Portkey Prompt Engineering
Version, test, and deploy prompts from a centralised library – iterate without touching your codebase or redeploying your application

Version Control

Track every change to every prompt with a full version history. Roll back to any prior version instantly with a diff view between any two revisions and complete author and timestamp attribution.

• Full version history retained indefinitely.
• One-click rollback to any prior version.
• Diff view between any two versions.
• Author and timestamp tracking on every change.

A/B Testing

Split traffic between prompt versions or models and compare quality metrics in real time. Promote the winning configuration with a single click – no code changes, no redeployment required.

• Traffic splitting by configurable percentage.
• Real-time quality, latency, and cost comparison.
• Promote winner with one click.
• Test prompt versions and model swaps simultaneously.

Template Variables

Use named dynamic variables in prompts to personalise at runtime without managing prompt string interpolation in application code. Preview prompts with test values before publishing.

• Named template variables with runtime injection.
• Preview with test values before deployment.
• Variables reusable across multiple prompts.
• No string interpolation required in application code.

Performance Analytics

Track response quality, latency, and cost per prompt version side by side. Historical trend views and integration with the eval pipeline enable data-driven decisions on what to deploy to production.

• Quality score per prompt version.
• Cost and latency tracked per prompt call.
• Historical trend view across versions.
• Linked to the Portkey eval pipeline.

Framework Compatibility

Works natively with LangChain, LlamaIndex, and any OpenAI-compatible client. Prompts are fetched from the library at runtime via the Portkey SDK or REST API with no framework-specific integration work required.

• LangChain and LlamaIndex native support.
• OpenAI-compatible drop-in.
• REST API for custom clients.
• Python and JavaScript/TypeScript SDK support.

Collaborative Prompt Library

A shared workspace where engineers, product managers, and data scientists can collaborate on prompt development without merge conflicts or deployment dependencies. Prompts are organised by project, with role-based access controlling who can view, edit, or publish each entry.

• Role-based access control per team and project.
• Comments and annotations on individual prompt versions.
• Shared workspace accessible across functions without engineering bottlenecks.
• Project-level organisation for multi-product environments.

Decouple Prompts from the Release Cycle

When prompts live in code, every improvement requires an engineering deployment. Portkey's Prompt Library separates prompt iteration from the release cycle entirely – teams can experiment, roll back, and promote prompts in production without waiting for a deployment window.

• Iterate prompts without a code deployment.
• Promote or roll back any version instantly.
• No application restart required on prompt change.
• Non-engineering team members can iterate independently.

Playground and Pre-Deployment Testing

Test any prompt directly in the Portkey playground before publishing. Run against multiple models, inject template variables with test values, and review the full response alongside cost and latency metrics – all before the change reaches production traffic.

Audit and Governance

Every prompt change is tracked with author attribution, timestamp, and a full diff. Combined with the Portkey AI Gateway's request logs, teams have a complete audit trail from prompt version to individual LLM call for compliance and incident review.

Portkey Prompt Engineering Specifications: