Sowgenx

A full-stack AI-powered tool to generate Statements of Work (SOW) using agentic workflows and memory-based retrieval.

Tech Stack

• Frontend: React, Typescript, shadcn-ui
• Backend: Python, Flask
• AI Orchestration: LangGraph
• Vector DB: pgvector (PostgreSQL)

System Architecture & Workflow

Agentic Workflow Overview

This application uses LangGraph to orchestrate multiple specialized AI agents in a workflow pattern:

1. Multi-Agent Architecture:

   • DraftingAgent: Creates the initial SOW based on user inputs
   • ComplianceAgent: Checks if the SOW meets compliance requirements
   • ValidationAgent: Validates the SOW structure and data
   • FormattingAgent: Converts the validated SOW to a formatted document

2. Workflow Graph: The agents form a directed graph that processes SOW generation in the following sequence:

   START → get_relevant_context → drafting_agent → compliance_agent → validation_agent → [conditional routing] → formatting_agent → END
   

3. Vector Database Integration:

   • Uses pgvector to store embeddings of sample SOW documents
   • Provides relevant context from similar SOWs during drafting
   • Enhances output quality by drawing from existing examples

4. How It Works in Practice:

   • User fills out project details in the UI
   • The LangGraph workflow retrieves relevant context from the vector database
   • Agents collaborate to draft, check compliance, validate, and format the SOW
   • If issues arise, the workflow loops back to refine the content
   • The final document is returned as both formatted text and a DOCX file

Key Features

• Intelligent Context Retrieval: Uses vector similarity search to find relevant examples
• Compliance Checking: Ensures SOWs meet required standards and regulations
• Validation & Error Recovery: Validates content structure with automatic error correction
• Chat-based Refinement: Users can refine the generated SOW through conversational interactions
• Document Export: Generates professionally formatted DOCX documents

UI Setup (React)

cd client
yarn install
yarn run dev

Make sure to update the API_BASE_URL in client\src\features\SOWGenerator\constants.tsx with the url server is running at.

Server Setup (Python + Flask + LangGraph)

cd server
# Create and activate a virtual environment
python -m venv venv
source venv/bin/activate  # On Windows, use: venv\Scripts\activate
pip install -r requirements.txt

1. Setup Environment Variables

Create a .env file in the root of /server based on .env-example:

cp .env-example .env

Update the required keys (e.g., OpenAI API Key, DB connection string).

2. Setup PostgreSQL with pgvector

Ensure your PostgreSQL instance has the pgvector extension enabled:

CREATE EXTENSION IF NOT EXISTS vector;

Create necessary tables using your preferred method or provided migrations (if any).

3. Run vector migrations

python generate_sample_embeddings.py

4. Run the Server

python app.py

The Flask server will start and handle LangGraph-based multi-agent interactions and vector DB retrieval.

---

Core Components

Server Components

• Agent Modules (/agents):

  • drafting_agent.py: Generates initial SOW content using LLM and vector context
  • compliance_agent.py: Checks for required fields and legal/regulatory compliance
  • validation_agent.py: Validates SOW structure and data formatting
  • formatting_agent.py: Converts validated SOW to formatted documents

• Graph Orchestration (/graph):

  • sow_graph.py: Defines the LangGraph workflow and state management

• Services (/services):

  • llm_service.py: Handles LLM interactions
  • vector_service.py: Manages vector database operations

UI Components

• SOW Generator (/features/SOWGenerator):
  • Form interface for user input
  • Document preview and chat-based refinement
  • Export functionality for generated documents

Notes

• Ensure both client and server are running simultaneously.
• You may need to allow CORS depending on deployment.
• This project assumes access to OpenAI or Hugging Face for LLM calls.
• The workflow has a built-in retry mechanism (max 10 attempts) for handling errors.