GraphML Book a demo →
Atlas 3 is live for enterprise workloads

Your data is a graph.Make your models read it that way.

GraphML turns the relationships buried in your documents, logs, transactions, and media into a live graph, then serves that graph to LLMs and multimodal models as grounded context. Fewer wrong answers, traceable evidence, and agents that remember what happened.

Book a demo → Read the benchmarks
0
Edges in production
0
Multi-hop accuracy lift
0
p95 retrieval
0
Pilots that renew
A ∈ Rn×n · adjacencyG = (V, E) · entities and relations
Traverse
Running in production inside
v1 the thesis

Chunks lose the answer.
Paths keep it.

Vector search returns text that looks like the question. It cannot tell you that the vendor in one contract owns the subsidiary named in another, or that the same device touched four accounts before the fraud posted. Those answers live in the connections between records, and connections are exactly what embeddings flatten away.

GraphML builds the connections first. We extract entities, relations, and time from every source you have, resolve duplicates, and keep the result queryable at billion-edge scale. Retrieval then returns a subgraph with a path you can audit, which is what turns a plausible answer into a defensible one.

Vector retrieval

Nearest neighbors in embedding space. The model has to guess how the passages relate, and it usually guesses confidently.

GraphML retrieval

A connected subgraph plus the shortest supporting path. The model reasons over structure, and every hop carries a citation.

v2 the platform

Four pieces. One graph underneath.

Adopt any piece on its own. Together they close the loop from raw source to grounded answer to measured quality, and every layer writes back into the same graph.

Atlas graph foundation model

A pretrained model for graphs. Atlas learns transferable structure from 1,200 public and licensed domains, so a new customer graph gets useful node, edge, and subgraph predictions on day one instead of after a labeling project.

  • Zero-shot link prediction, node typing, and anomaly scoring
  • Text, image, video, audio, and tabular records all become typed nodes
  • Fine-tunes on a single GPU hour for domain adaptation
Zero-shot AUC within 3.1 points of a task-specific GNN trained from scratch
Relay graph-grounded retrieval

Retrieval that returns paths, neighborhoods, and community summaries instead of a bag of passages. Relay plans the traversal, samples with degree-aware weighting so hub nodes stop swallowing the budget, and hands the model a compact evidence set.

  • Path, neighborhood, and community-level retrieval in one call
  • Degree-aware sampling that recovers low-degree answer entities
  • Drop-in endpoint for your existing agent framework
Recall@10 on 3-hop questions improves +34% over hybrid vector search
Recall agent memory

A temporal memory layer for agents. Every observation becomes a timestamped claim on the graph, so an agent can answer what it knew, when it knew it, and which action followed. Contradictions get flagged rather than silently overwritten.

  • Bitemporal edges with validity windows and provenance
  • Shared memory across a fleet of agents, scoped by permission
  • Conflict detection when a new claim contradicts an old one
Cuts repeated tool calls per task by 2.4× in long-running agent runs
Lens evaluation and control

Quality you can show an auditor. Lens scores every answer against the path that produced it, tracks drift as the graph changes, and blocks responses whose supporting evidence falls below the threshold your risk team sets.

  • Path-level attribution for each generated claim
  • Regression suites that replay production traffic after a model swap
  • Policy gates for unsupported answers and stale evidence
Unsupported claims caught before delivery: 96.4% at a 2% abstain rate
v3 results

Benchmark sheet

Numbers we publish, run on public benchmarks plus one held-out enterprise corpus. Same generator model, same token budget, same hardware for every method.

release · atlas-3.1 / relay-2.4
generator · frozen, 32k context
budget · 4,000 evidence tokens
hardware · 8×H100, single region
Figure 1

Answer accuracy on multi-hop question answering

0 20 40 60 80 100 82.6 HotpotQA 61.9 MuSiQue 84.1 2WikiMHQA 79.3 Enterprise-QA Vector RAG Hybrid BM25+vec Agentic search Relay (ours)
Figure 1. Answer F1 across three public multi-hop benchmarks and one held-out enterprise corpus. All methods share the same generator and evidence budget. Graph-grounded retrieval gains the most where questions require three or more hops, since the supporting entities are rarely similar to the question text.
Figure 2

Transfer improves with pretraining breadth

0.70 0.80 0.90 1.00 supervised GNN per graph (0.962) shallow embeddings 10 50 150 400 800 1200 Atlas zero-shot pretraining domains
Figure 2. Mean AUC over 34 held-out graphs as the number of pretraining domains grows. Atlas closes most of the gap to a supervised model trained separately on every target graph, with no target labels at all.
Figure 3

