3.11.5 Trigger Matching Mechanism

D-Moss Agent uses a trigger keyword matching mechanism to determine which skills are loaded into the current conversation context. This section describes how this mechanism works and how it differs from RAG (Retrieval-Augmented Generation).

Processing Workflow in D-Moss

When Studio starts up, the SkillManager inside D-Moss Agent performs the following initialization steps:

1. Scans all files matching skills/**/SKILL.md (recursive glob)
2. Parses the frontmatter of each SKILL.md file
3. Builds an in-memory index using fields such as trigger, category, and risk
4. Exposes the /api/skills endpoint for frontend skill listing
5. Provides an inject() method for on-demand skill loading during conversations

Each time a user sends a message in AI Dock:

1. SkillManager scans the trigger keywords of all installed skills
2. Matches relevant skills (e.g., if the user says "USB camera black screen," it matches rdk-x5-camera-debug)
3. Injects the body content of the matched SKILL.md file into the model's context
4. The model follows the instructions in SKILL.md to make decisions and execute actions (possibly invoking tools like device_exec, device_file_read, etc.)
5. The response to the user already includes a "targeted solution + specific commands"

Differences from RAG

The skill mechanism fundamentally differs from RAG (Retrieval-Augmented Generation):

Dimension	RAG	Skill (Trigger Matching)
Trigger Method	Vector similarity matching	String keyword matching
Loading Timing	Attempts retrieval on every conversation turn	Loads only when a trigger is matched
Context Consumption	May inject numerous fragments	Injects only the full SKILL.md of matched skills
Interpretability	Opaque (vector similarity is hard to explain)	Transparent (clearly shows which trigger led to which skill being loaded)
Suitable Use Case	Semantic search across massive document collections	Structured operational playbook libraries

The skill mechanism is better suited for scenarios where you want to "teach the AI a specific workflow or pattern." RAG is more appropriate for "finding answers within large volumes of documents." Both approaches can coexist—RDK Studio supports both mechanisms simultaneously (RAG for built-in documentation retrieval, and skills for operational strategies).

Matching Precision

Trigger matching performs exact string matching without semantic understanding. This means:

Trigger Word(s)	User Message	Matched?
USB camera	"What should I do if the USB camera screen is black?"	Matched (exact string match)
USB camera	"The camera on my board is not displaying any image"	Not matched (semantically equivalent but textually different)
USB camera, camera, cam	"The camera on my board is not displaying any image"	Matched (trigger includes "camera")

To improve matching accuracy, we recommend the following when designing triggers:

• List common synonyms and variants (in Chinese, English, and abbreviations)
• Include distinctive strings from error messages (e.g., code -6, SIGABRT)
• Incorporate colloquial expressions developers commonly use when describing issues

Verifying Skills Loaded in Current Session

Type /skills in AI Dock:

/skills

Example output:

Skills loaded in current session:
- rdk-device-ops (Core Device Operations)
- rdk-x5-camera-debug (X5 Camera Debugging)
- rdk-developer-docs (RDK Official Documentation Retrieval)

If you've installed a skill but it doesn't appear in the /skills output, it means its trigger didn't match the current conversation. Simply update the trigger field in the SKILL.md file and restart Studio to rebuild the index.