Published by Gigabrain Labs · Updated
What is a knowledge graph?
A knowledge graph represents entities and relationships. A project, a person and a decision can be entities. A relationship states something specific, such as which review a project depends on.
In our fictional example, Project Alder depends on the Kestrel access review. The review follows the Monday schedule. The schedule contains an inspection date.
| Connection | Supporting source |
|---|---|
| Alder depends on the review. | project.md |
| The review follows the schedule. | review.md |
| The schedule contains the inspection date. | schedule.md |
A graph database is one storage and query implementation. A knowledge graph is the information model. The terms describe different responsibilities.
Inspect a connected source and its correction
We reproduced the local memory-service demonstration against the System and Full 1.1.0 archives. Hybrid retrieval and graph expansion reached the connected schedule.
After the disposable source changed from 18 October to 25 October, its old receipt failed verification. Maintenance removed the stale path until the fixture supplied updated support.
The refreshed result quoted 25 October. Inspect the source quotations, relationship path and archive identities.
The runner authors this graph as a fictional fixture. The run does not measure extraction by an authoring model or the quality of a generated answer.
Graph reranking, expansion and bounded context traversal are distinct mechanisms. Check which interface your integration uses. The retrieval comparison explains the differences.
Follow the connections behind an answer
Gigabrain’s knowledge graph connects people, projects, decisions and source material through typed relationships. Those connections help your agent follow dependencies and changes. Graph search can also find linked passages beyond the first search results. Each relationship needs supporting evidence. A connected node alone does not prove a claim is true.
A supplier, a project and a changed deadline
| Source statement | Entity / relationship | What the source tells you |
|---|---|---|
| Aster supplies Project North. | Aster → supplies → Project North | Supported by the named project note. |
| The amended delivery date is 18 October. | Amendment → supersedes → original delivery date | Use the amendment and retain the original as history. |
| No warranty duration is recorded. | Warranty duration → unknown | The system should identify missing evidence, not invent a value. |
Go back to the source when something changes
A useful graph lets you move from a relationship back to the source text, distinguish current facts from superseded ones and keep unrelated projects separate. Try the synthetic example, then read the architecture guide.
Tell your agent how two notes relate
A wikilink connects notes. A typed edge says how they relate. The shipped model supports supersedes, contradicts, depends_on, caused, decided_by and relates_to, with their inverse relationships. For example, a dated amendment can supersede an original source note.
supersedes:
- "[[Original delivery note]]"Declare that frontmatter on the amendment note, using the exact source-note title. Source-level supersession marks quotations from the old source as superseded. It does not delete the original. Check the edges with the shipped graph_edges validator before relying on them. The graph cannot decide which conflicting account is true for you.
Find evidence beyond the first results
The expanded search starts with found passages and named people, places or projects. It follows the kinds of links you allow within your chosen project. Each added result includes source text and the path that led to it. Results show who wrote each link, whether you reviewed it and where the search stopped. A link explains why a source appeared. The source must still support the answer.
Keep changes visible
When a source changes, its old evidence no longer passes the source check. Corrections can replace old facts while preserving their history. Combining two named records needs your approval. You can split them again later. Removing a record does not erase copies in backups or Git history.
