📄 Document Compliance Agent

A multi-agent AI application that takes in documents, processes them, and checks for compliance against user-defined rules. Made using LangGraph for the agents and Streamlit for the UI – Made by Prathamesh Bheemanathi

---

⚙️ First Option: Run locally

1. Clone the Repository

git clone https://github.com/PB811/document-compliance-ai-agent-main.git
cd document-compliance-ai-agent

2. Create and Activate Virtual Environment

🖥 macOS / Linux:

python3 -m venv .venv
source .venv/bin/activate

🪟 Windows (CMD):

python -m venv .venv
.venv\Scripts\activate

🪟 Windows (PowerShell):

python -m venv .venv
.venv\Scripts\Activate.ps1

3. Create a .env file

Create a file named .env and add the following variables:

AZURE_DOCUMENT_INTELLIGENCE_ENDPOINT= "your_azure_endpoint_here"
AZURE_DOCUMENT_INTELLIGENCE_KEY= "your_azure_key_here"
OPENAI_API_KEY= "your_openai_key_here"
MISTRAL_API_KEY="your_mistralai_api_key" # You can use Mistral or any other API provider.

4. Install Dependencies

pip install -r requirements.txt

5. Run the App

streamlit run streamlit_app.py

---

🧠 Architecture

Application Pipeline

--> Document Processing

1. User uploads files.
2. Text is extracted using pdfplumber and Azure Document Intelligence OCR.
3. Text is passed to DocumentProcessingAgent to clean and store in ChromaDB.

--> Compliance Checking

1. User inputs compliance rules.
2. Rules go to ComplianceCheckAgent, which retrieves relevant docs and checks compliance.
3. A final report is generated and displayed.

Agents

The program uses 2 LangGraph agents – the DocumentProcessingAgent and the ComplianceCheckAgent.

1. DocumentProcessingAgent

• Takes the extracted text, cleans it, and stores it.
• Node: document_processing_agent (powered by OpenAI's GPT-4.1)
• Tools:
  • text_cleaning_tool: Uses LLM to clean raw data.
  • store_in_vectordb_tool: Stores cleaned text in ChromaDB.

2. ComplianceCheckAgent

• Takes compliance rules, retrieves relevant docs from the DB, and prepares a report.
• Node: compliance_checking_agent (also powered by GPT-4.1)
• Tools:
  • find_relevant_docs_tool: Regex-based tool to extract document names from rules.
  • retrieve_docs_from_vectordb_tool: Fetches document text from ChromaDB.

---

📘 Description

The Document Compliance Agent is designed to:

• Process PDF and image documents to extract text.
• Store extracted content in a vector database.
• Check documents against user-defined compliance rules.
• Provide a clear, consolidated compliance report in pdf or csv format.

The application supports multiple file uploads, AI-powered processing, and an interactive Streamlit interface for compliance verification and reporting.

Remember to switch tabs to the compliance checker after processing the documents.

Reasoning for various choices

• Langgraph was used because it's heavily customizable and allows great control over the agent. Tool calling and binding is easily handled, states can be easily customized, and it has a constantly improving library of features.
• pdfplumber used because some pdfs have directly embedded text so it saves the cost of an api call to Azure's OCR
• Azure Document Intelligence was used because of it's high accuracy of extraction and generous free tier
• ChromaDB is used because it's a lightweight vector database and easy to use, hence is perfect for this demo.
• The separation of the two pipelines is to reduce complexity in coding, and make it easy to prepare a user friendly UI

🧩 Assumptions, Limitations & Future Work

Assumptions:

• Users will upload PDFs or images of reasonable size with recognizable text.
• Compliance rules mention the exact filename with the extension (like .pdf or .png) since the extensions are used for the Regex functions

Limitations:

• The app currently supports English documents only.
• Too large files are unable to be handled since the Azure OCR api has a file size limit
• Did not have time to consider edge cases, so there might be errors

Future Improvements:

• A true CAG can be implemented in the future for improved accuracy, possibly with the help of a locally hosted LLM and a custom written caching function.
• Add support for multi-language OCR and compliance rules.
• Introduce human-in-the-loop corrections (for example if the agent is confused about a particular rule or about which file, it can ask for clarification
• Improve UI (this demo uses just streamlit which while convenient has it's limitations. Maybe in the future it can be remade with a React frontend

---