Changelog - 2025-11-28 (#4)

Standalone CLI Tools for Post-Generation Enrichment & Evaluation

Type: New Feature Priority: HIGH

Summary

Added three new CLI tools for post-generation memo enhancement, plus improved title extraction in the export tool. These tools enable targeted improvements without regenerating the entire memo:

New Tool: Citation Enrichment (cli/enrich_citations.py)

Purpose

Add inline citations to a section WITHOUT rewriting the content. Uses Perplexity Sonar Pro to find authoritative sources for existing claims.

Key distinction from improve_section.py: This tool only adds citation markers ([^1], [^2]) to support existing text. It does NOT rewrite or expand content.

Usage

# By company name (finds latest version)
python -m cli.enrich_citations "Sava" "Market Context"

# Specific version
python -m cli.enrich_citations "Sava" "Market Context" --version v0.0.2

# Direct path
python -m cli.enrich_citations output/Sava-v0.0.2 "Market Context"

# Skip reassembly (just update section file)
python -m cli.enrich_citations "Sava" "Team" --no-reassemble

Features

• Preserves original content verbatim

• Only adds citation markers where claims need sourcing

• Automatically reassembles 4-final-draft.md after enrichment

• Shows citation count before/after

• Supports all 10 direct investment + 10 fund commitment sections

Implementation Details

• Uses src/agents/citation_enrichment.enrich_section_with_citations() for API call

• Delegates final draft assembly to cli/assemble_draft.assemble_final_draft()

• Requires PERPLEXITY_API_KEY environment variable

New Tool: Link Enrichment (cli/enrich_links.py)

Purpose

Add hyperlinks to a section for organizations (investors, gov bodies, partners, competitors, universities) and social profiles (LinkedIn, Twitter/X) for people mentioned.

Usage

# Full enrichment (orgs + socials)
python -m cli.enrich_links "Sava" "Team"

# Organization links only
python -m cli.enrich_links "Sava" "Market Context" --orgs-only

# Social profile links only
python -m cli.enrich_links "Sava" "Team" --socials-only

# Specific version
python -m cli.enrich_links "Sava" "Team" --version v0.0.2

Two-Stage Enrichment Process

Stage 1: Organization Links (Claude)

• Uses Claude to identify and link organizations mentioned

• Links investors, government bodies, partners, competitors, universities

• Only links FIRST mention of each entity

• Uses official websites (not Wikipedia/LinkedIn/Crunchbase)

Stage 2: Social Profile Links (Tavily)

• Extracts people names from section (pattern matching)

• Searches for LinkedIn profiles via Tavily

• Also searches Twitter/X profiles

• Adds links in format: Name ([LinkedIn](url), [X](url))

People Extraction Patterns

The tool identifies team members using these patterns:

1. **Name – Role** (bold name with em-dash and role)

2. CEO John Smith (role prefix pattern)

3. John Smith (CEO) (name with role in parentheses)

Implementation Details

• Requires ANTHROPIC_API_KEY for organization links

• Requires TAVILY_API_KEY for social profile searches

• Claude model: claude-sonnet-4-5-20250929 (configurable via DEFAULT_MODEL)

New Tool: Memo Evaluation (cli/evaluate_memo.py)

Purpose

Evaluate the quality of each section in a memo, providing per-section scores, fact-checking, and improvement suggestions.

Usage

# Evaluate latest version
python -m cli.evaluate_memo "Sava"

# Specific version
python -m cli.evaluate_memo "Sava" --version v0.0.2

# Direct path
python -m cli.evaluate_memo output/Sava-v0.0.2

# Brief output (scores only)
python -m cli.evaluate_memo "Sava" --brief

# Custom output file
python -m cli.evaluate_memo "Sava" --output /path/to/eval.json

Evaluation Criteria (0-10 Scale)

Each section is scored on four dimensions:

Scoring Guidelines:

• 0-3: Poor (generic, vague, unsourced, promotional)

• 4-6: Adequate (basic coverage, some specifics)

• 7-8: Good (relevant details, mostly specific, well-sourced)

• 9-10: Exceptional (rare - requires unique insights throughout)

Output

Console Output:

