Key Concepts

Klu is simple and intuitive, but there are a few concepts you should know.

This guide provides a high-level overview of Klu's key concepts from actions to the workspace. By the time you are finished, you will understand everything available in Klu.

1. Context — Files, sites, or data stored in the Klu Vector DB, used in RAG.
2. Apps — Organize your projects and Actions into Apps.
3. Actions — Prototype, collaborate, and version generative Actions in the Studio.
4. Data — Generations logged, along with feedback, and meta data.

Actions and Data enable downstream workflows to analyze and optimize model generations.

1. Insights — AI performance, feedback, and usage over time.
2. Evaluate — Run evaluations to measure and manage performance.
3. Optimize — Improve performance over time with LLM fine-tuning and A/B experimentation.

How Klu Works

Klu provides a single abstraction for common Generative AI functionality when designing, deploying, and optimizing functionality built on large language models. Klu provides a single endpoint and SDK for all generative actions.

Klu Studio is the GUI to design, deploy, and optimize your AI apps.

Klu Engine is the API to call for generative actions. Manage your AI apps completely without a GUI.

Workspace

Connected Model Providers, Authorized Integrations, Skills, and Context are shared across all Apps within a Workspace. Workspaces have an owner and you can invite members to a Workspace.

Anyone can create a new Workspace, but it's likely you have one Workspace per organization or team.

Connections & Integrations

Klu Context organizes data in the Klu Vector DB for search, meta filtering, and retrieval. Refresh Context programatically or set automatic refreshes in the Klu App.

Apps

Organize Actions (Completion, Assistant, Workflow, Worker) into logical projects with an App. Apps provide observability for contained Actions, organize generated Data, and Optimization via Fine-tuning or A/B Experiments.

Actions

Klu supports four Action types: Completion, Assistant, Workflow, and Worker. Each Action contains a Prompt Template, Model Config, and additional configuration linking Context or Skills to the Action, as well as Output parsing guidelines. Actions inherit Workspace defaults for the LLM Provider and Model.

Action Capabilities

Data

Klu collects Data from generative Actions, making it easy to search, filter, organize, rate, and correct generations. Each Data point includes the prompt, completion, and meta data (prompt, user, source, etc). Rate (positive, negative), label (actions, issues), or provide feedback (correct, commentary) to Data points. Filter Data into a Dataset, a collection of Data used for evaluations, fine-tuning, export, or other optimization activities.

Manage Action feedback, versions, and optimization by organizing Data into Datasets. Datasets are used for evaluations, model fine-tuning, export, or other optimization activities.

Feedback

Klu collects feedback from users to improve Action generations. Feedback includes ratings, corrections, issues, and user behavior. Gathering user feedback on generations available through API and SDK. Providing internal feedback is available through Klu in the Data section. Labeled data used for RLHF (Reinforcement Learning Human Feedback) and fine-tuning.

Optimization

Klu enables users to optimize models using labeled data. Fine-tune model performance using Datasets to train custom models. Experiment with two Actions against a reward metric to determine the best performing Action. Klu automatically handles the split test and gathers signals for the winner against the reward metric.