encoding of a field. Tolerant of spelling variations.

pipeline = ReclinkPipeline.builder().block_phonetic("last_name", algorithm="soundex")

Supported algorithms: "soundex", "metaphone", "double_metaphone", "nysiis", "caverphone", "cologne_phonetic", "beider_morse", "phonex", "mra", "daitch_mokotoff".

Sorted Neighborhood

Sorts records by a key field and compares within a sliding window.

pipeline = ReclinkPipeline.builder().block_sorted_neighborhood("last_name", window_size=5)

Q-gram Blocking

Uses character q-gram MinHash signatures for approximate blocking.

pipeline = ReclinkPipeline.builder().block_qgram("name", q=2, threshold=0.5)

LSH Blocking

Locality-sensitive hashing for high-dimensional similarity search.

pipeline = ReclinkPipeline.builder().block_lsh("name", num_bands=10, band_size=5)

Hybrid Blocking

Combines multiple strategies with union (any match) or intersection (all must match) semantics.

# Union mode: candidates from ANY strategy (increases recall)
# Intersection mode: candidates from ALL strategies (increases precision)

Hybrid blocking is useful when no single strategy captures all true matches. Union mode is more common; intersection mode is useful when precision is critical.

Date Blocking

Groups records by date field at a configurable resolution.

pipeline = ReclinkPipeline.builder().block_date("birth_date", resolution="year")

Resolutions: "year", "month", "day".

Numeric Blocking

Buckets records by numeric field values.

pipeline = ReclinkPipeline.builder().block_numeric("age", bucket_size=5)