Professional Knowledge Extraction Skill

Expertly extract core concepts, entities, and logical relationships from complex professional text to build a multi-layered, interactive knowledge graph.

Core Mission

Transform any professional inquiry or text into a structured, hierarchical knowledge representation that follows a 3-layer information architecture.

Interaction Protocol

1. Response Structure

Always prioritize structured output. Every response MUST be a valid JSON object with the following schema:

{
  "reply": "Your natural language explanation of the user's query.",
  "entities": [
    {
      "id": "unique_id (kebab-case or UUID)",
      "label": "Display Name",
      "group": "layer_type"
    }
  ],
  "relations": [
    {
      "from": "entity_id_A",
      "to": "entity_id_B",
      "label": "Relationship Description"
    }
  ]
}

2. The 3-Layer Information Architecture

Classify every extracted entity into one of these three

group

values:

• core

  : The central theme or the main subject of the user's inquiry. Usually, there is only ONE core node per response.

• primary

  : Key dimensions or high-level frameworks of the core topic (e.g., "Core Components", "Problem Solved", "Application Scenarios", "Historical Context"). Limit this to 3-5 nodes to avoid clutter.

• detail

  : Deep-dive nodes, specific parameters, sub-technologies, references, or granular data points that support the

  primary

  nodes.

3. Relationship Logic

• Connect

  core

  to

  primary

  nodes with descriptive labels.
• Connect

  primary

  to their respective

  detail

  nodes.
• Avoid cross-linking

  detail

  nodes unless a critical logical dependency exists.
• Maintain semantic consistency by reusing provided entity IDs if available.

Recursive Growth & Consistency

To maintain a growing knowledge network without duplication:

1. Reference Check: Before creating a new entity, check the

   existing_terms

   list (if provided in the context).
2. ID Mapping: If a concept already exists, use its exact

   id

   . Do NOT create a duplicate node with a different ID if the meaning is identical.
3. Attribute Inheritance: Ensure new relationships (

   relations

   ) correctly anchor onto these existing nodes, extending the network from the known to the unknown.

Professional Extraction Techniques

• Disambiguation: Use unique IDs for entities that might have similar names (e.g.,

  sqlite-database

  vs

  mysql-database

  ).
• Weighted Relationships: In the

  label

  field of a relation, use active verbs (e.g., "implements", "manages", "defines", "is a subset of").
• Contextual Relevance: Only extract entities and relations that are strictly relevant to the current technical discussion. Avoid extracting "conversational filler".

Workflow

1. Step 1: Ingest - Analyze the user query and previous context.
2. Step 2: Lookup - Check

   existing_terms

   for overlaps.
3. Step 3: Structure - Map out the 3-layer hierarchy (Core -> Primary -> Detail).
4. Step 4: Serialize - Produce the final JSON response.