Skip to content

Repository files navigation

Akalin — LLM Prompt Evaluation Pipeline

Check at akalin.space

Akalin is an enterprise-grade interactive platform designed to build, benchmark, track, and auto-refine LLM system prompts and agentic tool workflows.

Akalin allows developers to test system prompts across test cases, evaluate model outputs with custom rubrics (scored 0.0–1.0), tune model-level behavioral parameters, track step-by-step tool invocation stack traces with token consumption, auto-calibrate evaluation rubrics via meta-LLMs, and run evaluations asynchronously with a PostgreSQL-backed worker queue.


Key Features

🎛️ Advanced Core Behavioral Parameters

  • Fine-Grained Parameter Tuning — Configure Temperature, Top-P, Top-K, and Max Tokens per evaluation preset or individual run.
  • Rubric Calibration Tuning — Apply custom behavioral parameters to Meta-LLMs during automated rubric calibration.
  • Toggleable UI & Run Auditing — Toggle advanced settings on demand in the frontend dashboard and inspect configured behavioral parameters on completed run detail pages.

🪄 Auto-Refine & Rubric Calibration

  • CSV Data Calibration — Auto-generate new evaluation rubrics from baseline training dataset CSVs.
  • Low-Score Meta-LLM Auto-Refinement — Automatically analyze historical low-scoring evaluation runs to produce refined, edge-case-resilient evaluation prompts.

🔍 Interactive Stack Tracing & Token Metrics

  • Vertical Execution Timelines — Step-by-step visual lifecycle tracing for agentic tool workflows (User InputAI Tool CallTool OutputAI Final Answer).
  • Granular Token Metrics — Real-time prompt (in), completion (out), and total token consumption breakdown per step.

🛠️ Mock Tool Calling & Blacklisting

  • Global & Project Tool Management — Define global mock tools with custom parameters and mock outputs, and map them to projects.
  • Per-Run Tool Blacklisting — Toggle active tools on or off per evaluation run to test LLM tool-calling logic and fallback behavior.

⚙️ Evaluation Presets & Config Management

  • Reusable Pipeline Presets — Save and reload pipeline configurations (System Prompt, Evaluation Rubric, Target Provider & Model, Evaluator Provider & Model, Pass Threshold, and Advanced Settings).
  • Quality Gates — Configurable pass/fail score thresholds (e.g., average score ≥ 0.8) with visual run status indicators.

🔑 Bring Your Own Key (BYOK) & Model Catalog

  • OpenAI-Compatible BYOK — Connect any provider (OpenAI, OpenRouter, Ollama, Anthropic proxy, local vLLM) with custom Base URLs, API Keys, and Custom Headers.
  • LLM Model Registry — Catalog of saved models with built-in connectivity test tools.

⚡ Async Background Queue & Cancellation

  • PostgreSQL Worker Pool — Asynchronous evaluation runner powered by PostgreSQL FOR UPDATE SKIP LOCKED locking for high concurrency.
  • Mid-Run Cancellation — Cancel pending or running evaluation jobs in real-time.

🔐 Auth & Account Isolation

  • Authentication & Security — User registration, JWT authentication, user account isolation, and profile management (profile details & password updates).
  • Granular Audit Trails — Track creation and modification history for prompts, presets, test cases, tools, and evaluation runs.

🏁 Version Milestones & Roadmap

  • v0.1.0 (Current)Database Schema Stability: Core database architecture and domain models are stable and locked in.
  • v1.0.0 (Target)Full Production Stability & Agent CLI:
    • API Stability: Semantic versioning and fully specified OpenAPI schemas.
    • Database Stability: Battle-tested zero-downtime migrations.
    • Frontend Stability: Pixel-perfect UI/UX and comprehensive E2E test coverage.
    • CLI for Agent Interaction: Command Line Interface allowing AI coding assistants and autonomous agents to trigger runs, fetch evaluation results, and calibrate rubrics programmatically.

Architecture

Akalin uses a decoupled, clean service-oriented architecture:

llm-evaluation-pipeline-dashboard/
├── backend/          # Go API server (Echo + sqlx + PostgreSQL)
│   ├── cmd/server/   # API entrypoint
│   └── internal/
│       ├── db/       # Database connection & schema setup
│       ├── middleware/# Auth & HTTP middleware
│       ├── models/   # Shared domain entities & DTOs
│       ├── validator/# Custom validator bridge
│       └── service/  # Clean domain services (auth, evaluation, evaluator, llmmodel, project, prompt, provider, rubric, testcase, tool, worker)
├── frontend/         # React + Vite + TypeScript + Shadcn UI
└── landing-page/     # Astro marketing website (akalin.space)

Quick Start

For detailed setup, local development, and deployment guides, refer to:

About

Akalin empowers developers to systematically evaluate LLM prompts with precision. Design, test, track, and grade system prompts against custom rubrics and programmatic checks with data-driven confidence.

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages