AI Infrastructure Documentation
Documentation for building, deploying, optimizing, and troubleshooting production AI infrastructure — from agents and inference to GPUs, observability, reliability, and cost.
Popular documentation
Section titled “Popular documentation”ProductOculisAI workload discovery and observability for production infrastructure. Deploy Oculis Collector to discover supported AI workloads and report operational telemetry, without putting Oculis in the model request path.Collector 1.1.0Gateway & model routingListeners, upstreams, and routing strategies across inference providers.Provider fallback & failoverCircuit breakers, overflow routing, and local-to-cloud failover.Latency tuning (TTFT & TPOT)Prefix caching, chunked prefill, KV cache quantization, FlashAttention.Capacity planning & VRAM sizingInteractive planner for weights, KV cache, and concurrency ceilings.Telemetry & observabilityOpenTelemetry export, per-hop tracing, cost attribution, alert rules.Oculis quickstartSee your first AI workload and its activity in Oculis after installing Oculis Collector.GPU, CUDA & memory troubleshootingCUDA OOM, KV cache exhaustion, NCCL timeouts, and Xid hardware faults.Guardrails & prompt-injection defensePII redaction, injection classification, and fail-open vs fail-closed.Oculis APIProgrammatic access to your Oculis data, and how requests are authenticated.
All documentation
Section titled “All documentation”Everything published across the hub, filterable by product and by subject — a page has both, and the two filters narrow independently. Products and categories appear here only once they have documentation behind them.
Product
Subject
- Benchmarking methodologyBuild a representative corpus, find the saturation knee, prove a change.
- Capacity planning & VRAM sizingInteractive planner for weights, KV cache, and concurrency ceilings.
- Changing embedding modelsRe-embedding a corpus and cutting over with an alias swap.
- Collector CLICheck Collector status, confirm the installed version, and remove the Collector.
- Credentials & accessDeploy tokens, API keys, and how member access is controlled.
- Gateway & model routingListeners, upstreams, and routing strategies across inference providers.
- GPU, CUDA & memory troubleshootingCUDA OOM, KV cache exhaustion, NCCL timeouts, and Xid hardware faults.
- Guardrails & prompt-injection defensePII redaction, injection classification, and fail-open vs fail-closed.
- Inference & streaming troubleshootingGateway timeouts, SSE drops, rate-limit cascades, context overflow.
- Latency tuning (TTFT & TPOT)Prefix caching, chunked prefill, KV cache quantization, FlashAttention.
- Networking & connectivityTLS verification, DNS, egress policy, and proxy configuration.
- Oculis APIProgrammatic access to your Oculis data, and how requests are authenticated.
- Oculis CollectorThe component that connects AI workloads and infrastructure to Oculis.
- Oculis compatibilitySupported platforms, architectures, and workload types for Oculis Collector 1.1.0.
- Oculis core conceptsCollectors, systems, agents, AI-capable applications, and runs.
- Oculis installationConnect Oculis to your infrastructure by adding your first Collector.
- Oculis quickstartSee your first AI workload and its activity in Oculis after installing Oculis Collector.
- Provider fallback & failoverCircuit breakers, overflow routing, and local-to-cloud failover.
- RAG retrieval configurationChunk budgets, hybrid search, reranking, and tenant isolation.
- Rate limits & token budgetsPer-tenant request rates, token ceilings, and spend caps.
- Telemetry & observabilityOpenTelemetry export, per-hop tracing, cost attribution, alert rules.
- Throughput & cost optimizationContinuous batching, speculative decoding, and semantic routing.
- Vector store integrationQdrant, Milvus, pgvector, and Pinecone connections and failure modes.
Nothing matches that filter yet.
Showing 23 of 23 documentation entries.