> ## Documentation Index > Fetch the complete documentation index at: https://docs.llmgrid.ai/llms.txt > Use this file to discover all available pages before exploring further. # FAQs > Frequently asked questions about LLMGrid features, usage, security, and operations. ## General ### What is LLMGrid? LLMGrid is an enterprise AI gateway and orchestration platform that provides centralized access, governance, routing, and observability for large language models, tools, and agents. *** ### Is LLMGrid tied to a single model or provider? No. LLMGrid is provider‑agnostic and designed to support multiple model providers, search tools, vector stores, and integrations behind a single, consistent API surface. *** ### Does LLMGrid require changes to existing OpenAI‑based applications? Minimal changes are required. Applications using OpenAI‑compatible SDKs typically only need to update the `base_url` to point to the LLMGrid proxy and use an LLMGrid API key. *** ## Access & Authentication ### What is a Virtual Key? A Virtual Key is an API key managed by LLMGrid that controls authentication, access scope, budgets, rate limits, routing behavior, and observability for requests. *** ### Can Virtual Keys be rotated or revoked? Yes. Virtual Keys can be rotated or revoked at any time without requiring application redeployment. *** ### How is access controlled? Access is controlled using a combination of: * Virtual Keys * Teams and organizations * Budgets and limits * Guardrails * Routing policies *** ## Models & Routing ### How does model routing work? Requests are routed based on configured routing strategies, access rules, and fallback policies. Routing can consider availability, performance, and governance constraints. *** ### Can I use model aliases? Yes. Model aliases allow applications to reference stable identifiers while underlying models or providers change. *** ### What happens if a model is unavailable? If fallback models are configured, LLMGrid automatically routes requests to the next available option. Otherwise, an error is returned. *** ## Guardrails & Safety ### What are Guardrails? Guardrails are policy enforcement mechanisms that inspect and control inputs, outputs, and tool execution to ensure safety, compliance, and governance. *** ### When do Guardrails run? Guardrails can run: * Before a model call * During execution * After a response is generated * Before tool or MCP execution * In logging‑only (audit) mode *** ### Can Guardrails be scoped? Yes. Guardrails can be enforced tenant‑wide or scoped to specific keys, teams, routes, or test scenarios. *** ## Usage, Cost & Budgets ### How is usage tracked? Usage is tracked at the request level, including tokens, execution time, routing outcomes, and metadata like keys, tags, and agents. *** ### How do Budgets work? Budgets define usage limits and rate limits that are enforced automatically. When a budget is exceeded, requests may be throttled or rejected. *** ### Can costs be adjusted or discounted? Yes. Cost Tracking supports applying percentage‑based discounts to provider costs, which are reflected in usage metrics and headers. *** ## Caching & Performance ### What does caching do? Caching stores responses for repeat or deterministic requests, reducing latency and avoiding repeated model calls. *** ### What cache backends are supported? LLMGrid supports Redis‑based caching with multiple deployment modes, including single‑node, cluster, sentinel, and semantic‑aware caching. *** ### Does caching affect model behavior? Caching short‑circuits model execution for cache hits but does not alter model output semantics. *** ## Search Tools & Vector Stores ### What are Search Tools? Search Tools allow models and agents to retrieve external or live information to ground responses and enable retrieval‑augmented generation (RAG). *** ### What are Vector Stores used for? Vector Stores store and retrieve embeddings for semantic search and RAG workflows. They are referenced by ID and managed centrally. *** ### Can vector stores be tested? Yes. Vector Stores include a test mode to validate connectivity and availability without impacting production traffic. *** ## Observability & Logging ### What types of logs are available? LLMGrid provides: * Request logs * Execution and model logs * Guardrail enforcement logs * Audit logs for administrative actions *** ### Can usage and logs be filtered? Yes. Logs and metrics can be filtered by time, model, key, team, organization, tag, agent, and status. *** ### Can observability data be exported? Yes. Observability data can be accessed programmatically and integrated with external analytics or monitoring systems. *** ## Security & Compliance ### How is data protected? LLMGrid enforces secure transport, masked credentials, scoped access, and policy‑based controls through Guardrails and access management. *** ### Is LLMGrid suitable for regulated environments? LLMGrid supports common enterprise compliance requirements through configuration, observability, and governance controls rather than hard‑coded logic. *** ### Who is responsible for compliance? LLMGrid provides security and governance tooling, while customers remain responsible for application‑level data handling and regulatory obligations. *** ## Troubleshooting ### My requests are failing—where should I look first? Start with: * Request Logs * Guardrail enforcement events * Budget or rate‑limit violations * Model availability and routing status *** ### Cache hit ratio is low—why? Common reasons include: * Non‑deterministic prompts * Missing or inconsistent routing keys * Semantic caching not enabled where appropriate *** ### A Guardrail is blocking traffic unexpectedly—what should I do? Review Guardrail logs in logging‑only mode first, adjust scope or thresholds, and validate changes using the Test Playground. *** ## Getting Help ### Where can I find more documentation? Refer to: * API Reference * Guardrails * Routing Settings * Observability * Security & Compliance *** ### How do I request a feature? Use the provided support or issue‑tracking links in the UI to submit feature requests or feedback. *** If you need help beyond these FAQs, consult the relevant feature documentation or contact your platform administrator.