Concepts
Pipe is the fastest way to turn ideas into AI. Pipe is like an AI feature. It is a high-level layer to Large Language Models (LLMs) that creates a personalized AI assistant for your queries.
Let's understand the key concepts of Pipe:
Meta
The Pipe meta defines its configuration. It contains the following information:
Type
The type of Pipe, i.e., generate or chat. The type of Pipe determines the behavior of the Pipe. For example:
- Generate Pipe is designed to generate LLM completions.
- Chat Pipe is designed to create a conversational AI assistants.
Stream mode
Handles whether the Pipe should stream the response or not. If enabled, the Pipe will stream the response in real-time.
Store messages
Pipe can store both prompts and their completions if the Store messages in Pipe meta is enabled on. Otherwise, only system prompts and few-shot messages will be saved. No completions, final prompts or variables will be retained to ensure privacy.
Moderate
Available only for OpenAI models. Moderation endpoint by OpenAI identifies harmful content. If enabled, Langbase blocks flagged requests automatically.
JSON
Enforces the completion to be in JSON format. If enabled, the completion will be in JSON format.
Variables
Any text written between {{}} in your prompt instructions acts as a variable to which you can assign different values using the variable section. Variables will appear once you add them using {{variableName}}.
On runtime, these variable will dynamically populate with the assigned values during execution
Safety
Define AI safety prompt for any LLM inside a Pipe. For instance, do not answer questions outside of the given context.
One of its use cases can be to ensure the LLM does not provide any sensitive information in its response from the provided context.
Experiments
They help you learn how your latest Pipe config will affect LLM response by running it against your previous generate requests.
One example of Experiments can be changing Pipe's LLM model to gemma-7b-it from gpt-4-turbo-preview to check how the response will look like.
Few-shot training
It helps AI LLM pick up and apply knowledge from just a handful of examples.
Pipe lets you define multiple user and AI assistant prompts and completion pairs that can be used to few-shot train any LLM.
Pipe level keysets
Pipe LLM keyset is specific to each individual pipe. When selected, the Pipe doesn't use the user/org LLM API keys but instead use the Pipe level keyset added to it in its settings.