Changelog - 2025-11-26 (#1)

Dataroom Analyzer Agent System & Export Improvements

Overview

Introduced a new multi-agent dataroom analyzer system for processing investment datarooms, plus improvements to trademark handling and export versioning.

New Feature: Dataroom Analyzer

Purpose

Analyze investment datarooms containing diverse document types (pitch decks, financials, legal docs, competitive analysis, etc.) and produce structured inventory with intelligent classification.

Architecture

Created new module at src/agents/dataroom/ with 5 files:

src/agents/dataroom/
├── __init__.py              # Package exports
├── analyzer.py              # Main orchestrator (8.9 KB)
├── dataroom_state.py        # TypedDict schemas (11 KB)
├── document_scanner.py      # Directory scanning (6.7 KB)
├── document_classifier.py   # 3-stage classification (17 KB)
└── extractors/              # Placeholder for Phase 2
    └── __init__.py

Three-Stage Classification System

The classifier uses a hierarchical approach for maximum accuracy:

Stage 1: Directory-Based (Highest Confidence)

• Parses directory names like "1 1_0 Executive Summary" or "4 4_0 GTM_Competitive Overview"

• Strips numeric prefixes and normalizes separators

• Maps to document types with 85-90% confidence

Stage 2: Filename Pattern Matching

• Regex patterns for common naming conventions

• Examples: "cap.*table" → cap_table, "battlecard" → competitive_analysis

• Confidence based on match specificity (60-85%)

Stage 3: LLM Content Classification (Optional)

• Uses Claude Sonnet for uncertain documents

• Extracts content samples (first 2000 chars)

• Only invoked when earlier stages have <70% confidence

Document Type Taxonomy

Priority 1 (Critical):

• pitch_deck - Investor presentations, executive summaries

• competitive_analysis - Battlecards, competitor comparisons, market maps

• financial_statements - P&L, balance sheets, historical financials

• financial_projections - Operating models, forecasts, budgets

• cap_table - Ownership structure, equity breakdown

Priority 2 (Important):

• term_sheet - Investment terms, SAFE notes

• team_bios - Founder/leadership backgrounds

• customer_list - Notable customers, case studies

• traction_metrics - KPIs, growth data

Priority 3 (Supporting):

• product_documentation - Architecture, specs, roadmaps

• marketing_collateral - Datasheets, brochures

• legal_document - Corporate docs, agreements

• investor_update - Board materials, governance

TypedDict Schemas

Defined 11 structured schemas in dataroom_state.py:

• DocumentInventoryItem - Individual document metadata

• FinancialData - Income statement, balance sheet, metrics

• CapTableData - Shareholders, options pool, SAFEs

• CompetitiveData - Competitors, positioning, SWOT

• TeamData - Founders, leadership, advisors

• TractionData - Customers, revenue, pipeline

• LegalDocData - Term sheets, investment terms

• DataroomAnalysis - Main analysis output container

Output Artifacts

Follows project conventions with versioned output:

output/Hydden-v0.0.1/
├── 0-dataroom-analysis.json  # Structured inventory (14 KB)
└── 0-dataroom-analysis.md    # Human-readable report (4.3 KB)

Markdown Report Includes:

• Document inventory summary by type

• Detailed document list with confidence indicators (🟢/🟡/🔴)

• File sizes, page counts, classification reasoning

• Data gaps (missing document types)

• Processing notes

Usage

# CLI usage
python -m src.agents.dataroom.analyzer "path/to/dataroom" "Company Name"

# Programmatic usage
from src.agents.dataroom import analyze_dataroom
analysis = analyze_dataroom("path/to/dataroom", "Company Name")

Testing Results

Tested on Hydden dataroom (20 documents):

• Classification accuracy: 100% (all 20 documents correctly typed)

• High confidence: 20/20 documents (100%)

• Processing time: 0.1 seconds

• Data gaps identified: Missing team_bios, financial_statements

Improvement: Trademark Path Handling

Problem

Trademark paths in company JSON files were being malformed when using absolute paths:

# Before (broken)
![Logo](../..//Users/mpstaton/code/.../logo.svg)

Solution

Updated src/agents/trademark_enrichment.py to properly handle:

• HTTP/HTTPS URLs: Pass through unchanged

• Absolute paths: Convert to relative from project root

• Relative paths: Prepend ../../ correctly for output directory navigation

# Now supports clean relative paths in data/*.json
"trademark_light": "data/Secure-Inputs/trademark__OWM--Light-Mode.svg"

Files Modified

• src/agents/trademark_enrichment.py (+15 lines)

Improvement: Export Auto-Versioning

Problem

Running export multiple times would overwrite the same HTML file.

Solution

Added automatic versioning to cli/export_branded.py:

• Detects if output file exists

• Appends .1, .2, etc. to filename

• Supports 4-segment versioning (e.g., OWM-v0.0.2.1.html)

# First export
OWM-v0.0.2.html

# Subsequent exports
OWM-v0.0.2.1.html
OWM-v0.0.2.2.html

Files Modified

• cli/export_branded.py (+11 lines)

Planning Document

Created comprehensive planning document for the dataroom analyzer system:

• context-vigilance/Dataroom-Analyzer-Agent-Plan.md

• Includes architecture, classification strategies, implementation phases

• Updated based on real Hydden dataroom analysis

Files Changed Summary

New Files (6)

Modified Files (2)

Documentation (1)

Next Steps (Phase 2)

The dataroom analyzer foundation is complete. Future phases:

1. Document Extractors - Build type-specific extractors:

   • competitive_extractor.py - Parse battlecards, synthesize competitive landscape

   • financial_extractor.py - Extract P&L, projections, metrics

   • cap_table_extractor.py - Parse ownership structure

2. Data Synthesis - Cross-document fact extraction and conflict resolution

3. Workflow Integration - Connect to main memo generation pipeline

Testing & Validation

Dataroom Analyzer

• ✅ Scans directories recursively

• ✅ Handles numbered directory patterns (N N_0 Category Name)

• ✅ Classifies with 3-stage confidence system

• ✅ Outputs to versioned output/ directory

• ✅ Generates both JSON and markdown artifacts

• ✅ Identifies data gaps

Export Improvements

• ✅ Trademark relative paths work correctly

• ✅ Auto-versioning prevents overwrites

• ✅ Both light and dark mode exports working

Contributors

Implementation by Claude Code (Anthropic) with user guidance.