AI knowledge graphs: the arrow is a claim too
Review AI-extracted knowledge graph relationships for direction, meaning, validity period and supporting source. Correct names alone are not enough.
Article language: English

In this article
In a diagram, an arrow labelled "manufactures" connects the company Linea to the Noma panel. The document says Linea uses it when building exhibition stands. The names are right, the panel exists, and so does the company. The arrow has simply promoted a user to manufacturer. That is a rather generous way to expand a business.
This fictional example shows a basic challenge when reviewing an AI-assisted knowledge graph. A graph records objects and their connections, which can later be used to find answers. A relationship can be wrong even when every object has been identified correctly. Checking the list of names is therefore not enough. Each meaningful connection needs to be read too.
Read the arrow as a complete sentence
Instead of starting with an attractive visual, write the result in ordinary language: "Linea manufactures the Noma panel." Identify who performs the action, the type of connection and the other object. Then find the passage in the source that actually supports the whole sentence. Highlighting both names in the same paragraph is the beginning of the check, not its end.
The RDF 1.1 Concepts specification describes a triple as its basic structure: subject, predicate and object. The predicate, the middle element, identifies a relationship between resources. This is one specific data model; not every company graph needs to use RDF. For review, however, a simple question helps: who has what relationship with whom?
In our sample, the answer is "Linea uses the Noma panel." Reversing the direction would give "Noma uses Linea", while changing the type to "manufactures" would create a different claim. Both errors are more than a visual problem. They can change a supplier list, a parts search result or a description of a partner's capabilities. The recorded relationship should mean the same as the accepted sentence.
First confirm the objects' identities. Two companies may have similar names, or a product may be confused with a product range. Checking false record matches made by AI (in Polish) explores this step. Only assess the arrow once the objects have been distinguished. The right relationship type attached to the wrong company still produces false information.
Distinguish the product, the individual item and the project. Using one panel at one event does not mean the company regularly works with the whole range. In the sample, describe the scope of the claim: which product, in what role and for which task. Without this, a correct arrow may be extended to every project simply because the data model does not allow a more specific record.
For the test, retain each relationship's identifier, both endpoints, its type name and a reference to the document. Read the result from the data used by the application, not only from the label on a diagram. The visual layer may shorten names or combine several connections into one line. A reviewer should be able to trace a customer's question to a specific record and its basis.
Four documents with the same names
Prepare a short test set using fictional data. Let Linea and Noma appear in every document, but in a different relationship each time:
- Linea uses the Noma panel when building stands. We expect a use relationship, without adding manufacture or ownership.
- Linea is considering using the Noma panel next season. This is an intention, not confirmed current use.
- Linea does not manufacture the Noma panel; it orders it from a named supplier. Negation must not become confirmation of manufacture.
- Until the end of last season, Linea used Noma, then chose another range. A historical relationship should not answer a question about current equipment without qualification.
Agree on the expected result before running extraction, the process of drawing information from text. If the graph schema cannot record intention, negation or a period, decide what happens to such a passage. It can be sent for a human decision. It should not be automatically forced into the nearest available relationship simply because the application has to draw something.
Also add a sentence reporting someone else's words: "At the meeting, the client asked whether Linea manufactures Noma." A question does not establish a relationship, even though it contains exactly the words a simple mechanism looks for. Similarly, test an old proposal quoted in a new document. The file's current date does not turn an old assumption into a current fact.
Do not judge only the number of connections found. A result enriched with three unsupported arrows may look impressive but make answers worse. In our set, distinguish a missing required relationship from an added relationship with no basis. Record cases where the tool correctly withheld a decision too. An empty list and a deliberate finding that there is insufficient evidence need different descriptions in the report.
Include material where the names appear next to each other without a shared relationship: an exhibitor list names Linea, while a separate box presents Noma. Proximity on a page does not prove product use. If text is extracted from a PDF, also examine the order in which columns are read. Joining two independent blocks can create a sentence the document's author never wrote.
The schema organizes types, while the source supports content
The Neo4j GraphRAG Python documentation describes guiding extraction with node and relationship types and patterns that connect a source object to a target. This is a useful way to define the expected structure. The permitted pattern "company uses product" does not itself prove that a particular company actually uses that product.
Build your relationship vocabulary with people who know the material. "Supplies", "sells", "manufactures" and "services" may concern the same pair of parties, but they are not interchangeable. Add a short definition and an example excluded by it to each term. If a relationship means "supplies for this project", do not later shorten it to the unconditional "is the company's supplier".
