guardrails · observability
Guardrails and observability live in the routing layer: requests are validated before reaching the model and validated again on the way back, while traces capture every span of agent planning, retrieval, tool calls and model inference.
Policies take effect by stage; on a hit you can block, rewrite or just warn. Every rule can be overridden per project, with custom regexes and model-based judges supported.
| Policy | Stage | On Hit | Description |
|---|---|---|---|
| PII Masking | Before input | Rewrite | Detects phone numbers, ID numbers, emails and other sensitive fields, then masks or blocks them automatically. |
| Prompt Injection Detection | Before input | Block | Detects jailbreak and instruction-injection patterns; hits are blocked and logged to traces. |
| Keyword Filter | Before input | Warn | Custom keyword lists that warn or block on hits, overridable per project. |
| Content Moderation | After output | Block | Scores model outputs for toxicity and policy violations; unsafe content is replaced with safe fallbacks. |
| JSON Schema Validation | After output | Rewrite | Enforces structured output against a schema and auto-retries on missing fields or type mismatches. |
| Hallucination Check | After output | Warn | Scores consistency between answers and citations; low-confidence results are flagged. |
A timeline example of a real agent call: 2.8s total, including planning, retrieval, model inference, tool calls and two guardrail checks.
agent.run
Agent
2840ms · Plans and dispatches all subtasks
guardrail: input
Guardrail
60ms · Input policy check (6 policies)
planner · gpt-4o
LLM
640ms · Task decomposition and tool selection
retriever.embed
Retrieval
280ms · Query embedding and recall
tool: search_docs
Tool
510ms · Search knowledge base docs
executor · claude-3.5
LLM
980ms · Generates the final answer
guardrail: output
Guardrail
90ms · Output review and schema validation
Aggregate tokens, latency, error rate and spend by model, project or user, with automatic anomaly alerts.
Sample production traces into eval sets, run one-click regressions after version changes, and compare quality vs cost.
Export via OTLP to self-hosted Langfuse, data warehouses or existing APM — your data stays yours.
billing · subscription
Start calling models with no upfront commitment. When guardrail checks and trace volume grow, or you need unified quota and discounts across providers, switch to a plan package or a multi-model Plan — all billed in one place.
For teams getting started with low call volume.
¥0
No monthly fee · billed per use
All catalog models at native pricing
Bundles Guardrails and full-chain observability for production.
coming soon
Monthly subscription · quota resets monthly
Includes guardrail check and span metering quota
Unified quota and discounts across providers.
Prepaid
Prepaid credits · valid for 12 months
Covers popular models from multiple providers
| Capability | Pay-as-you-go | Guardrails & Observability | Multi-model Plan |
|---|---|---|---|
| Model catalog access | Full | Full | Popular models |
| Guardrails | — | All policies | All policies included |
| Trace observability | Basic logs | 30-day retention | 30-day retention |
| Automated evaluation | — | Included | Included |
| Unified discounts | — | — | Tiered discounts |
| Support | Community | Priority tickets | Dedicated CSM |
Model costs follow each provider's native pricing; guardrails are metered per check and observability per span.
Plan quotas reset monthly; overage automatically falls back to pay-as-you-go rates without service interruption.
Upgrade or downgrade anytime in the console, prorated daily; data is retained for 30 days after cancellation for export.