Portkey AI Observability
Full-stack LLM observability – real-time insights, tracing, cost attribution, and eval pipelines

Detailed Request Logging

Every request and response is recorded automatically at the gateway layer with over 40 metadata fields. Filter, search, and export logs by any attribute to debug issues, audit usage, and understand application behaviour in production.

• Cost and latency captured per request.
• 40+ metadata fields including model, tokens, user ID, environment, and custom tags.
• Filter logs by any attribute combination.
• Exportable log data in JSON and CSV formats.

Distributed Tracing

Monitor the full lifecycle of LLM requests and agentic workflows in a unified, chronological trace view – from the first prompt to the final response. Portkey captures every intermediate step including tool calls, sub-agent invocations, and retrieval operations.

• End-to-end request traces across multi-step workflows.
• Tool call and sub-agent tracing with parent–child linking.
• Timeline view of every execution step.
• Cost and token usage linked to each trace node.

Feedback & Eval Pipelines

Collect structured feedback at the request or conversation level and run evaluation pipelines directly on live production traffic. Portkey supports RAGAS, G-Eval, and custom eval frameworks, enabling continuous quality measurement without a separate infrastructure.

• Human and AI feedback collection at request or session level.
• RAGAS, G-Eval, and custom eval pipeline support.
• Hallucination and faithfulness scoring.
• Eval results feed into RLHF and fine-tuning workflows.

Advanced Log Filtering

Filter production logs across 40+ dimensions including model, cost, latency, user, environment, custom metadata, and eval scores. Pinpoint issues in seconds rather than trawling through raw API logs.

• Filter across 40+ dimensions simultaneously.
• Full-text search across request and response payloads.
• Save and share custom filter views.
• Export filtered results for downstream analysis.

FinOps & Cost Dashboards

Real-time cost attribution broken down by model, team, user, and project. Identify cost leaks early, enforce per-team and per-project budgets, and drive more efficient AI workflows with historical spend trend analysis.

• Cost breakdown per model, team, user, and project.
• Real-time spend dashboards with configurable time ranges.
• Budget guardrail alerts (soft and hard thresholds).
• Historical trend analysis and exportable cost reports.

Observability Integrations

Export traces and metrics to your existing observability stack without changing your application code. Portkey provides one-click integrations with all major LLM observability and monitoring platforms.

• Langfuse, Datadog, Arize, and Weights & Biases integrations.
• OpenTelemetry and Grafana export support.
• One-click integration – no code changes required.
• Works with any LLM framework or application stack.

SIEM and Data Warehouse Export

Export all observability data – logs, traces, eval results, and cost attribution – to your SIEM or data warehouse for long-term retention, compliance reporting, and cross-system analytics.

Custom Metadata and Tagging

Attach custom metadata to every request at instrumentation time: environment, feature flag, experiment ID, user segment, or any other dimension your team tracks. All custom tags are fully filterable and exportable.

Portkey Observability Specifications:

Documentation:

View the Portkey AI Observability Documentation (External Link).

View the Portkey Distributed Tracing Guide (External Link).

View the Portkey Feedback & Eval Pipelines Reference (External Link).