kdeps

Appearance

Graph Dependency ​

Kdeps utilizes its custom graph library, Kartographer, to manage resource dependencies. Kartographer enables the traversal and resolution of dependent nodes, allowing Kdeps to orchestrate the execution order of resources.

This capability makes Kdeps particularly well-suited for building context-aware RAG (Retrieval-Augmented Generation) AI agents that require chaining large language models (LLMs) and other components. It also allows reusing and remixing other AI agents.

Defining Dependencies ​

To construct a dependency graph, you must define resource dependencies in the resource's requires configuration. Additionally, the target node should be specified in the workflow's targetActionID parameter.

Here’s an example of how to define a resource’s dependencies using requires:

apl
requires {
    "resourceID1"
    "resourceID2"
    "resourceID3"
}

Understanding Graph Dependencies ​

When your workflow's targetActionID is set to a target, say JSONResponder, and your resources have defined dependencies.

Kartographer processes the graph by executing:

  1. All the dependent nodes first.
  2. Finally, the specified target node, such as JSONResponder.

For instance, given the following dependency chain:

LLMResourceJSON -> PythonResource -> JSONResponder

Kartographer ensures that each node is executed in sequential order. However, by design, a resource ID cannot be reused within the same workflow. This avoids complex dependency problems such as circular dependencies. If you need to reuse a resource, it should be under a unique ID, as shown below:

LLMResourceJSON -> PythonResource -> LLMResourceJSON2 -> JSONResponder