Section Scores
┌─────────────────────┬───────┬──────────┬───────┬─────────────┬──────────┬──────┬─────────┐
│ Section             │ Words │ Citations│ Depth │ Specificity │ Citations│ Tone │ Overall │
├─────────────────────┼───────┼──────────┼───────┼─────────────┼──────────┼──────┼─────────┤
│ Executive Summary   │   245 │        4 │   7.0 │         6.5 │      8.0 │  7.5 │     7.2 │
│ Market Context      │   890 │       12 │   8.0 │         7.5 │      8.5 │  8.0 │     8.0 │
│ ...                 │       │          │       │             │          │      │         │
└─────────────────────┴───────┴──────────┴───────┴─────────────┴──────────┴──────┴─────────┘

Overall Score: 7.3/10 (B)

JSON Output (evaluation.json):

• Per-section scores and metrics

• Strengths, issues, and suggestions per section

• Claim analysis (total, cited, uncited, suspicious)

• Top issues across all sections

• Priority improvements (from lowest-scoring sections)

• Letter grade (A through F)

Fact-Checking

The tool extracts factual claims that should be cited:

• Metrics: "500 customers", "$2M ARR"

• Financial figures: "$10M raised"

• Percentages: "40% growth"

• Market sizes: "$5B market"

• Growth rates: "150% YoY"

• Dates: "founded in 2020"

Claims without citations are flagged as potential issues.

Exit Codes

• 0: Overall score >= 6.0

• 1: Overall score < 6.0

Improvement: Export Title Extraction (cli/export_branded.py)

Problem

The extract_title_from_markdown() function was incorrectly extracting company names when the markdown file contained "Investment Memo" text later in the document (e.g., in section content).

Solution

Implemented priority-based extraction with explicit boundaries:

Priority Order:

1. Directory name pattern (most reliable): {Company}-v{version}/4-final-draft.md → "Company Name"

2. Explicit header (first 500 chars only): Investment Memo: Company at file start

3. First H1 (first 1000 chars only): # Title header near top

4. Filename fallback: Clean up filename stem

Code Change

# Priority 1: Extract from parent directory name (most reliable)
parent_dir = md_path.parent.name
dir_match = re.match(r'^(.+?)-v\d+\.\d+\.\d+$', parent_dir)
if dir_match:
    company = dir_match.group(1).replace('-', ' ')
    return f"{company} - Investment Memo", company

# Priority 2: Look for "Investment Memo" at START of file only
first_500 = content[:500]
company_match = re.search(r'^Investment Memo[:\s]+(.+?)(?:\n|$)', first_500, ...)

This prevents false matches from content deep in the document.

Files Changed

New Files (3)

Modified Files (1)

Section Name Mapping

All three new tools use consistent section name mapping:

Direct Investment (10 sections):

• "Executive Summary" → 01-executive-summary.md

• "Business Overview" → 02-business-overview.md

• "Market Context" → 03-market-context.md

• "Team" → 04-team.md

• "Technology & Product" → 05-technology--product.md

• "Traction & Milestones" → 06-traction--milestones.md

• "Funding & Terms" → 07-funding--terms.md

• "Risks & Mitigations" → 08-risks--mitigations.md

• "Investment Thesis" → 09-investment-thesis.md

• "Recommendation" → 10-recommendation.md

Fund Commitment (additional sections):

• "GP Background & Track Record" → 02-gp-background--track-record.md

• "Fund Strategy & Thesis" → 03-fund-strategy--thesis.md

• (etc.)

Workflow Integration

These tools are designed for post-generation refinement:

# 1. Generate initial memo
python -m src.main "Sava" --type direct --mode consider

# 2. Evaluate quality
python -m cli.evaluate_memo "Sava"
# Output: Overall Score: 6.2/10 (C+)
# Priority: [Team] Add specific metrics for team experience

# 3. Enrich weak sections
python -m cli.enrich_citations "Sava" "Team"
python -m cli.enrich_links "Sava" "Team"

# 4. Re-evaluate
python -m cli.evaluate_memo "Sava"
# Output: Overall Score: 7.4/10 (B)

# 5. Export
python -m cli.export_branded output/Sava-v0.0.1/4-final-draft.md --brand hypernova

Dependencies

Related Changes

This changelog is part of improvements made on 2025-11-28:

1. 2025-11-28_01.md - Initial Sava memo generation

2. 2025-11-28_02.md - Entity disambiguation system

3. 2025-11-28_03.md - Canonical draft assembly tool

4. 2025-11-28_04.md - Standalone CLI tools (this file)

Contributors

Implementation by Claude Code (Anthropic) with user guidance.