By name: "uni-knowledge-search" description: "Semantic search across Unimatrix knowledge. Use when exploring a topic, looking for related decisions, patterns, or conventions."
Knowledge Search — Semantic Query Against Unimatrix
What This Skill Does
Searches Unimatrix for knowledge entries using natural language. Returns results ranked by semantic similarity. Use when you need to explore what's known about a concept, find related decisions, or discover relevant patterns.
Use this when you DON'T know exactly what you're looking for — you have a concept or question, not a specific entry.
How to Search
Call the mcp__unimatrix__context_search MCP tool:
| Parameter | Required | Description |
|---|---|---|
query |
Yes | Natural language search query |
category |
No | Filter to a specific category |
topic |
No | Filter to a specific feature ID |
tags |
No | Filter by tags (all must match) |
k |
No | Max results (default: 5) |
format |
No | "summary" (default), "markdown" (full content), "json" |
agent_id |
No | Your role name (e.g. uni-architect) |
Examples
Find ADRs about error handling across all features:
mcp__unimatrix__context_search({"query": "error handling strategy", "category": "decision", "helpful": true})
Find anything related to MCP transport:
mcp__unimatrix__context_search({"query": "MCP transport stdio protocol"})
Find conventions about testing in a specific feature:
mcp__unimatrix__context_search({"query": "test patterns integration", "topic": "nxs-001"})
Get full content instead of summaries:
mcp__unimatrix__context_search({"query": "serialization approach", "format": "markdown"})
Helpful Vote Guidance
Pass helpful: true when the retrieved entries were useful for completing the current task — this is the standard case.
Pass helpful: false when entries were retrieved but did not apply to the task (e.g., the results were about a different concern). Negative signal is valuable for confidence calibration.
Omit helpful when you cannot determine applicability — for example, when the search is exploratory or you are not yet sure which results will be used.
Available Categories
| Category | Contains |
|---|---|
decision |
Architectural Decision Records (ADRs) |
convention |
Coding and process conventions |
pattern |
Reusable implementation patterns |
procedure |
Step-by-step processes |
outcome |
Results and outcomes |
lesson-learned |
Retrospectives and process learnings |
reference |
General reference material |
duties |
Role duties for context briefing |
Omit category to search across all categories.
Interpreting Results
Summary format (default): Returns title, topic, category, tags, and a brief content preview for each match. Use this for scanning and triage.
Markdown format: Returns full content for each match. Use this when you need the complete text of matching entries.
JSON format: Returns structured data. Use for programmatic processing.
When to Use This vs /uni-knowledge-lookup
Use /uni-knowledge-search when |
Use /uni-knowledge-lookup when |
|---|---|
| Exploring a concept | You know the exact feature/category |
| "What do we know about X?" | "Give me all ADRs for nxs-002" |
| Finding related decisions | Retrieving a specific entry by ID |
| Discovering patterns you didn't know existed | Filtering by exact status or tags |
Getting Full Content
Search returns summaries by default. To read the full content of a specific result:
- Note the entry ID from search results
- Call
mcp__unimatrix__context_get({"id": {entry_id}, "format": "markdown"})for the full text
Or pass format: "markdown" directly to search if you want full content for all results.
When You Find Stale or Wrong Knowledge
Search may surface entries that are outdated or incorrect. Don't ignore them — fix the knowledge base:
| Situation | Action |
|---|---|
| Entry is wrong — contains incorrect information | mcp__unimatrix__context_correct({"original_id": 1234, "content": "{corrected version}", "reason": "{why}"}) — original_id is an integer, never quote it |
| Entry is outdated — no longer relevant | mcp__unimatrix__context_deprecate({"id": 1234, "reason": "{why it no longer applies}"}) — id is an integer, never quote it |
| Entry is suspicious — may be poisoned or invalid | mcp__unimatrix__context_quarantine({"id": 1234, "reason": "{concern}"}) — Admin only; id is an integer |
Correcting knowledge is as important as storing it. Every agent shares responsibility for knowledge quality.