Knowledge and connection graphs
How a pile of records and documents becomes a connected map of people, companies, and deals.
A knowledge graph stores information as things and the connections between them, not as rows in a table or loose paragraphs of text. Google put the idea in one line that still lands: things, not strings.
Jump to the source papers →Things, and how they connect.
Nodes are the things. Edges are the connections. The value lives in the edges.
Each thing is a node: a person, a company, a deal, a product, a place. Each connection is a labeled link: works at, attended, reports to, decided on. The atomic unit is a triple, written as a subject, a relation, and an object.
(Maria Lopez) works at (Acme)
(Acme) evaluating (Premium plan)
A light schema or ontology says what kinds of things and links are allowed, so the graph stays consistent. Compare that with the two structures you already know. A table stores the attributes of one kind of row but hides the cross-connections. Free text holds everything and structures nothing. A graph makes the connections explicit and queryable, so you can follow a path across many hops. The comprehensive map of the field, from data models and schema through construction and quality, is the Hogan et al. survey.
| Subject | Relationship | Object |
|---|---|---|
| Maria Lopez | works at | Acme |
| Acme | evaluating | Premium plan |
| Marcus Chen | approves | Renewal |
The world is messier than the map.
Real data is noisy, duplicated, and incomplete. Turning it into a clean graph is the whole challenge.
The same thing wears many names. "Bob Smith," "Robert Smith," and "R. Smith at Acme" may be one person, or three. Records arrive from many systems in many shapes, and deciding what is actually the same real-world thing is the central difficulty. This is the problem of entity resolution, and the end-to-end view at scale is the Christophides et al. survey.
Scale rules out the naive approach. Comparing every record with every other is quadratic. Real systems use blocking and indexing to compare only plausible candidates.
Words are ambiguous. A mention like "Apple," "Jordan," or "Mercury" can point to many things. Picking the right one needs the surrounding context.
Graphs are always incomplete. Most true facts are never written down, so a graph built only from what was stated will have gaps. And the world keeps moving: people change jobs, deals change stage, companies merge. A graph that is right today drifts tomorrow unless it is kept current.
Bob Smith · Robert Smith · R. Smith at Acme
- Check
- Compare identifiers and contextual evidence.
- Link
- Attach confirmed mentions to the same entity ID.
- Keep separate
- Preserve unresolved candidates rather than forcing a merge.
Five stages turn mess into structure.
You clicked into Research, so here are the real names. Each stage maps to a paper in the strip above.
Knowledge graph construction
Pull entities and relationships out of structured sources like a database or CRM, and unstructured ones like documents, email, and transcripts, using information extraction (named-entity recognition and relation extraction). Then map them onto the schema and write the triples.
named-entity recognition · relation extraction · schema mapping · Hogan et al.Entity resolution
Also called entity matching, deduplication, or record linkage. The pipeline runs in steps: blocking groups likely-matching candidates so you avoid all-pairs comparison, comparison scores similarity across attributes, matching decides same versus different, and clustering collapses a group of records into one canonical entity. This is the literal "many messy records become one clean node" step.
blocking · comparison · matching · clustering · Christophides et al.Entity linking
Take a mention inside a piece of text and connect it to the right canonical node in the graph. The stages are mention detection, candidate generation, and disambiguation using context. Modern systems encode the mention, its context, and each candidate as vectors and pick the closest match.
The clean distinction for readers: entity resolution reconciles record to record, while entity linking reconciles a mention-in-text to a node. Both answer "which thing is this, really."
mention detection · candidate generation · disambiguation · Sevgili et al.Link prediction
Also called knowledge graph completion: infer the missing edges from the structure that already exists. Knowledge graph embeddings place every entity and relation as a vector so that plausible triples score high and implausible ones score low. The seminal, easy-to-explain example is TransE, which models a relation as a translation in vector space, so head plus relation should land near tail. If two contacts share many of the same connections, the graph can propose a likely-but-unstated link. These are suggestions, worth confirming, never certainties.
embeddings · translation in vector space · Bordes et al. (TransE) · Ji et al.Graph-augmented retrieval
Use the graph to ground an answer. Instead of retrieving only loose chunks of text, retrieve the relevant entities, their neighborhood, and summaries of related clusters, then let a language model answer from that. This is what lets a system connect the dots across many documents and answer broad questions, while keeping the answer traceable back to specific nodes and edges. It pairs naturally with retrieval-augmented generation.
graph index · community summaries · grounded answers · Edge et al. (GraphRAG)Apple announced a new laptop.
| Candidate | Fit to this context |
|---|---|
| Apple, the fruit | Does not fit the product announcement |
| Apple Inc. | Fits the company and product context |
| Known relationship | Evidence |
|---|---|
| Maria → Acme | Verified contact record |
| Marcus → Acme | Verified contact record |
| Maria → same renewal as Marcus | Candidate relationship only |
Shared neighbors can suggest a link, but do not establish it.
