Regulatory Document Parser Skill

Specialized capability to parse regulatory documents (PDF, DOCX) using semtools and extract structured section hierarchies for pharmaceutical/biotech dossiers.

When to Use This Skill

Invoke this skill when:

• Analyzing regulatory document templates (ICH modules, eCTD sections)
• Extracting table of contents or section hierarchies
• Parsing PDF/DOCX files to structured markdown
• Identifying numbered sections (1.1, 2.6.2, A.3.1)
• Processing multiple documents in bulk
• Preparing templates for content generation workflows

Tool Usage Philosophy

Primary Tool: Semantic Search (search)

Semantic search is your PRIMARY discovery tool:

• Finds relevant content even with different wording
• Discovers sections without knowing exact patterns
• Allows iterative refinement (broad → specific queries)
• Essential for exploring unfamiliar document structures

Secondary Tool: Exact Matching (grep)

Grep is SECONDARY, used AFTER semantic search:

• Validates findings from search results
• Extracts precise numbering patterns
• Ensures formatting accuracy
• Only use directly if you already know exact patterns

Recommended Pattern: Search → Grep

# Step 1: Semantic discovery (PRIMARY)
search "table of contents sections" ~/.parse/template.md --n-lines 20 --top-k 10
search "module 2 clinical nonclinical" ~/.parse/template.md --n-lines 15 --top-k 15

# Step 2: Exact extraction (SECONDARY)
grep -E '^\s*[0-9]+\.[0-9]+' ~/.parse/template.md

# Step 3: Combine results
# Use search context + grep precision for complete picture

Decision Tree:

• Need to find sections? → Start with search
• Found relevant areas? → Use grep to extract exact patterns
• Already know exact format? → Can use grep directly (rare)

Core Commands

parse - Document to Markdown Conversion

Syntax:

parse "<file-path>"

Behavior:

• Converts PDF/DOCX to clean markdown
• Output cached at ~/.parse/<filename>.md
• Handles tables, hierarchies, multi-column layouts
• First parse is slow (1-10s), subsequent access instant

Examples:

# Single file
parse "template_abc/document.pdf"

# Bulk processing
find . -name "*.pdf" | xargs parse

search - Semantic Discovery (PRIMARY TOOL)

Syntax:

search "query" <files> --n-lines N --top-k K --max-distance D

Key Options:

• --n-lines N: Context lines (10-20 recommended)
• --top-k K: Number of results (3-15 typical)
• --max-distance D: Similarity threshold (0.0=perfect, 0.3=good)

When to Use: ALWAYS start with search for section discovery

Iterative Pattern:

# 1. Broad discovery
search "table of contents sections" ~/.parse/template.md --n-lines 20 --top-k 10

# 2. Refine
search "module 2 clinical nonclinical" ~/.parse/template.md --n-lines 15 --top-k 15

# 3. Target specifics
search "pharmacology toxicology" ~/.parse/template.md --n-lines 12 --top-k 10

grep - Exact Pattern Extraction (SECONDARY TOOL)

When to Use: AFTER semantic search, to extract precise patterns

Common Patterns:

# Numbered sections
grep -E '^\s*[0-9]+\.[0-9]+' ~/.parse/template.md

# Deep hierarchies
grep -E '^\s*[0-9]+\.[0-9]+\.[0-9]+' ~/.parse/template.md

# ICH modules
grep -E '^Module\s+[0-9]' ~/.parse/template.md

workspace - Performance Optimization

Commands:

export SEMTOOLS_WORKSPACE=dossierflow-templates
workspace use dossierflow-templates  # 10x faster subsequent searches
workspace status                      # Check stats
workspace prune                       # Clean stale files

When to Use: Repeated searches on same document set

Section Numbering Patterns

Common Patterns to Recognize:

# Numbered sections: 1.1, 2.5, 3.2.1
^\s*\d+\.\d+

# Deep hierarchies: 2.6.2.4.1
^\s*\d+\.\d+\.\d+\.\d+

# Lettered sections: A.1, B.2.3
^\s*[A-Z]\.\d+

# ICH modules: Module 2.5, 3.2.S.1
^Module\s+\d+|^3\.2\.[SP]

Output Format

For each extracted section, provide structured metadata:

{
  "sections": [
    {
      "title": "2.5 Clinical Overview",
      "summary": "Provides integrated analysis of clinical data including study design, patient populations, efficacy results, and safety profiles. Synthesizes findings across all clinical studies.",
      "originalHeading": "2.5 Clinical Overview"
    },
    {
      "title": "2.6.2 Pharmacodynamics",
      "summary": "Describes pharmacodynamic properties including mechanism of action, dose-response relationships, and therapeutic effects. References nonclinical and clinical PD studies.",
      "originalHeading": "2.6.2 Pharmacodynamics"
    }
  ]
}

Summary Guidelines:

• 2-3 sentences describing expected section content
• Reference typical evidence requirements (studies, data, analyses)
• Use regulatory terminology (ICH, FDA, EMA)
• Note cross-references to other modules when relevant

Error Handling

Parse Failures:

ls -la template.pdf        # Verify file exists
find . -name "*.pdf"       # Find available PDFs
ls -la ~/.parse/           # Check cache

Search No Results:

cat ~/.parse/template.md | head -100  # Verify parsed
search "section" ~/.parse/template.md --top-k 20 --max-distance 0.5  # Broaden
grep -i "keyword" ~/.parse/template.md  # Fallback to exact match

JSON Validation:

• All sections must have: title, summary, originalHeading
• No trailing commas, use double quotes
• Preserve section order from template

Best Practices

1. Search first, grep second - Always start with semantic search
2. Iterate queries - Broad → specific → targeted
3. Parse once - Parsed files cached at ~/.parse/
4. Use workspaces - 10x faster for repeated searches
5. Validate JSON - Check structure before returning
6. Preserve hierarchy - Maintain exact numbering from template

Complete Workflow Examples

For detailed end-to-end workflows with real examples, see semtools-examples.md:

• Conference paper analysis (900+ PDFs with iterative search)
• Regulatory template processing (search → grep pattern)
• Multi-template comparison
• Workspace optimization strategies

---

Note: This Skill synthesizes semtools best practices. See semtools-examples.md for complete workflows.