Knowledge graph for AI agents: entities, edges, citations
A knowledge graph for AI agents gives answers sources: entities, relationships and citations, shown on a synthetic family office with forty entities.
Why AI agents need a knowledge graph#
An AI agent needs a knowledge graph because the questions people ask it are about things and the links between them: a trust, the LLC it owns, the fund that LLC committed to, the person who signs for it. A pile of retrieved text stores the words but drops identity, relationships and where each fact came from. A graph keeps all three, so the agent can follow the links to an answer and cite the record behind every step.
The problem with a pile of text#
Most agents today answer from retrieval-augmented generation (RAG). The idea goes back to a 2020 paper by Patrick Lewis and colleagues, which paired a language model with "a dense vector index of Wikipedia, accessed with a pre-trained neural retriever." The authors framed it as a way to tackle two open problems for models that rely on their weights alone: "providing provenance for their decisions" and "updating their world knowledge."
RAG works well when one passage holds the answer. It struggles when the answer is spread across several documents. Microsoft Research named the problem in its February 2024 GraphRAG post: "Baseline RAG struggles to connect the dots. This happens when answering a question requires traversing disparate pieces of information." The same post adds that baseline RAG "performs poorly when being asked to holistically understand summarized semantic concepts over large data collections."
Stuffing more text into the prompt does not fix it. Nelson Liu and colleagues found that model "performance is often highest when relevant information occurs at the beginning or end of the input context, and significantly degrades when models must access relevant information in the middle of long contexts." A bigger haystack gives the model more places to lose the needle.
Entities, relationships and citations#
A knowledge graph stores three things that plain text leaves implicit.
| What the agent needs | A folder or vector store | A knowledge graph |
|---|---|---|
| Identity | "Dana Ito", "dana.ito" and "D. Ito" are three unrelated strings | One node for the person, whatever the spelling |
| Relationships | Two facts look related if they sit near each other in a file or in embedding space | Typed edges: owned by, signed by, committed to, attended by |
| Provenance | A chunk, maybe with a filename | An id that resolves to the exact source record |
The table follows Mitosis Labs' developer docs, which describe the gap this way: "the text is stored, but nothing about the text is stored." Add more documents to a vector store and identity gets worse, because you get more spellings and more near-duplicate chunks competing for the same slots in a result set.
Citations follow from provenance. When every node and edge points back to the email, document or record it came from, an answer built from them can carry those pointers. Microsoft's team made the same point about their system: "GraphRAG provides the provenance, or source grounding information, as it generates each response."
An illustration: a family office with forty entities#
The family office below is invented for this post. The names, entities and numbers do not describe any real family or client.
Picture the Varrow family office. Three generations, forty legal entities: a dynasty trust and four grantor trusts, eleven holding LLCs, six limited partnerships for direct investments, a private foundation, an operating company, and commitments to fourteen outside funds. Two trustees, an outside counsel, a CFO and a controller sign documents for different subsets of those entities.
The CFO asks an agent: "Which of our entities owe a capital call to Fund IV this quarter, and who has to sign for each one?"
With plain RAG, the agent searches for text that looks like the question. It finds the Fund IV capital call notice, maybe the subscription agreement, maybe an email thread. The notice names "Varrow Ventures II LLC". It does not say that this LLC is owned 60/40 by two grantor trusts, or that one trustee resigned in March and a successor signs now. Those facts live in an operating agreement, a trust amendment and a board resolution, none of which uses the words "capital call" or "Fund IV". The agent never retrieves them, so it either guesses or stops short.
With a graph, the same question becomes a walk:
- Start at the node for Fund IV.
- Follow the "committed to" edges back to the entities that hold commitments.
- From each entity, follow "owned by" edges to the trusts above it.
- From each entity and trust, follow "authorized signatory" edges, filtered to the people whose authority is current.
- Attach the source record for every edge used: the subscription agreement, the operating agreement, the trust amendment, the resolution.
The answer names three entities, the signer for each, and links to the four documents that establish it. The CFO can check any line in seconds. If the March amendment is wrong, she can see that the answer depends on it.
That last part matters as much as the answer. In a family office, the cost of a wrong answer is a missed call, a late wire or a document signed by someone without authority. An answer you can trace back to its source is one your team can rely on, and one an auditor or outside counsel can verify.
How a graph gets built#
Two broad approaches exist, and many systems mix them.
