Sarvantaryamin

Sarvāntaryāmin (सर्वान्तर्यामिन्) is derived from:

• Sarva (सर्व): Everything or All.
• Antar (अन्तः): Inside or Within.
• Yāmin (यामि): Controller, Guide, or Mover (from root yam).

It translates to "The Inner Controller of Everything" or "One who controls everything from within". In philosophy, it refers to the all-pervading universal Self that dwells within, guiding all thoughts and actions.

Sarvantaryamin is an "AI Swiss Army Knife" for Enterprise, designed to run fully offline/locally using Ollama. It serves as a private, secure, and intelligent interface to your enterprise data and services.

Vision

To provide a secure, local-first AI agent that can seamlessly interact with:

1. SQL Databases: Querying structured business data.
2. APIs: Invoking enterprise services via OpenAPI specifications.
3. Documents (RAG): Answering questions from technical PDFs, docs, and standards.

Architecture

The project follows a clean MVC (Model-View-Controller) architecture with Hexagonal/Clean Architecture principles:

• web/ (Primary Adapter): FastAPI entry points, routers, and dependencies.
• usecase/ (Application Layer): Business logic and orchestration of agents.
• domain/ (Domain Layer): Core entities and interfaces.
• adapters/ (Secondary Adapters):
  • agents/: specialized AI agents (SQL, API, RAG).
  • infrastructure/: Database connections, Configuration.

Prerequisites

• Docker & Docker Compose: Recommended for running the full stack.
• Ollama: Installed locally with required models (see Model Requirements below).

Model Requirements

Sarvantaryami uses two types of models for optimal performance:

Model Type	Purpose	Used By
LLM (Chat)	SQL generation, API reasoning, RAG answering, general queries	All agents
Embeddings	Intent routing, document retrieval, semantic search	Router, RAG Agent

Local Development (Mac M2/M3, 16-32GB RAM)

Component	Model	Size	Command
LLM	llama3.2:3b	~2GB	ollama pull llama3.2:3b
Embeddings	nomic-embed-text	~274MB	ollama pull nomic-embed-text

Why these models?

• Fast inference on Apple Silicon (~20-40 tok/s)
• Low memory footprint, leaves room for other services
• Good enough accuracy for testing all 4 use cases (SQL, API, RAG, General)

Production (Medium Enterprise, 32GB+ RAM, optional GPU)

Component	Model	Size	Command
LLM	llama3.1:8b-instruct	~5GB	ollama pull llama3.1:8b-instruct
Embeddings	nomic-embed-text	~274MB	ollama pull nomic-embed-text

Hardware Recommendation:

• Minimum: 32GB RAM, no GPU (CPU inference ~5-10 tok/s)
• Recommended: 32GB RAM + RTX 4060 Ti 16GB (~$400) for 25+ tok/s
• Optimal: 64GB RAM + RTX 4080 16GB (~$1000) for 50+ tok/s

Why Llama 3.1 8B?

• 128K context window (excellent for long PDFs in RAG)
• Superior instruction following and reasoning
• Enterprise-proven, well-tested

Quick Start (Docker)

The easiest way to run Sarvantaryamin is via Docker Compose.

1. Pull Required Models:

   # For local development
   ollama pull llama3.2:3b
   ollama pull nomic-embed-text

2. Configure Environment: Create a .env file in the root directory.

   cp .env.example .env

   Set DB_URI to point to your target database.

3. Run with Docker: The docker-compose.yml builds the service using Dockerfile.agent (named agent_service in compose).

   docker compose up -d

   This will start:

   • Ollama Service (Local LLM)
   • Qdrant (Vector Store)
   • Agent Service (The Sarvantaryamin Agent App)

Configuration

Configuration is managed in infrastructure/config.py and sourced from .env.

Variable	Description	Default
LLM_MODEL	Chat/reasoning model name	llama3.2:3b
EMBEDDING_MODEL	Embedding model name	nomic-embed-text
DB_URI	Database connection string	postgresql+psycopg2://...
OLLAMA_BASE_URL	Ollama server URL	http://localhost:11434
QDRANT_HOST	Qdrant vector DB host	localhost

Development Setup (Testing)

If you need a test database and services to validate the agents, this project includes a reference setup in pg_service/.

To start the test environment (Infrastructure):

cd pg_service
docker compose up -d

This spins up a complete ecosystem with Postgres (entity_service), Keycloak, etc., which the Agent can be configured to interact with for testing purposes.

Minimal Setup (Entity Service only):

cd pg_service
docker compose up -d entity_db_pg_service entity_pg_service kc_pg_service kc_db_pg_service kafka_pg_service zk_pg_service

Usage

Interactive API Documentation: http://localhost:8010/docs (port 8010 for Docker/local).

Capabilities

• Agent Interaction: POST /api/v1/agent/ask - Main entry point.

  • "Show me all countries in the database" → SQL Agent
  • "Get currency details for USD" → API Agent
  • "How does the ISO 8583 protocol work?" → RAG Agent (after uploading docs)
  • "Hello, what can you do?" → General LLM

• Document Upload: POST /api/v1/agent/upload_doc - Upload PDFs/Text for RAG.