Knowledge graph vs document store

A knowledge graph represents subjects, their properties, and their relationships in a form that supports traversal and structured questions. It is useful when a question crosses several entity types or requires explicit relationship semantics. A graph can also integrate assertions from multiple sources. Its nodes and edges, however, are interpretations of source material: a relation in the graph needs provenance, scope, and time if it will support a consequential conclusion. A well-formed graph is not evidence by itself.

A document store retains documents and associated metadata as the main unit of access. It is appropriate when the document is the authoritative artifact, users need to read full context, or structure varies too much for early modeling. Search can make documents findable without predefining every relation. Yet repeated cross-document questions may require manual interpretation, and a document-level citation can be too broad to support a specific assertion. The choice concerns the unit of query and review, not a universal ranking of storage technologies.

A fictional analyst asks which companies in a collection supply a given reagent to hospitals that also run a certain trial. In a document store, the analyst searches, opens many documents and assembles the answer by hand. Each document is read in its own context, which is a strength when the wording matters.

In a knowledge graph, supplier, product, hospital and trial are nodes, and supplies, runs and located-in are edges. The question becomes a path query. The graph answers quickly, and it is only as good as the extraction that created each edge. A reviewer who doubts one edge needs to reach the document and the sentence behind it. A graph that kept no link to the source is a map without a legend, and a document store without structure is a library without a catalog. Both are needed for different moments of the same investigation.

How often are questions of the same shape asked? Do the questions cross several entity types? Is the unit of review the document or the statement? Who will govern identifiers and relation meanings? How will the graph be kept in step with changing documents? A modest practice is to start with a document store, find the questions that recur, and model only those relations in a graph, each with a pointer to the supporting passage.

The decision is easiest to make with a small pilot. Take ten questions that people actually ask, answer them from the documents, and note how long each took and where the answer was ambiguous. Then model only the entities and relations those questions need, extract them with citations, and answer the same questions from the graph. Compare the answers, the time, and the number of questions where the graph answer needed a trip back to the source. If the graph is faster and the trips back are rare, the modelling effort is justified for that question family. If not, the documents and a good search are the better investment.