Microsoft's GraphRAG uses a language model to build the graph. In the team's words, "The LLM processes the entire private dataset, creating references to all entities and relationships within the source data, which are then used to create an LLM-generated knowledge graph." The paper by Darren Edge and colleagues then clusters related entities and pregenerates "community summaries for all groups of closely related entities." On global questions over large text collections, the authors report "substantial improvements over a conventional RAG baseline for both the comprehensiveness and diversity of generated answers."
Mitosis, the knowledge graph Mitosis Labs builds for AI agents, does the expensive work once, when data arrives. Its docs describe each incoming item being embedded, with typed entities and edges extracted from it and written to the graph "with a record of how each one was derived." Extraction rules normalize values such as emails, phone numbers and URLs, so two spellings of the same person land on one node. Each edge also records how it was produced, and the docs rank those methods: "A declared field outranks an exact lexical match, which outranks a geometric similarity, which outranks an LLM extraction." Retrieval then combines vector search, full-text search and a one-hop graph expansion, and returns scored, cited results. The docs state that "There is no LLM at query time": your agent gets the evidence and does the reasoning itself.
The two designs answer different questions. GraphRAG's community summaries suit "what are the main themes across this corpus?" A derivation-tagged graph suits "who signs for this entity today, and which document says so?" For an agent that acts on a family's affairs, the second kind of question comes up every day.
What the published numbers show#
Mitosis Labs publishes its benchmark results at mitosislabs.ai/research. The page calls the product Mitosis Cortex, and it reports an independent evaluation, the Agentic Memory Index, run by Verging Labs: "Verging Labs ran 8 agent memory systems through 2,176 tasks across 56 simulated working sessions. We did not run it, fund it, or see the questions. Cortex finished first among products, second overall, and answered nothing incorrectly." The page lists "0.0% Incorrect answers across 272 tasks" for Cortex and dates those figures to Verging Labs' revision of 3 August 2026.
The same page shows a self-reported LongMemEval result of 91.7%, marked "prelim.", and says to read it differently because Mitosis Labs ran it. Read the method section before you rely on either figure. The page publishes the protocol so you can argue with it.
Where this leads#
If your agent works across many entities, people and documents, the graph under it decides whether its answers carry sources you can check. Private SuperIntelligence, the AI concierge Mitosis Labs runs for family offices, enterprises and private clients, runs on Mitosis. It works across a family's own email, documents and data and answers with citations, and each deployment is built for one client.
Frequently asked questions#
What is a knowledge graph for AI agents?#
It is a store of the entities in your data (people, companies, trusts, funds, documents) and the typed relationships between them, with each fact linked to its source. An agent queries it to find connected facts and cite where each one came from.
What is GraphRAG?#
GraphRAG is Microsoft Research's approach to retrieval-augmented generation over a knowledge graph. A language model extracts entities and relationships from your documents, groups related entities into communities, and summarizes each community so the system can answer questions about a whole dataset.
Does a knowledge graph replace a vector database?#
No. Many systems use both. Vector search finds text that resembles the question, and the graph adds identity, relationships and provenance. Mitosis, for example, fuses vector search, full-text search and a one-hop graph expansion in one query.
Why does a family office need a graph more than most organizations?#
A family office holds many legal entities with layered ownership and different signers, and the facts that connect them sit in separate documents. Questions about obligations or authority need those links followed, which plain text search does not do.
Do citations stop an AI agent from making things up?#
Citations let a person check every claim against its source, so a wrong answer gets caught before anyone acts on it. They make errors visible. You still need to read the source when the stakes are high.
Sources#
- Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (Lewis et al., arXiv:2005.11401, 2020)
- GraphRAG: Unlocking LLM discovery on narrative private data (Microsoft Research blog, 2024-02-13)
- From Local to Global: A Graph RAG Approach to Query-Focused Summarization (Edge et al., arXiv:2404.16130, 2024, revised 2025)
- Lost in the Middle: How Language Models Use Long Contexts (Liu et al., arXiv:2307.03172, 2023)
- What Cortex is (Mitosis Labs developer docs, accessed 2026-10-15)
- Why not files (Mitosis Labs developer docs, accessed 2026-10-15)
- Agent memory research and benchmarks (Mitosis Labs, last updated 2026-08-03, accessed 2026-10-15)
- Private SuperIntelligence for family offices (Private SuperIntelligence, accessed 2026-10-15)