Why the structure is worth the work.
Once the connections are explicit, a whole class of questions becomes answerable.
Connect-the-dots answers
Multi-hop questions, like who at this account connects to that deal, become answerable because the path exists in the graph.
One clean view of each thing
Deduplicated and disambiguated, so "which Bob is this" stops being a question.
Grounded answers with their sources attached
Because every claim traces to a node or an edge, an answer can show where it came from, which reduces unsupported guessing.
Surfacing likely-missing facts
Link prediction proposes connections worth confirming.
A structure you can govern
Permissions can sit on nodes and edges, and separate graphs can be kept partitioned, so access and isolation are properties of the structure rather than afterthoughts.
Who approves Acme’s renewal?
- Acme
Has renewal R-482.
- Renewal R-482
Approval relationship points to Marcus Chen.
- Source
The relationship cites the current approval document.
An honest forward look.
The field is moving fast, and the labs are candid about what is still maturing.
Tighter language-model and graph loops. Models help build and query graphs, and graphs ground models. The graph-augmented retrieval line of work is recent, and the authors are explicit that the evaluation frameworks are still maturing. That candor reads as credible: see the closing discussion in the GraphRAG paper.
Time-aware graphs. Facts that carry a validity period, so the graph can reason about change instead of pretending today is forever.
Richer extraction and schema learning. Multimodal entities, and neuro-symbolic reasoning that combines learned embeddings with explicit rules. For a structured tour of representation learning and knowledge acquisition, the Ji et al. survey goes deeper than TransE.
Privacy-aware, access-controlled retrieval. Where the relevant subgraph respects who is allowed to see what, so access is part of the query rather than a filter bolted on after.
A seller's world is a connection problem.
We build a voice co-pilot, so this field is close to home. Here is what it means for the work, at the level of what a rep actually feels.
The same contact shows up across email, the calendar, the CRM, and past calls under slightly different names. Every account is a web of people and threads. The outcomes this research points at are the outcomes a strong co-pilot is judged on:
A clean, deduplicated view of each account
So the record is trustworthy and you are never guessing which "Bob" you are looking at. Clean CRM.
Answers that show their sources
So a rep can trust them at a glance instead of taking them on faith. Deep context.
The whole account legible at once
The threads connected instead of scattered across a dozen places.
Each client's information kept separate
One account's picture never mixes with another's, the way two partitioned graphs stay apart.
When a co-pilot keeps your accounts this clean and answers with sources, the admin stops being your job. Prism is our own model, built for sales conversations.
- Maria Lopez
- Works at Acme. Source: verified contact record.
- Premium plan
- Under evaluation. Source: latest meeting note.
- Renewal
- June target. Source: current renewal record.
- Marcus Chen
- Approver. Source: approval document.
The papers behind this page.
The readable map of the whole field, from data models and schema through construction, refinement, quality, and embeddings. A natural starting point for "what is a knowledge graph."
The end-to-end view of entity resolution at scale: the blocking, comparison, matching, and clustering steps that turn many messy records into one clean entity.
Systematizes how modern systems link a mention in text to the right canonical entity: the candidate-generation, context-encoding, and disambiguation pipeline at the heart of "which thing is this."
The seminal, easy-to-explain idea that a relationship can be a translation in vector space (head plus relation is close to tail). The anchor of a whole family of knowledge-graph-embedding methods for link prediction.
Shows how building a graph index over a private corpus and summarizing communities of related entities lets a language model answer broad, connect-the-dots questions that plain chunk retrieval misses, with traceable structure.
A structured tour of representation learning, knowledge acquisition and completion, and applications. Good for readers who want to go deeper on link prediction and embeddings beyond TransE.
Keep reading.
Sibling explainers in the same series.
Back to the Research index.
Talk. Done.
Reading about clean, connected accounts is one thing. Talk to Prism and watch a note build itself from your own call.
No app required. Just your phone number.