Mandarum Cuvex — RAG and memory at GPU speed.
A unified memory engine that runs embedding, RAPIDS/cuVS vector search, and reranking in one GPU-accelerated path — with permission-aware filtering so retrieval respects per-user and per-agent access, and freshness sync from 200+ connectors.
Why your RAG stack is slow, fragile, and leaking access.
CPU vector search blows tail latency
At production QPS and corpus size, CPU ANN search misses latency budgets. Cuvex runs billion-scale search on GPU in single-digit milliseconds.
A DB, an embedder, a reranker
Three services with three hops add latency and cost. Cuvex fuses embed, search, and rerank into one accelerated path co-located with serving.
RAG ignores access control
Retrieval that bypasses ACLs is a breach. Cuvex enforces row-level access at query time, so an agent recalls only what its caller may see.
What Cuvex does.
Billion-scale GPU vector search via cuVS.
RAPIDS cuVS runs approximate nearest-neighbor search on NVIDIA GPUs at sub-10ms p99 and 100x CPU throughput. Hybrid dense + BM25 with GPU reranking, multi-tenant isolated.
- → RAPIDS cuVS GPU ANN search
- → Sub-10ms p99 at 100M+ vectors
- → Hybrid dense + BM25 + GPU rerank
cuvex.retrieve({ query: query, model: "nemo-retriever-e5", top_k: 10, acl: "enforce" }); // GPU ANN, permission-checked
Confluence · Notion · Drive
Salesforce · Jira · Snowflake
cuVS GPU ANN + knowledge graph
Row-level ACL at query time
NeMo Retriever grounding, org-wide.
Connect documents, tickets, CRM, and warehouses into a NeMo Retriever-indexed fabric. Models ground on authoritative context instead of hallucinating it, freshness-synced from 200+ connectors.
- → NeMo Retriever embedding + reranking
- → Entity resolution across 200+ sources
- → Incremental freshness sync via CDC
Permission-aware by construction.
Every vector and document carries access rules enforced at GPU retrieval time. Recall is scoped to the caller — no exceptions, with right-to-be-forgotten built in.
- → Row-level ACL on GPU retrieval
- → Tenant + user isolation
- → Right-to-be-forgotten built in
Cuvex, in production.
One product. One platform.
Cuvex grounds the model. Voltra serves it fast. Praxon governs what it recalls and who can see it. Edgeon runs the same GPU memory pipeline on Jetson for on-device grounding. Twinex simulates retrieval load before it hits prod. Better together, every time.
Mandarum Voltra
Agentic scheduling and autoscaling that lifts GPU utilization and cuts $/token across H100/H200/Blackwell fleets.
Explore Voltra →
Mandarum Praxon
Run autonomous agents with per-action identity, NeMo Guardrails, durable execution, and an immutable audit trail.
Explore Praxon →
Mandarum Edgeon
Serve governed AI on NVIDIA Jetson fleets with cloud-parity policy, telemetry, and safe OTA.
Explore Edgeon →
Mandarum Twinex
Simulate GPU inference fleets on Omniverse to plan cost, capacity, and serving topology before production.
Explore Twinex →Give your models memory worth having.
See sub-10ms GPU retrieval, ACL-enforced recall, and NeMo Retriever grounding on your own corpus. Start free, or let the team connect your first data source.