I'm designing a serving stack
Start with capacity planning to size it before you buy hardware.
AI documentation
Practical documentation for running AI in production: deploying and operating inference, routing traffic across providers, grounding models with retrieval, sizing GPUs, tuning latency and cost, securing access, and diagnosing failures when they happen.
It is written for the problems that only appear once traffic is real — saturation, cost, reliability, and observability — rather than for choosing a model. Where a task is best done with a SelaWare product, the guide says so; everything else is vendor-neutral.
Documentation areas
Section titled “Documentation areas”Each area is a self-contained set of pages. Open one and the left navigation narrows to just that area, so you are never scrolling past the rest of the collection to find the next page.
Gateway & RoutingRouting across inference providers: listeners, upstreams, routing strategies, circuit breakers, fallback, and failover.
RAG & Vector DatabasesRetrieval-augmented generation: vector store connections, chunk budgets, hybrid search, reranking, embedding changes, and tenant isolation.
Security & GovernanceAuthentication, scopes and roles, OIDC, guardrails, PII redaction, prompt-injection defense, rate limits, and spend budgets.
Telemetry & ObservabilityOpenTelemetry export, per-hop tracing, cost attribution, and the alerts that fire before an incident rather than during it.
Capacity PlanningSizing GPU memory and concurrency for a serving configuration, then proving the plan with a reproducible benchmark.
Performance TuningTime to first token, inter-token latency, batching, KV cache quantization, speculative decoding, and cost per token.
TroubleshootingDiagnosing production failures: a searchable symptom matrix, error-code index, GPU and CUDA faults, streaming drops, and connectivity.
Looking for product documentation — installing the collector, the CLI, SDKs, or the configuration schema? That lives in Oculis documentation.
Start where you are
Section titled “Start where you are”I'm deploying to production
Work through gateway & routing and security & governance. Every guide states its prerequisites up front and ends with a command that proves the change took effect.
Something is broken right now
Go straight to the diagnostic matrix. Search it by error code, symptom, or a log line pasted from your terminal.
I need it faster or cheaper
Start with benchmarking so you can prove the change worked, then apply the performance tuning playbooks.
Five minutes to first request
Section titled “Five minutes to first request”The fastest path to a working gateway, using the Oculis Collector:
1 — Install
A single container, a Helm chart, or a binary. See installation.
2 — Point at an upstream
vLLM, Ollama, OpenAI, Anthropic, or Bedrock. See gateway & routing.
3 — Verify
oculis gateway test proves the route end to end before you send real traffic.
4 — Add resilience
Configure fallback and failover so one provider outage is not yours.
How these docs are organized
Section titled “How these docs are organized”Every task-oriented page in this collection follows the same shape, so you always know where to look:
- Compatibility strip — version, hardware, and restart impact, before you read a word.
- Prerequisites — exact versions, ports, and permissions, plus one pre-flight command.
- Steps — with tabs for Python, TypeScript, CLI/YAML, Helm, and cURL where they apply.
- Verification — the command that proves it worked, and the output you should see.
- Edge cases — known incompatibilities and the settings that quietly bite.
- Troubleshooting links — the specific failure modes this change tends to produce.
If a page is missing any of these, that is a bug. The style guide explains the contract in full.