parishad 1.0.1

A cost-aware, local-first council of heterogeneous LLMs for reliable reasoning, coding, and factual correctness

A cost-aware, local-first council of heterogeneous LLMs for reliable reasoning, coding, and factual correctness.

---

📖 Overview

Parishad (Hindi: Council) orchestrates multiple local language models into a structured reasoning pipeline. Instead of relying on a single massive model, Parishad coordinates specialized, smaller models (2B-34B parameters) into functional roles based on traditional Indian administrative titles.

This approach achieves higher reliability and lower latency on consumer hardware (Mac Silicon, NVIDIA GPUs) by assigning tasks to the most appropriate model slot (Small/Medium/Big) tailored to your compute budget.

✨ Key Features

• 🏛️ Heterogeneous Model Council: Seamlessly orchestrate varying model sizes (e.g., Phi-3 for syntax, Llama-3 for reasoning, Mixtral for judgment).
• 🖥️ Beautiful TUI Dashboard: Real-time interactive terminal interface with visual role tracking, progress bars, and emoji indicators.
• 👁️ Vision Capabilities: Integrated PerceptionTool allows the council to "see" and analyze images via local VLMs.
• 💰 Cost-Aware Execution: Define token budgets per query; the council adapts its strategy to stay within limits.
• 🛡️ Structured Verification: The Prerak (Checker) role actively challenges hallucinations using deterministic tools and cross-examination.
• 🔌 Local-First Backends: Native support for Ollama, Llama.cpp, MLX, and Transformers.

🏛️ The Three Sabhas (Council Configurations)

Parishad offers three distinct council sizes to balance speed, cost, and depth:

1. Laghu Sabha (Fast Council)

• Roles: 5 (Darbari, Majumdar, Sainik, Prerak, Raja)
• Use Case: Quick queries, simple code generation, factual checks.
• Models: Optimized for Small/Mid models.

2. Madhya Sabha (Core Council)

• Roles: 8 (Adds Sar Senapati [Coordinator], Sacheev [Secretary], Dandadhyaksha [Magistrate])
• Use Case: Balanced reasoning for standard development tasks.
• Models: Uses Mid/Big models for deeper analysis.

3. Maha Sabha (Extended Council)

• Roles: 10 (Adds Pantapradhan [Prime Minister], Vidushak [Lateral Thinker])
• Use Case: Complex system design, creative problem solving, and strategic planning.
• Models: Full spectrum orchestration (Small + Mid + Big).

🏗️ Architecture

Parishad organizes models into a directed graph of roles:

┌───────────┐      ┌───────────┐      ┌───────────┐
│  DARBARI  │ ──►  │ MAJUMDAR  │ ──►  │  SAINIK   │
│ (Refiner) │      │ (Planner) │      │ (Worker)  │
└───────────┘      └───────────┘      └───────────┘
                                            │
                                            ▼
┌───────────┐      ┌───────────┐      ┌───────────┐
│   RAJA    │ ◄──  │  PRERAK   │ ◄──  │ SYSTEM &  │
│  (Judge)  │      │ (Checker) │      │   TOOLS   │
└───────────┘      └───────────┘      └───────────┘

• Darbari: Normalizes vague user queries into structured task specifications.
• Majumdar: Decomposes complex tasks into stepwise execution plans.
• Sainik: Executes steps, writing code or generating content using tools.
• Prerak: Validates output against facts, schema, or safety policies.
• Raja: Synthesizes the final answer and assigns a confidence score.

📦 Installation

Standard Install

pip install parishad

Optimized Install Options

Choose the backend that matches your hardware:

# Apple Silicon (M1/M2/M3) - Recommended
pip install "parishad[mlx]"

# NVIDIA GPU / Local Parsing
pip install "parishad[local,cuda]"

# With Retrieval (RAG) capabilities
pip install "parishad[retrieval]"

# For Developers (Linting, Testing)
pip install "parishad[dev]"

🚀 Quick Start

1. Launch & Setup

Run the main command to initialize the council. This automatically handles permissions, scans for models (Ollama/HF), and launches the TUI.

parishad

Pro Tip: Use /sabha to switch between council sizes (e.g., laghu for speed, maha for deep reasoning) and /scan to find new local models.

2. CLI Direct Execution

Run a specific task without entering the interactive shell.

parishad run --sabha madhyam "Analyze the 'data.csv' file and plot trends"

3. Python API

Integrate Parishad into your own applications.

from parishad import Parishad

# Initialize the council
council = Parishad(config="core")

# Execute a complex query
result = council.run("Design a scalable REST API architecture for a bookstore.")

print(f"Final Answer: {result.final_answer}")
print(f"Confidence: {result.confidence_score}/10")
# Full execution trace available in result.trace

🎮 Interactive Features

Unlock the full power of the council with these TUI capabilities:

⚡ Slash Commands

Control the session without leaving the chat interface:

• /sabha [mode]: Switch council configuration instantly (e.g., /sabha laghu for speed, /sabha maha for power).
• /scan: Re-scan your system for newly downloaded models.
• /roles: View currently active roles, their models, and slot assignments.
• /history: View past queries and load previous traces.
• /clear: Reset the conversation context.

📎 Context Awareness (@ Mentions)

Parishad allows you to "chat with your files". Directly reference local files in your query:

• @path/to/file.py: Reads the file and adds it to the context.
• @src/: Scans the directory structure.

Example:

"Analyze @src/main.py and refactor the ErrorHandler class."

🤝 Contributing

Contributions are welcome! Please follow these steps:

1. Fork the repository.
2. Install dev dependencies: pip install -e ".[dev]"
3. Ensure code style compliance: black . && ruff check .
4. Submit a Pull Request.

📄 License

This project is licensed under the MIT License.

---

_{Built with ❤️ by Om Mohite & Ashish Prajapati. Trusted by developers for local agentic workflows.}