Company Brain: Knowledge Graphs and GraphRAG

Company knowledge that supports the work

Company Brain is the name we use for the company context behind agent work. In the current product, Company Knowledge is the place to work with that knowledge. It helps connect selected documents, records, and relationships to the questions and tasks your team brings to Lightbulb.

A second brain is a useful analogy for organizing knowledge so it can be found and reused. For a company, that also means respecting which sources a person can access, keeping context tied to its source, and reviewing important conclusions. Company Brain supports the work; people remain responsible for the decisions.

GraphRAG means graph-based retrieval-augmented generation. Retrieval finds material relevant to a request before an AI model produces a response. A knowledge graph adds entities and relationships to the retrieval process, so connected facts can supply context alongside text passages found through vector search.

Entity Extraction

Represent people, organizations, projects, and other entities from supported sources, with relationships that help connect relevant knowledge.

Connected Context

Follow relationships between entities to gather context across connected documents and records. The result depends on the available sources and graph structure.

Hybrid Search

Blend dense vector retrieval with graph traversal and optional re-ranking. Teams can compare hybrid and single-mode retrieval against a representative, customer-approved evaluation set.

Document Ingestion

Upload supported PDFs, Word documents, spreadsheets, and HTML. The ingestion pipeline can chunk, embed, extract entities, and index content; processing time varies with document size, format, model, and deployment capacity.

Context Windows

Assemble selected passages and graph facts within the model's context budget so the agent can work with relevant material.

Company Knowledge

Organize company sources and context for reuse across supported questions and workflows, with access determined by the applicable permissions.

When a graph helps retrieval

Vector search finds passages with similar meaning. A knowledge graph represents connections, such as a supplier linked to an invoice or a project linked to a policy. A question about which policy applies to a project can benefit from both relevant passages and those relationships. GraphRAG combines these sources of context; retrieval quality still needs to be checked against representative questions and trusted source material.

Graph engineering starts with the sources

Graph engineering is the work of choosing useful entities and relationships, linking them to sources, and checking whether retrieval answers the right questions. Start with a bounded collection of company documents, decide which relationships matter, and check freshness and access requirements. Supported ingestion can prepare text, embeddings, and entities; source quality and configuration shape the resulting knowledge.

Retrieval in Action

When an agent needs context, the retrieval service can return ranked chunks with available provenance metadata. Depending on configuration, queries may be expanded, searched against vector indexes, enriched with graph neighbors, and re-ranked. Source references, scores, and graph paths can help reviewers inspect retrieval, but they do not guarantee factual accuracy or complete explainability.