Latency holds as the graph grows

10M edges 41 ms 100M edges 78 ms 1B edges 180 ms 10B edges 402 ms p95, 3-hop traversal, single region, warm cache
Figure 3. p95 end-to-end retrieval latency for a 3-hop traversal with community summarization. Growth is sublinear in edge count because traversal is bounded by the degree-aware sampling budget rather than by graph size.
Figure 4

What each component contributes

ConfigurationF1Δp95
Full Relay pipeline79.3—180 ms
− degree-aware sampling72.1−7.2166 ms
− community summaries74.8−4.5141 ms
− temporal edges76.0−3.3174 ms
− entity resolution68.4−10.9158 ms
− path reranking75.2−4.1129 ms
Figure 4. Ablation on the enterprise corpus. Entity resolution and degree-aware sampling carry the most weight. Removing components buys back a little latency and costs far more accuracy than it returns.
Reproducibility

Every number here ships with a rerun script

Customers get the harness, the prompts, the seeds, and the evaluation code. Point it at your own corpus and you will see your own numbers before you sign anything.

+34%recall@10, 3-hop
−61%unsupported claims
2.4×fewer tool calls
41Bedges served
v4 use cases

Where relationships are the answer

Six deployments that share one shape. The question spans records, the records are owned by different teams, and the answer has to survive review.

Find the ring, not the transaction

Rule engines score payments one at a time, so coordinated fraud slips through as a set of individually unremarkable events. We link accounts, devices, addresses, merchants, and counterparties into one graph and let Atlas score the structure. Investigators get a ranked ring with the path that connects it, and a narrative the model wrote from that path alone.

The same graph answers ownership questions for sanctions screening, where the exposure is four corporate layers deep and nothing in the text of any single filing says so.

3.1×more rings surfaced
−44%false positives
12 minmedian case prep
flagged ring · 5 accounts · 2 shared devices · score 0.94
Highlighted cycle is the returned evidence subgraph. Every edge carries a source record and a timestamp.

Connect the literature to the assay

Discovery teams keep papers, patents, internal assays, and trial records in separate places, then ask questions that cross all four. We build a knowledge graph over targets, compounds, pathways, phenotypes, and cohorts, and ground the model in it so a hypothesis arrives with the chain of evidence attached rather than a citation the model invented.

Multimodal nodes matter here. Figures, gels, and imaging panels enter the graph as first-class evidence rather than as unreadable attachments.

7.4×faster target triage
0 fabricatedcitations in review
18Mentities resolved
paperstargetsassayscohorts
A single hypothesis path crossing four source layers, returned as one retrievable unit.

Reason about identity and blast radius

Security data is already a graph and almost nobody queries it as one. Identities, machines, permissions, repositories, and agent credentials become nodes, and reachability becomes a question you can ask in plain language. The graph answers what an attacker reaches from a compromised token, and which of those paths a policy change would close.

Autonomous agents make this urgent. Recall keeps a permission-scoped memory per agent, so an agent cannot read what its identity was never granted.

−68%time to blast radius
4,200paths closed in Q1
Real timepermission diffing
tokenci runnersvc accountprod store
Shortest privilege path from a leaked token to a production data store, recomputed on every permission change.

See past your tier-one suppliers

Your direct suppliers are in a spreadsheet. Their suppliers are not, and that is where the disruption starts. We assemble a supplier graph from contracts, shipping records, filings, and news, then propagate risk through it so a port closure or an export rule maps to the parts and revenue it actually threatens.

Ask a scenario in plain language and get the exposed set with the dependency chain that explains it.

6 tiersdepth resolved
−29%expedite spend
2 hrsscenario turnaround
tier 4 disruptionyour line
Risk propagated from a tier-four plant to the finished good, with the dependency chain intact.

Agents that remember the org

Most enterprise copilots answer from whatever a search index happened to return, which is why they contradict themselves across a week. Recall gives a fleet of agents one shared graph of people, projects, decisions, documents, and tickets, with time on every edge. An agent can cite the decision, name the owner, and say when it changed.

Permissions come from your identity provider and are enforced at traversal time, so a retrieved path never crosses a boundary the user cannot cross.

2.4×fewer tool calls
91%answers with a source
40+connectors
recall plannerresearchercoderreviewerrouteranalyst
One memory graph, six agents, permission-scoped reads and append-only writes.

Personalize the first session, not the hundredth

Recommenders work well for users with long histories and fail for everyone else. The graph fixes the cold start by borrowing structure: a new user inherits signal from similar sessions, similar items, and similar contexts, so the first page is already relevant. Atlas scores the user-item-context graph directly instead of relying on a static embedding table.

