Glossary · AI & automation
Knowledge graph meaning
A knowledge graph stores facts as entities joined by named relationships. See how it works, an illustrative lending example and how it differs from RAG.
How does a knowledge graph work?
Every fact is written as a small statement with three parts: a subject, a relationship and an object. Using a made-up firm, "Ramesh Traders has GSTIN X", "Ramesh is a director of Ramesh Traders" and "Ramesh Traders pays rent into account Y" are three such statements. The W3C calls these triples. Each entity becomes a node, each relationship becomes a labelled edge, and a question becomes a walk along the edges: which other companies share a director with this borrower, or which applications list the same mobile number.
The value is in the relationships, not the storage. A spreadsheet can hold the same names and numbers, but the connection between row 14 in one sheet and row 203 in another lives only in an analyst's head.
Where do knowledge graphs show up in Indian lending and operations?
Credit and risk teams use the idea, often without the name, when they connect a borrower to its promoters, group companies, guarantors, GSTINs, bank accounts and addresses. Company master data from the MCA portal, GST registrations and bank statements each describe part of the picture. Linking them shows group exposure, circular transactions between related firms, or a guarantor who already backs several other loans. Operations teams use the same structure to link vehicles, drivers, trips and invoices.
A graph is only one way to keep evidence traceable. AutoCredit, credit appraisal software for NBFCs, drafts a memo in which every line cites the document it came from, so a credit officer can follow a figure back to the statement or return behind it. The same discipline applies to any memo, as the credit appraisal memo format guide explains.
What does a knowledge graph catch in a worked example?
Illustrative example, not customer data. A lender receives three small business loan applications in one week from three different firms. Read one by one, each looks clean. Placed in a graph, two links appear: all three list the same contact mobile number, and two of them share a director. The graph has five nodes that matter (three firms, one director, one phone number) and five edges between them. The finding is not proof of fraud. It is a reason for the credit officer to ask why, before sanctioning three loans that may really be one exposure.
How is a knowledge graph different from a database, vector search or RAG?
| Approach | What it is good at |
|---|---|
| Relational table | Fixed records with known columns, such as a loan ledger |
| Knowledge graph | Explicit, named links between entities that you can follow several hops deep |
| Vector embeddings | Finding text that means something similar, even with different words |
| Retrieval-augmented generation | Feeding retrieved passages, from any of the above, to a language model so its answer is based on them |
They are not rivals. A system can use a graph to find related entities and vector search to find the relevant paragraphs, then pass both to a model.
What mistakes do teams make with knowledge graphs?
- Skipping entity resolution. "R. K. Sharma" and "Rajesh Kumar Sharma" become two nodes unless someone decides, with evidence, that they are one person. Bad merges create false links.
- Treating a link as a verdict. A shared address may be a family business or a co-working space. The graph raises a question; a person answers it.
- Forgetting dates. A director who resigned three years ago is not a current link. Relationships need start and end dates.
- Loading personal data without a purpose. Linking more fields than the decision needs adds privacy risk under the DPDP Act. Confirm what you may hold with your legal adviser.
For the limits on what software should decide inside a loan file, see what AI should never do in a loan file.
Related terms
Read next
Where this shows up in our work. AutoCredit: Statements, GST, bureau and AA into one file; a memo against your policy bands; same-day field verification. Book a 30-minute working session with an engineer.
Questions
What is a knowledge graph in simple words?
A knowledge graph is a map of facts in which things such as people, companies and accounts are dots, and the relationships between them are labelled lines. You answer questions by following the lines from one dot to the next.
Is a knowledge graph the same as a graph database?
Not quite. A knowledge graph is the model of entities and relationships; a graph database is one kind of software used to store and query it. The same graph can also be kept in ordinary tables with link records.
How is a knowledge graph different from RAG?
A knowledge graph structures facts and their links. RAG is a method of retrieving relevant material and giving it to a language model before it answers. A graph can be one of the sources that RAG retrieves from.