The Knowledge Problem No One Talks About
Every business has a knowledge problem. The answer to 'how do we handle refunds?' lives in three places: a Slack message from two years ago, a paragraph in someone's head, and a Google Doc no one can find. The answer to 'what does this integration do?' lives in the engineer who built it and nowhere else.
This is not just an organizational problem — it is an AI problem. When you build AI agents for your business, they need a source of truth to answer questions from. If that source of truth does not exist, the agents guess. And guessing is worse than saying 'I do not know.'
I build knowledge bases that solve both problems at once: a linked vault that your team can search and maintain, and that your AI agents can query through MCP to give grounded, accurate answers.
Why Obsidian: Linked, Local, and Yours
I use Obsidian as the knowledge base platform for three reasons. First, linked knowledge: Obsidian's wiki-link syntax creates a graph of connections between notes, so 'Refund Policy' links to 'Billing Process' links to 'Payment Provider Integration.' This graph is how human knowledge actually works — things connect to other things.
Second, local-first: the vault lives as plain markdown files on your machine or your team's shared drive. There is no vendor lock-in, no API dependency, and no risk of a SaaS shutdown taking your knowledge with it.
Third, machine-readable: because the vault is just markdown files in a directory, an MCP server can read them, index them, and serve them to AI agents without any proprietary integration. The AI reads the same files your team reads.
Knowledge note structure with linking
# Refund Policy
**Status**: Active (updated 2025-03-15)
**Owner**: [[Team/Operations]]
**Related**: [[Billing Process]], [[Payment Provider Integration]]
## Process
1. Customer requests refund via [[Support Channels]]
2. [[Support Agent]] checks eligibility against criteria below
3. If eligible, process through [[Payment Provider Integration]]
4. Log in [[CRM/Interactions]] with refund tag
## Eligibility Criteria
- Within 30 days of purchase
- Service not fully delivered
- No prior refund on same account in 90 days
## Escalation
Refunds over $500 require [[Team/Management]] approval.
See [[Approval Workflows]] for the process.
## AI Agent Instructions
When a customer asks about refunds, reference this policy.
Do NOT promise a refund — say "I will check your eligibility
and get back to you." Then escalate to [[Support Agent/Human]].The note includes instructions for the AI agent directly in the knowledge base. The agent reads the same note the human reads, and the policy is maintained in one place.
The Linking Strategy: Making Knowledge Findable
A knowledge base without links is just a folder of documents. The linking strategy is what makes it a system. I organize the vault into four layers: Entity notes (people, companies, tools), Process notes (how things work), Policy notes (rules and decisions), and Reference notes (external links and documentation).
Every note links to its related notes using Obsidian's wiki-link syntax. A process note links to the tools it uses, the policies it follows, and the people who own it. This creates a navigable graph where you can start at any node and find your way to any connected piece of knowledge.
I also add YAML frontmatter to every note with metadata: status, owner, last-updated date, and tags. This metadata powers the MCP server's filtering — the AI agent can query for 'all active policies related to billing' and get exactly that.
MCP Integration: Connecting the Vault to Your AI
The knowledge base connects to AI agents through a custom MCP server that reads the vault directory. The server exposes tools like search_knowledge, get_note, and list_related. When an AI agent needs to answer a question, it calls search_knowledge with the question, gets back the most relevant notes, and uses their content to formulate an answer.
The search uses a combination of full-text search and semantic similarity. Full-text catches exact matches (product names, policy terms). Semantic similarity catches conceptual matches ('how to handle an angry customer' finds the de-escalation policy even if it does not use those exact words).
The MCP server respects the same permission model as other tools: read-only by default, with write access only for agents that are explicitly authorized to update the knowledge base. Every query is logged for audit and improvement purposes.