Policy-Aware RAG Multi-Agent AI System

This project is a local Retrieval-Augmented Generation (RAG) system with a built-in governance agent that filters sensitive or restricted information.
It allows you to:

• Load and manage multiple local corpora (datasets)
• Multi-agent orchestration
• Shared memory to recall previous queries and reasoning chains
• Block or approve queries based on governance rules
• Interact with the system using REST APIs
• Easily run locally in Docker or manually with Python
• Separated and Real-Time Logs system in each agent
• Automatic PDF Document Upload & Vector Store Building
• Fully Customizable Configuration

git clone https://github.com/pooyaphoenix/RAG-Gateway-with-Governance-Agent.git
cd RAG-Gateway-with-Governance-Agent

Build and run with Docker

docker compose up --build

Or run locally (without Docker)

python3 -m venv venv
source venv/bin/activate       # (Windows: venv\Scripts\activate)
pip install -r requirements.txt
python ra3g.py --api-port 8010 --ui-port 8501

• --api-port: Port for FastAPI backend (default: 8010)
• --ui-port: Port for Streamlit frontend (default: 8501)

All project settings are centralized in config.yml to make the system easy to configure and maintain.

Per-agent confidence thresholds can be tuned via the THRESHOLDS dictionary:

THRESHOLDS = {
    "retriever": 0.6,
    "reasoner": 0.7,
}

curl -X 'POST' \
  'http://127.0.0.1:8000/query' \
  -H 'accept: application/json' \
  -H 'session-id: default' \
  -H 'Content-Type: application/json' \
  -d '{
  "query": "What are the benefits of regular hand washing?",
  "top_k": 5
}'

Response

{
  "query": "What are the benefits of regular hand washing?",
  "answer": "Preventing the spread of infections is one of the simplest ways.",
  "governance": {
    "approved": true,
    "reason": "approved"
  },
  "trace": [
    {
      "index": 0,
      "note": "relevant passage about benefits of hand washing"
    },
    {
      "index": 4,
      "note": "direct mention of hand washing benefits"
    }
  ],
  "retrieved": [
    {
      "id": "corpus_medical_general.txt#p0",
      "text": "...",
      "source": "corpus_medical_general.txt",
      "score": 0.5118035674095154
    },
    {
      "id": "corpus_medical_governance.txt#p0",
      "text": "...",
      "score": 0.036167800426483154
    },
    {
      "id": "patient_record_001.txt#p0",
      "text": "...",
      "source": "patient_record_001.txt",
      "score": -0.003944195806980133
    }
  ],
  "confidence": 0.512,
  "session_id": "default"
}

❌ Block queries containing personal names, phone numbers, addresses, or medical records. ❌ Block queries asking for confidential corporate data. ✅ Allow general medical, scientific, or educational questions.

Also you can see Swagger in your local address: http://localhost:8010/docs

Place your text files (.txt or .md) in data/corpus/ directory.

Automatic Index Building (Default): By default, the system automatically builds the FAISS index on startup if it doesn't exist. This feature is controlled by AUTO_BUILD_FAISS = True in app/config.py.

When the FastAPI server starts, if the index is missing:

• The system will automatically scan data/corpus/ for documents
• Build and save the FAISS index to app/index.faiss and app/index_meta.pkl
• Log the indexing process

Manual Index Building (Optional): If you prefer to build the index manually or need to rebuild after updating corpus files:

python indexer.py --corpus data/corpus

Configuration:

• Set AUTO_BUILD_FAISS = False in app/config.py to disable auto-building
• Configure the corpus directory via CORPUS_DIR = "data/corpus" in app/config.py

[email protected]