---

name: extract description: Extract structured information (email, url, date, entity) from unstructured text with schema + anti-hallucination guardrails. license: MIT OR Apache-2.0 compatibility: Native (CLI, llama.cpp), Browser (WebLLM), Edge (Deno HTTP LLM) metadata: targets: ["email", "url", "date", "entity", "name"] version: "1.0.0" guardrails: ["json-schema", "anti-hallucination"] allowed-tools: ""

Skill: Extraction

1. Skill Name

extract

2. Description

Extract structured information from unstructured text or content. This skill transforms free-form text into validated, structured JSON output.

Supported extraction targets:

• email - Email addresses
• url - URLs and web links
• date - Dates in various formats
• entity - Named entities (people, organizations, locations)
• name - Person names (first name, last name, full name)

3. When the Agent Should Use This Skill

The agent MUST invoke this skill when:

1. The user explicitly requests extraction of structured data from text
2. The task requires identifying specific patterns (emails, URLs, dates) in content
3. The user provides unstructured text and asks for specific information types

The agent MUST NOT invoke this skill when:

1. The text is already structured (JSON, CSV, etc.)
2. The user asks for summarization or transformation (not extraction)
3. No specific extraction target is identifiable

4. Inputs

Input Validation Rules

• text must be non-empty
• text must be at least 1 character
• target must be one of the supported types
• Unknown targets cause immediate failure

5. Outputs

Output is always JSON with the following structure:

Copy{
  "<target>": <extracted_value>
}

Output by Target Type

Output Guarantees

• Output is always valid JSON
• Output always contains the requested target field
• Empty extractions return empty arrays, not null
• No free-form text in output

6. Failure Modes

This skill fails explicitly in these cases:

Failure is First-Class

• All failures are explicit
• No silent empty results
• No hidden retries without policy
• Failures propagate to the host for handling

7. Guardrail Expectations

The Extraction skill enforces these guardrails:

Structural Guardrails

1. JSON Validity: Output must parse as valid JSON
2. Schema Conformance: Output must contain the target field
3. Type Correctness: Field values must match expected types

Semantic Guardrails

1. No Hallucination: Extracted values must appear in input text
2. No Fabrication: Cannot invent emails, URLs, or dates not present
3. Plausibility: Extracted patterns must match expected formats

Guardrail Rejection Triggers

CopyREJECT if:
  - Output is empty
  - Output is not valid JSON
  - Output missing target field
  - Extracted value not found in source text
  - Value format doesn't match target type

8. Host Execution Notes

CLI Host

Copy# Execution
agent extract --text "Contact: hello@agent.rs" --target "email"

# Success output (stdout)
{"email": "hello@agent.rs"}

# Failure output (stderr) + exit code 1
Error: InvalidTarget - unknown target 'phone'

Browser Host

Copy// Execution
const result = await agent.invokeSkill('extract', {
  text: 'Contact: hello@agent.rs',
  target: 'email'
});

// Success
{ email: 'hello@agent.rs' }

// Failure - throws SkillError
SkillError: InvalidTarget - unknown target 'phone'

Edge Host

Copy// HTTP Request
POST /skill/extract
Content-Type: application/json
{ "text": "Contact: hello@agent.rs", "target": "email" }

// Success Response (200)
{ "email": "hello@agent.rs" }

// Failure Response (400)
{ "error": "InvalidTarget", "message": "unknown target 'phone'" }

9. Examples

Example 1: Email Extraction

Input:

Copy{
  "text": "For support, email us at support@agent.rs or sales@agent.rs",
  "target": "email"
}

Output:

Copy{
  "email": ["support@agent.rs", "sales@agent.rs"]
}

Example 2: URL Extraction

Input:

Copy{
  "text": "Visit https://agent.rs or check https://github.com/hwclass/agent.rs",
  "target": "url"
}

Output:

Copy{
  "url": ["https://agent.rs", "https://github.com/hwclass/agent.rs"]
}

Example 3: Date Extraction

Input:

Copy{
  "text": "The meeting is scheduled for January 15, 2024 at 3pm",
  "target": "date"
}

Output:

Copy{
  "date": "2024-01-15"
}

Example 4: Entity Extraction

Input:

Copy{
  "text": "John Smith from Anthropic met with Sarah at Google headquarters",
  "target": "entity"
}

Output:

Copy{
  "entity": {
    "people": ["John Smith", "Sarah"],
    "organizations": ["Anthropic", "Google"],
    "locations": ["Google headquarters"]
  }
}

Example 5: Name Extraction

Input:

Copy{
  "text": "The report was prepared by Dr. Jane Smith and reviewed by Michael Johnson.",
  "target": "name"
}

Output:

Copy{
  "name": ["Dr. Jane Smith", "Michael Johnson"]
}

Example 6: No Match (Valid Result)

Input:

Copy{
  "text": "This text contains no email addresses",
  "target": "email"
}

Output:

Copy{
  "email": []
}

Example 7: Guardrail Rejection

Input:

Copy{
  "text": "Contact us anytime",
  "target": "email"
}

Invalid LLM Output (REJECTED):

Copy{
  "email": "contact@example.com"
}

Rejection Reason: contact@example.com does not appear in source text - hallucination detected.

---

Contract Version: 1.0.0 Last Updated: 2024