Skip to content

tech-diligence-report-generator.md

Latest commit

 

History

History
144 lines (107 loc) · 4.4 KB

tech-diligence-report-generator.md

File metadata and controls

144 lines (107 loc) · 4.4 KB

Tech Diligence Report Generator

A document analysis system that uses vision-language models (VLMs) to extract technical specifications from PDFs with full provenance tracking.

Overview

The Tech Diligence Report Generator processes PDF documents by:

  1. Extracting text from each page
  2. Converting each page to a high-quality image for VLM analysis
  3. Using GPT-4o vision to analyze both text and visual content (diagrams, tables, schematics)
  4. Tracking provenance (source document, page number, relevant text) for every claim

Components

PDF Processing Pipeline

PdfImageExtractor (app/services/pdf_image_extractor.rb)

  • Converts PDF pages to PNG images using pdftoppm (poppler-utils)
  • 150 DPI for good quality while keeping file sizes manageable
  • Supports single page or batch extraction

DocumentProcessingJob (app/jobs/document_processing_job.rb)

  • Extracts text using pdf-reader gem
  • Generates page images and stores as Active Storage blobs
  • Stores references in page_images JSON column: { "1" => "signed_blob_id", ... }

TechDiligenceAgent

Agent (app/agents/tech_diligence_agent.rb)

  • Inherits from ApplicationAgent with has_context for provenance tracking
  • Supports vision API via api_version: :chat
  • Four main actions:
    • answer_questions - Answer specific technical questions
    • analyze_pages - Deep analysis of specific pages
    • extract_specs - Extract structured specifications
    • verify_claims - Verify claims against document evidence

Views (app/views/tech_diligence_agent/)

  • answer_questions.text.erb - Q&A prompt with citation requirements
  • analyze_pages.text.erb - Page analysis prompt
  • extract_specs.text.erb - Structured spec extraction
  • verify_claims.text.erb - Claim verification prompt

Provenance Tracking

AgentFragment (app/models/agent_fragment.rb)

  • New fragment types: citation, spec_extraction
  • Metadata JSON includes:
    • page_number - Source page
    • source_document - Document filename
    • confidence - Extraction confidence score
  • Scopes: citations, with_page_reference(page_num)

API Endpoints

POST /assistants/tech_diligence/questions
  document_id: integer (required)
  questions: array of strings
  use_vision: boolean (default: true)

POST /assistants/tech_diligence/pages
  document_id: integer (required)
  page_numbers: array of integers
  focus_areas: array of strings

POST /assistants/tech_diligence/specs
  document_id: integer (required)
  spec_categories: array (default: [cpu, memory, storage, connectivity, power])

POST /assistants/tech_diligence/verify
  document_id: integer (required)
  claims: array of strings

All endpoints return { stream_id: string } for ActionCable streaming.

Usage

CLI Testing

# Process a PDF (extract text + page images)
rake tech_diligence:process_pdf[/path/to/motherboard_manual.pdf]

# Ask questions about a document
rake tech_diligence:ask[123,"What is the CPU socket type?"]

# Extract all specifications
rake tech_diligence:extract_specs[123]

# Run full test suite with example questions
rake tech_diligence:test_motherboard[123]

Programmatic Usage

# Load and process a document
document = Document.find(123)

# Ask questions with vision
agent = TechDiligenceAgent.with(
  document: document,
  questions: [
    "What is the CPU socket type?",
    "What is the maximum RAM speed?",
    "Does it have optical audio?"
  ],
  use_vision: true
)

response = agent.answer_questions.generate
puts response.message.content

# Check provenance
agent.context.fragments.citations.each do |citation|
  puts "Page #{citation.source_page_number}: #{citation.generated_content}"
end

Citation Format

The agent is instructed to cite sources in a parseable format:

The CPU socket type is LGA 1700 [Source: motherboard_manual.pdf, Page 12]

These citations are automatically parsed and stored as AgentFragment records with full metadata.

Requirements

  • pdftoppm (from poppler-utils): brew install poppler
  • pdf-reader gem (already in Gemfile)
  • image_processing gem (already in Gemfile)
  • OpenAI API key with GPT-4o vision access

Future Enhancements

  1. Line-level provenance - Track specific line numbers within pages
  2. Confidence scoring - ML-based confidence for extracted claims
  3. Multi-document analysis - Compare specs across multiple documents
  4. Export to structured formats - JSON, CSV, or custom report templates
⚡