Atlas
Resolves complex multi-hop enterprise reasoning by traversing knowledge graph entities first, surpassing conventional vector distance limitations.
Engineering highlights
The core architectural breakthroughs.
High-scale engineering demand distilled into four verified production milestones.
Relational path reasoning across 38,000,000 supply chain entities
Eliminated combinatorial path explosion via A* heuristic search
Critical tier-3 supplier disruption identified in near real time
96% faster than historical multi-database federated queries
Executive overview
The engineering challenge & solution.
Supply chain executives manage 38 million components, suppliers, shipping lanes, and factory contracts. Vector search failed completely on relational questions like: 'Which Tier-3 sub-tier suppliers in East Asia provide titanium fasteners to factories affected by the recent maritime port strike?' Pure embeddings could not trace multi-hop relationship chains.
We engineered a hybrid Graph-RAG architecture uniting Neo4j graph databases with pgvector semantic indexes. Natural language queries are parsed into structured entity seeds and Cypher graph queries. The graph engine traverses supply relationships, extracting the precise relational sub-graph before passing verified facts to the reasoning LLM.
Engineering governance
The three non-negotiables.
These are the architectural constraints that shaped every boundary — no trade-offs or compromises permitted.
5-Hop Relational Traversal
Traverse up to 5 tiers of supplier and component relationships in sub-second response times.
Deterministic Path Provenance
Every answer must output the exact graph path (Nodes & Edges) validating how the conclusion was reached.
Streaming Graph Synchronization
Real-time updates from SAP ERP, Oracle databases, and maritime AIS shipping feeds without graph lock pauses.
Eliminating Combinatorial Explosion in 5-Hop Graph Expansions with 38M Entities
Expanding 5 hops across 38 million nodes frequently resulted in millions of candidate relationship paths, exhausting database memory and causing 45-second query timeouts.
We implemented an entity-ranking heuristic and graph pruning algorithm. Before graph traversal, vector similarity scores filter the candidate seeds to the top 15 most relevant entry nodes. The Cypher traversal applies bi-directional A* search bounded by relationship edge weights, reducing the explored node space from 1.2M nodes to 420 relevant entities.
Request lifecycle
How data moves through Atlas.
Production verification
Measured outcomes in production.
Every metric below was captured under real production traffic and audited against historical baseline data.
Operational resilience
Automated safety guardrails.
Systems fail in production. Here is how Atlas survives network partitions, upstream timeouts, and anomalous inputs without human intervention.
Cypher AST Safety Validator
Validates synthesized graph queries to prevent unbounded or mutating queries.
Bounded Expansion Timeouts
Traversals automatically terminate and return partial valid sub-graphs at 1.2s.
Continuous Graph Integrity Checks
Streaming Kafka consumers reconcile orphaned nodes and broken foreign keys.
Next step
Building a system with this level of demand?
Bring us the constraint that keeps your engineering leadership up at night — the latency ceiling, the compliance perimeter, or the unyielding reliability requirement.