Check the direction recorded in the data too. Changing labels in the diagram can make it readable while a query still travels from the wrong object. Add questions in both directions to the test: "Which panels does Linea use?" and "Who uses Noma?" The expected answers should come from the same correct relationship, not two separately added demonstration answers.

Assign the supporting source to the relationship, not just to the whole object. A document about a company may support its address and sector but not its role with a particular product. Record the passage, its location and the version of the material. This extends knowledge architecture in a company (in Polish): the source owner and its status still matter after information moves into a graph.
Aeneia Audit
Prepare a few relationships that are supposed to affect answers or the team's work, together with their source documents. Such a sample allows the direction, meaning and basis of AI-extracted connections to be assessed. Check a sample of knowledge graph relationships
If one connection has several sources, distinguish independent confirmation from successive copies of the same report. Three documents may repeat one unverified description. Their number does not automatically increase confidence in the relationship. Show in the record whether a new basis was added or merely another occurrence of the same information. Do not discard provenance when combining similar records.
Decide how to respond when a document is deleted or withdrawn. The disappearance of a file alone does not prove that the described collaboration never existed. It may, however, prevent the team from checking the basis. In the report, distinguish a relationship challenged after evaluating the content from one whose source has become unavailable. Leaving both labelled "current" makes a later discussion of answer reliability harder.
Time should describe the relationship, not the run date
The arrow's creation date says when the system recorded it. The document's publication date says when the material was made available. The period of the described collaboration is a third piece of information. When a source describes a project from two years ago, importing it today does not make it a current contract. Mark these three moments separately in the test, even if the user sees only a short explanation on screen.
Create a case involving an ended relationship. An old document describes use of Noma; a newer note confirms a change of equipment. Check the answer to a historical question and one about the current situation. Replacing the old arrow with a new one can erase needed history, while leaving both without periods can suggest simultaneous use of two ranges. The rule depends on what the graph is intended to do.
Record exactly what "currently" means in the user's question. When reconstructing an old offer, the reference point may be its preparation date; for a new assignment, it may be the day of the answer. The same historical data can produce different correct results depending on that reference point. The system should disclose the point it has adopted instead of silently equating every question with the last import.
Add a relationship with an unknown end date. Missing information about an end does not by itself confirm that the collaboration continues. The team should decide how to mark this state and when to ask for the material to be checked. "Current status not confirmed" is more useful than arbitrarily extending the relationship to today. Do not substitute the last index refresh date for this decision.
Keep contradictions between sources visible for assessment. A newer document may matter more for a particular task, but recency alone does not make it correct. Check the author, scope and status of the document. It may describe another branch or only one project. The decision should state why a relationship was accepted, limited or held for confirmation.
Check an arrow's error in the answer that uses it
Correcting the graph is not enough if the application still answers from an earlier result. Prepare a question about Noma's manufacturer before and after the correction. In the second run, we expect an answer consistent with the available material, not a fluent repetition of the incorrect association. For RAG answers (in Polish), also check which passage actually reached the model.
See whether the relationship was duplicated during import. Two identical descriptions of one panel's use do not mean two projects or two customers. Counting connections requires a clear unit. A similar trap appears in the difference between the number of rows and the number of cases in SQL (in Polish). A query result can be technically correct while answering a different question.
Another case is an inference that goes too far. Linea works with a company called Oś, and Oś uses Noma. Those two statements do not automatically mean Linea also uses the panel. If a solution derives additional connections, separate them from relationships read directly from the source. A reviewer should be able to tell which statement has direct support and which was produced by an inference rule.
When correcting one relationship, check dependent outputs. An incorrect attribution of manufacture may already have entered a supplier list or company description. Fixing the arrow does not confirm that those materials have been refreshed. Choose one such output and establish how it will be recalculated or sent for correction. This helps assess the work needed to repair one error without expanding the test to every use of the graph.
For testing, you can prepare synthetic data with consistent relationships and events (in Polish). Label it clearly and do not mix it with the company's database of confirmed partners. A fictional sample tests the tool's behaviour; it does not confirm a real collaboration. Keep the expected decisions outside the material the model may consult during extraction.
Finally, choose one important relationship and trace the route from document to user decision. Can you identify the parties, read the meaning, assess the period and open the supporting source? Does a lack of evidence remain visible? Such a limited scope can be discussed through Aeneia Audit. The test ends with the claim matching the material, not with the number of colourful connections on screen.








