The research behind connected AI knowledge

Why source context matters

An AI model needs the right material to answer questions about your work. Retrieval finds that material. A knowledge graph records connections that can help guide the search. The studies below test different ways to use these ideas.

From Local to Global

Microsoft Research · 2024-04

The system links named people, places and topics. It then summarizes related groups for questions spanning many documents.

The study reports more comprehensive and diverse answers to global questions than its conventional RAG baseline.

Gigabrain follows bounded source relationships. It does not implement this paper’s full community-summary pipeline.

Microsoft Research publication; no separate funding statement verified.

Original research PDF. Text and figures unchanged. Microsoft Research.

Introducing Contextual Retrieval

Anthropic · 2024-09-19

The system adds context to each passage before indexing. It combines meaning-based search with BM25 keyword scoring, then sorts results again by relevance.

Its strongest reported configuration reduced top-20 retrieval failures from 5.7% to 1.9% on its evaluation.

Gigabrain uses a different lexical and graph pipeline. This result does not measure Gigabrain.

Vendor-authored engineering evaluation.

PDF copy of the web article. Layout adapted; article text and figures retained. Anthropic.

Reducing Hallucinations in Complex Question Answering

NICD, Newcastle University

AI agents searched linked documents to answer 510 selected Wikipedia questions. The team repeated each test three times.

The median fine-grained truthfulness score rose from 35 to 63, an 80% relative increase.

This is not 80% accuracy. A separate coarse measure reported more incorrect answers alongside more correct answers and fewer refusals.

Neo4j funded the study. The authors state that funding did not influence their research.

Original research PDF. Text and figures unchanged. NICD, Newcastle University.

Lost in the Middle

Liu and colleagues · 2023-07-06

Researchers moved relevant facts between the start, middle and end of long inputs. They tested document questions and key-value lookup, which finds the value linked to a given label.

The evaluated models often performed worse when relevant information appeared in the middle of a long context.

The results concern the models studied in 2023. They do not establish every current model’s behavior.

Funding statement not verified in this source review.

Original research PDF. Text and figures unchanged. Liu and colleagues.

How Gigabrain uses connected knowledge

Hybrid retrieval finds exact wording and related meaning. Graph expansion follows connections within your project to find more source passages. Persistent memory keeps your knowledge available between sessions. Citation checks let you inspect the source text.

Research supports exploring these methods. Its results do not measure Gigabrain’s accuracy or time savings. Your agent, sources and task affect the result.

Explore the product mechanisms or see Gigabrain in action.

Put your knowledge to work

The optional Gigabrain Loop Pack adds 24 prepared automations to System. Use your knowledge for daily briefs, meeting records and follow-up drafts. Choose the required tools, permissions and schedule during setup. You review drafts before sending.

See the 24 loops or get Gigabrain System.