PolyAgents - Multi-Agent System (Gemini + Qwen3)

Overview

A multi-agent system with orchestration, intelligent consensus, and persistent memory. Supports cloud agents (Gemini) and a local model (Qwen3-0.6B) for summarization and fusion of responses.

Key Features

• Orchestration of Gemini (cloud) agents with local consensus and synthesis
• Local model Qwen3-0.6B for summaries and fusion ("thinking" mode)
• Real-time streaming via WebSocket
• Memory: Redis, PostgreSQL, Qdrant (optional)
• Advanced security: API Key, JWT, rate-limit
• Robust healthcheck (waits for local model to load)

Requirements

• Python >=3.11
• Docker & Docker Compose
• Redis, PostgreSQL
• Qdrant (optional)
• Local model: Qwen/Qwen3-0.6B (HuggingFace)

Quickstart

git clone <repository-url>
cd PolyAgents-main
cp env.example .env  # configure variables
pip install -r requirements.txt
# or
# docker-compose -f docker-compose.dev.yml up -d --build

Environment (.env) excerpt

LOCAL_LLM_MODEL=Qwen/Qwen3-0.6B
HUGGINGFACE_TOKEN=hf_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
API_HOST=0.0.0.0
API_PORT=8000
NUM_AGENTS=3
CONSENSUS_ALGORITHM=synthesis
GEMINI_API_KEY=your_gemini_api_key_here
GEMINI_MODEL=gemini-pro
REDIS_HOST=redis
POSTGRES_HOST=postgres
POSTGRES_USER=postgres
POSTGRES_PASSWORD=yourpassword
POSTGRES_DB=polyagents
JWT_SECRET_KEY=your_very_secure_secret_key_here
API_KEY_ENABLED=true
RATE_LIMITING_ENABLED=true
# (see env.example for more)

Notable changes

• Local model Qwen3-0.6B: used for summarizing and fusing cloud agent (Gemini) responses in "thinking" mode.
• Parametric loading: local model is downloaded from HuggingFace using LOCAL_LLM_MODEL and, if needed, HUGGINGFACE_TOKEN.
• Robust healthcheck: container is healthy only after the local model is fully loaded.
• Healthcheck timeout increased: see docker-compose.dev.yml (start_period=240s).
• requirements.txt updated: transformers>=4.41.0, torch>=2.1, accelerate, safetensors.

Run

# Development (hot reload, robust healthcheck)
docker-compose -f docker-compose.dev.yml up -d --build
# Or backend only
python -m app.main

Troubleshooting

• If the container is unhealthy, check logs: local model loading may take 1-3 minutes.
• If you get HuggingFace/tokenizer errors, update transformers and torch as per requirements.
• For private models, set HUGGINGFACE_TOKEN.

---

PolyAgents - Sistema Multi-Agente (Gemini + Qwen3)

Descrizione

Sistema multi-agente con orchestrazione, consenso intelligente e memoria persistente. Supporta agenti cloud (Gemini) e un modello locale (Qwen3-0.6B) per riassunti e fusione delle risposte.

Caratteristiche principali

• Orchestrazione di agenti Gemini (cloud) con consenso e sintesi locale
• Modello locale Qwen3-0.6B per riassunti e fusione (modalità thinking)
• Streaming risposte via WebSocket
• Memoria: Redis, PostgreSQL, Qdrant (opzionale)
• Sicurezza avanzata: API Key, JWT, rate-limit
• Healthcheck robusto (attende caricamento modello locale)

Requisiti

• Python >=3.11
• Docker & Docker Compose
• Redis, PostgreSQL
• Qdrant (opzionale)
• Modello locale: Qwen/Qwen3-0.6B (HuggingFace)

Installazione rapida

git clone <repository-url>
cd PolyAgents-main
cp env.example .env  # configura le variabili
pip install -r requirements.txt
# oppure
# docker-compose -f docker-compose.dev.yml up -d --build

Estratto configurazione (.env)

LOCAL_LLM_MODEL=Qwen/Qwen3-0.6B
HUGGINGFACE_TOKEN=hf_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
API_HOST=0.0.0.0
API_PORT=8000
NUM_AGENTS=3
CONSENSUS_ALGORITHM=synthesis
GEMINI_API_KEY=your_gemini_api_key_here
GEMINI_MODEL=gemini-pro
REDIS_HOST=redis
POSTGRES_HOST=postgres
POSTGRES_USER=postgres
POSTGRES_PASSWORD=yourpassword
POSTGRES_DB=polyagents
JWT_SECRET_KEY=your_very_secure_secret_key_here
API_KEY_ENABLED=true
RATE_LIMITING_ENABLED=true
# (vedi env.example per altre variabili)

Novità e modifiche principali

• Modello locale Qwen3-0.6B: usato per riassumere e fondere le risposte degli agenti cloud (Gemini) in modalità thinking.
• Caricamento parametrico: il modello locale viene scaricato da HuggingFace usando LOCAL_LLM_MODEL e, se necessario, HUGGINGFACE_TOKEN.
• Healthcheck robusto: il container è healthy solo dopo il caricamento completo del modello locale.
• Timeout healthcheck aumentato: vedi docker-compose.dev.yml (start_period=240s).
• requirements.txt aggiornato: transformers>=4.41.0, torch>=2.1, accelerate, safetensors.

Avvio

# Sviluppo (hot reload, healthcheck robusto)
docker-compose -f docker-compose.dev.yml up -d --build
# Oppure solo backend
python -m app.main

Troubleshooting

• Se il container va in unhealthy, controlla i log: il caricamento del modello locale può richiedere 1-3 minuti.
• Se hai errori di HuggingFace/tokenizer, aggiorna transformers e torch come da requirements.
• Per modelli privati, imposta HUGGINGFACE_TOKEN.

---

Per supporto o richieste, apri una issue o contatta il maintainer.