The same graph feeds generated content, so promotions and product copy stay consistent with what the customer has actually seen.

+21%cold-start CTR
+8.6%revenue per session
<60 msscoring budget
userssessionsitems new
A first-session user scored through session and item structure rather than a sparse history.
v5 runtime

How it runs in your stack

One managed control plane, one graph engine, and an endpoint your agents already know how to call. Deploy in our cloud, in your VPC, or fully disconnected.

SOURCESINGESTGRAPH ENGINESERVINGCONSUMERS data warehouse, lakedocs, tickets, email video, audio, imagesevent and log streams third-party feeds entity and relationextraction, any modality entity resolutionblocking, matching, merge temporal versioningvalidity windows, provenance ATLAS graph engine billion-edge store · incremental embeddingsdegree-aware sampling · community index Relay retrieval APIRecall memory APILens evaluation agentscopilotsanalytics feedback and new claims write back to the graph
Deploy where the data lives

Managed cloud, your VPC, on-premise, or air-gapped. The control plane never needs a copy of your records.

Streaming updates

Change data capture keeps the graph current within seconds. Embeddings update incrementally, not by full rebuild.

Bring your own model

Works with any frontier or open-weight model. Swap generators without touching retrieval, and replay the suite to compare.

Talk to it however you like

REST, Python and TypeScript SDKs, MCP server, and native adapters for the common agent frameworks.

v6 trust

Built for the review board

The graph carries provenance, permissions, and time. That is what lets a regulated team put a model in front of a decision that matters.

certified
SOC 2 Type II and ISO 27001

Independently audited annually. HIPAA and GDPR support, with data residency in eleven regions.

isolation
Your data stays yours

No customer data is used to train shared models, ever. Single-tenant storage and per-tenant keys by default.

access
Permissions at traversal time

Row and edge-level access inherited from your identity provider. A path is filtered before it reaches the model.

provenance
Every claim has a source

Each returned edge names the record, the extractor version, and the time window it holds for.

reliability
99.95% served

Multi-region failover, read replicas, and a published status page. Latency budgets are contractual.

control
Evaluate before you trust

Lens gates unsupported answers, tracks drift, and exports an audit trail your risk team can read without us.

v7 team

Two decades of graph research,
now shipping

GraphML was founded by one of the most cited researchers in graph machine learning. The work behind Atlas started long before the category had a name, and most of it is public.

Ryan Rossi
Founder and Chief Executive
PhD, Purdue University
Previously Senior Staff Research Scientist, Adobe Research
EmailScholarLinkedIn

Ryan builds research that ends up in shipped products. His work spans graph representation learning, retrieval-augmented generation, personalization, agentic reasoning, and multimodal models, and he has moved fifteen research results into products used by hundreds of millions of people.

He sets the direction for Atlas and Relay, from pretraining strategy through the retrieval evaluations we publish, and still reviews the benchmark runs himself.

230+publications
100+patents
19k+citations
15+research to product transfers
0people across research and engineering
0papers written by the team
0granted and filed patents
3offices, San Jose and remote
selected work by our founder

A small sample of the research the platform is built on. All of it is public, and the open graph repository we maintain is free to use.

  • 2021Graph Neural Networks with Heterophilyseparating ego and neighbor signal when connected nodes disagreeAAAI
  • 2020On Proximity and Structural Role-based Embeddings in Networksmisconceptions, techniques, and applicationsACM TKDD
  • 2018Deep Inductive Network Representation Learninginductive features that transfer across graphs, the idea behind AtlasWWW Companion
  • 2018Learning Role-based Graph Embeddingsstructural roles rather than proximity aloneIJCAI StarAI
  • 2015The Network Data Repository with Interactive Graph Analytics and Visualizationthe open graph collection now used across the fieldAAAI

We are hiring

Research and engineering roles, mostly senior, mostly hard. If you have shipped graph infrastructure at scale or trained models on structured data, write to us even if nothing below matches.

Research Scientist, graph foundation models Research Scientist, multimodal extraction Staff Engineer, graph storage Staff Engineer, retrieval serving Forward Deployed Engineer Design Engineer Head of Developer Relations
v8 contact

Bring us a question
your search stack cannot answer.

We run a two-week evaluation on your data. You keep the harness, the numbers, and the graph we build, whether or not you buy anything.

14 days
to a scored pilot
Your VPC
or ours
hello@graphml.com
San Jose, California
We reply within one business day. No sequence, no drip, one human.