LLM-Powered-Negotiated-Detail-Extraction-from-Signed-Agreements

Automates extraction of contract data from DocuSign-signed agreements into a Google Spreadsheet. When a contract is signed via DocuSign, the system receives a webhook, downloads the PDF, uploads it to Google Drive, extracts 32 structured fields using an LLM, and appends the data as a row in the master spreadsheet.

Pipeline

DocuSign envelope-completed webhook
  -> Deduplication check (SQLite with TTL)
  -> Download signed PDF from DocuSign
  -> Upload PDF to Google Drive (public view URL)
  -> LLM extracts 32 fields from PDF -> JSON
  -> Append row to Google Spreadsheet

Project Structure

src/
  main.py                       # Flask app, webhook endpoint, pipeline orchestration
  constants.py                  # Extraction columns, defaults, config constants
  logger.py                     # Wrapper methods for log generation

  data_sources/
    docusign.py                 # JWT auth, webhook handling, PDF download
    google.py                   # Sheets API (read/write), Drive API (upload/download)

  db/
    db.py                       # SQLite wrapper with TTL for envelope deduplication

  llm/
    llm_interface.py                 # OpenAI Responses API: prompt generation + contract content extraction
    prompt_template_analysis.md      # Meta-prompt: analyzes blank templates -> generates extraction prompt
    prompt_contract_info_extract.md  # Generated extraction prompt (used at runtime)
                                     # Prompts have been deleted due to confidentiality. 

tests/
  test.py                       # Unit tests

DOCS/
  contract_templates/           # Blank contract template PDFs (downloaded from Google Drive)

How It Works

Prompt Generation (one-time setup)

1. download_templates() fetches blank contract templates from a Google Drive folder (supports PDFs, Google Docs, Word files)
2. generate_extraction_prompt() uploads all templates with prompt_template_analysis.md to the LLM
3. The LLM analyzes template structures and produces prompt_contract_info_extract.md -- a self-contained extraction prompt with per-contract-type rules and few-shot examples

Contract Extraction (per webhook)

1. extract_contract_info() uploads the signed PDF with the generated extraction prompt
2. The LLM identifies the contract type (MSA, NDA, BAA, DPA, Pilot, Partner, Contractor, etc.) and extracts all 32 fields
3. Response is validated (all keys present, JSON format) with configurable defaults for optional fields
4. Row is appended to the spreadsheet aligned to the header

Extracted Fields

Company Name, DBA, Entity Type, Country, Price Increase terms, Term dates, Auto-Renewal, Renewal Terms, Termination Notice, Pricing Model, Payment terms, Late Payment Fees, Liability Cap, Indemnification Scope, Governing Law, Arbitration/Litigation venue, DPA/HIPAA applicability, Assignment Restrictions, Non-Solicitation, Agreement Status, Data Training rights, Internal Owner, Renewal Action Required, Notes, and Document Link.

Setup

Requirements

• Python 3.12+
• Dependencies: pip install -r requirements.txt

Environment Variables

Copy .env.example to .env and fill in:

Variable	Description
FLASK_PORT	Server port
DOCUSIGN_INTEGRATION_KEY	DocuSign app integration key
DOCUSIGN_USER_ID	DocuSign impersonated user ID
DOCUSIGN_PRIVATE_KEY	RSA private key for JWT auth
DOCUSIGN_AUTH_URL	DocuSign auth server URL
DOCUSIGN_CONSENT_REDIRECT_URI	OAuth consent redirect
DOCUSIGN_HMAC_SECRET	Webhook HMAC verification secret
GOOGLE_CLIENT_ID	Google OAuth client ID
GOOGLE_CLIENT_SECRET	Google OAuth client secret
GOOGLE_REFRESH_TOKEN	Google OAuth refresh token
GOOGLE_SPREADSHEET_ID	Target spreadsheet ID
GOOGLE_SPREADSHEET_TAB_NAME	Target tab name
TEMPLATES_GOOGLE_DRIVE_URL	Google Drive folder URL with contract templates
OPENAI_API_KEY	OpenAI API key
OPENAI_MODEL	Model name (e.g. gpt-4o)

Run

python -m src.main

Tests

python -m unittest tests.test -v

Generate Extraction Prompt

Uncomment parse_contract_templates() in main.py or call it directly to re-generate the extraction prompt from current templates.