Research notes: knowledge graph construction

This Advanced note supports the runnable source-linked construction guide. For a first mixed-methods assignment, use Your first research project.

Preserve meaning while combining records

Give documents and source-specific observations stable keys. Keep the original passage distinct from an extracted mention, resolved entity, coding decision, and research conclusion. A node named Claim is an ordinary graph record; that label does not activate immutable native assertion semantics.

MERGE supports repeat imports under an explicit key. It does not establish that two names refer to the same person or that two documents are independent sources. Duplicate model calls must not accumulate as independent confirmations.

Store disagreements and new reviews separately instead of overwriting one source field or retaining only the largest confidence score. Record the meaning of a score and its producer; a normalized score is not necessarily a calibrated probability. See recording an inquiry.

Supported construction and retrieval

Use native add_node / add_edge, parameterized Cypher, or the bulk construction API. Framework outputs need explicit translation; the Python facade has no add_graph_documents(). Use Arrow results through execute(...).to_pylist() or to_pandas() as needed. There are no to_dicts() / to_dataframe() facade methods.

For candidate entity matches, use search and an explicit identity rule. Similar names alone do not justify deleting an alias or merging all of its relationships. Original mentions remain useful even when resolved to one entity.

Retain the basis of the result

Keep source locators and original input data alongside reproducible construction code. A URL alone cannot reproduce changed or unavailable content. Use durable projects for working state, portable interchange for transfer, and knowledge contracts when richer native evidence/status semantics are needed.