Why might building a knowledge graph in two passes, extract first and then merge duplicates, beat doing it all at once?

This explores why splitting knowledge-graph construction into an extraction pass and a separate cleanup pass that merges duplicate entities might produce a better graph than extracting everything in one shot. None of the retrieved notes tests that comparison directly. What follows is the closest neighbouring evidence, and it is partly my inference, not a finding from the corpus.

The nearest match is a general pattern: one pass finds candidates, and a second pass decides whether they are really the same thing. Can verification separate structural near-misses from topical matches? describes a pipeline with cheap, broad recall followed by a small learned verifier. The verifier looks at full token-by-token similarity patterns, so it can reject structural near-misses that only look topically alike. Deduplication in a graph has the same shape. 'Same topic' is not 'same entity', and a broad extraction pass can't reliably tell them apart. If the first pass is allowed to be generous and a second pass owns the question of identity, duplicates get merged without also merging things that merely sound alike.

The extraction side helps explain why one pass isn't enough. Can smaller language models outperform larger ones at graph extraction? argues that graph extraction is mostly a reading task that depends on language comprehension, not remembered facts. A domain-adapted 7B model beat a 32B one at it. A model reading one chunk at a time has no way to know that 'the company' in one chunk is the 'Acme' named in another. Reconciling across chunks is a different job from reading, and it can be handed to a different step or a different, cheaper model.

There is also a noise argument. Can graph structure patterns outperform direct edge signals in noisy data? builds product-substitute graphs from patterns across several edges instead of single edges, because several independent noisy signals rarely line up by accident. A second pass works the same way in spirit. An entity or relation that survives separate extractions and merging is less likely to be a one-off hallucination. This is an analogy from a different domain, not evidence about LLM extraction.

The stakes are downstream. Can community detection enable RAG systems to answer global corpus questions? partitions the entity graph into communities, and Can knowledge graphs enable multi-hop reasoning in one retrieval step? walks it to find multi-hop paths. Both assume each node means one thing. My reading is that duplicate nodes would split a topic across communities or dilute the paths a walk can follow, though the notes don't measure this. Can query-time graph construction replace pre-built knowledge graphs? takes the opposite route and builds a small graph at query time, which avoids corpus-wide extraction and cleanup costs. If you want a direct answer on dedup, the corpus needs a note on entity resolution in graph construction.