Preselection / Blocking (new step)

[HamedBabaei]

Problem O(N²) comparisons are infeasible for real ontologies.

Task: Introduce a BlockingPipeline before matching.

What to implement

• SBERT / BM25 top-N candidate selection
• HNSW index (FAISS / hnswlib)
• Configurable candidate size

Pipeline:

blocking → matching → reranking

The blocking goal is not to compare everything to everything. First, reduce the search space, and then do the alignment.

naive approach:

for each source concept:
    compare with every target concept

OntoAligner approach

Step 1: Blocking (cheap filtering)
    reduce 10,000 targets → top 100 candidates

Step 2: Matching (smarter model)
    run alignment only on those 100

Step 3: Reranking (final refinement)
    pick best match from those 100

[HamedBabaei]

Dear @abhishekreddyallu , I need multiple examples of how the reranking strategy that we implemented can be applied independently of the aligner. This is just needed for the possibility of adding such technique. We need to generalize the reranking.