Findola Platform · Retrieval engine
The GPU retrieval engine that serves permission-aware, cited answers in under 2 seconds.
Retrova is Findola's owned retrieval and reranking inference engine. It turns a question into the right passages, fast, at mid-market cost. It replaces third-party embedding and rerank APIs with self-hosted, fine-tuned models running on NVIDIA GPUs and served as standardized microservices.
Why Retrova exists.
Retrieval quality and latency make or break enterprise search. Off-the-shelf API embeddings and rerankers are generic, expensive at volume, impossible to tune per corpus, and cannot run air-gapped. Hitting under 2 seconds over millions of permission-scoped documents is not possible without accelerated inference.
What it does
Retrova runs billion-scale, permission-filtered hybrid retrieval (BM25 plus GPU dense plus late-interaction rerank) over Cortexa's index and returns ranked, citation-backed evidence. Its retrieval, reranker, and verifier models are fine-tuned with NeMo on Findola's data flywheel and compiled with TensorRT for sub-2s latency, then served through NIM and Triton so any product calls one fast inference layer.
What ships in Retrova.
Every capability below maps to a real workload that needs accelerated compute.
Hybrid retrieval
Sparse, GPU dense, and late-interaction reranking in one path.
Permission-filtered ANN
Filtering happens at query time, so there is no leakage.
Distilled hot path
Models that are 10 to 100 times cheaper, with a quality fallback.
Evidence assembly
Citations with confidence scores, and abstention when retrieval is weak.
NIM-packaged
Runs in the cloud, in a VPC, and air-gapped inside Vaulta.
Architecture.
A query is permission-filtered, then cuVS recalls candidates over Cortexa vectors, a TensorRT reranker orders them, and evidence with citations returns to Findola or Agentiq. Models are trained by Lumind with NeMo and served by Triton and NIM. Prompt-injection defense uses NeMo Guardrails.
NVIDIA hardware
- L40S / A100. Production retrieval and rerank inference, batched for low latency.
- H100 / H200. Embedding and reranker fine-tuning on large corpora.
- RTX workstations. Fast eval and development of retrieval models.
NVIDIA SDKs & libraries
- cuVS / RAPIDS RAFT. GPU billion-vector ANN with metadata and permission filters.
- TensorRT / TensorRT-LLM. Compile reranker and reasoning to hit the SLA at cost.
- NeMo Retriever. Customize embeddings and reranker on customer corpora.
NVIDIA software
- NIM. A standardized inference microservice every product calls, on-prem capable.
- Triton. Multi-model serving of rerank, verifier, and distilled paths together.
Retrova is GPU-essential.
Billion-vector ANN, late-interaction reranking, and transformer inference are GPU-bound. The under-2s SLA and the $5 per user per month economics are only achievable together with TensorRT and cuVS. CPU serving blows either latency or cost.
The moat.
Per-customer fine-tunes and the flywheel make retrieval quality compound, so switching back to generic APIs is a measurable downgrade. Retrova is deeply coupled to Cortexa's permission graph.
Retrova turns Findola's "we call APIs" weakness into a GPU-essential, NVIDIA-served inference core with owned model IP. It is the single biggest lever on the acceptance score.
Ecosystem link: Retrova is the retrieval brainstem. It reads Cortexa's graph, feeds Findola search and Agentiq agents, deploys inside Vaulta, and is trained by Lumind.
See Retrova on your own data.
Connect three apps and Findola will answer your first question in under ten minutes.