In this section, we explore the detailed architecture and individual components that implement agentic AI for knowledge work. These components work together to create a seamless, intelligent workflow that transforms how we handle information and knowledge tasks.
File Storage
File storage with Markdown format containing 3 categories - Prompts and workflows (PnW) - User-created files (as input to PnW) - AI-generated files (as output to PnW)
This allows us to build AI workflow that looks as follows:
%%🖋 Edit in Excalidraw%%
Applications using the storage - Obsidian (PKM) : for users' interaction w/ files - Git (VCS) : for keeping the history of changes - Claude Code : for batch / complex workflows - Cursor : for interactive AI usage scenarios
User Documents
Documents can be imported from various sources 1. Obsidian Web Clipper (link) 2. Obsidian Importer (link) 3. Readwise (link) Google Docs support exporting to .md format
Prompts
Prompts are building blocks of PKM workflows - Unit of execution and evaluation Prompts have well-defined input(s) & output(s) - 1:1 / 1:N / N:1 / N:N based on number of input & output More about PKM Prompts
Example prompts and respective types:
%%🖋 Edit in Excalidraw%%
Batch Workflows
Workflow combine multiple prompts (and scripts[^1]) to complete a task - With various dependencies (e.g. sequential or parallel) Each step can be run by different LLM or Agent - Workflow orchestrator can determine the optimal workflow More about PKM [^1]: Python or shell scripts for programmable tasks
Batch workflow example:
%%🖋 Edit in Excalidraw%%
On-demand Processing
However, batch processing can't meet the demand of - Real-time user request (e.g. for PKM Agent) - Non-predetermined workloads (e.g. ad-hoc research) Therefore, we support on-demand processing - Which process unit knowledge tasks (UKT) in task queue - Main components - Knowledge Task Generator (KTG) - Knowledge Task Processor (KTP)
%%