PolyAgent

简体中文

PolyAgent is a multi-agent workspace for knowledge-heavy service workflows. It combines grounded QA, document summarization, formal writing, and step-by-step guidance in one chat interface.

The current repository ships with Chinese prompts, example data, and UI copy, while the architecture itself remains domain-agnostic. You can adapt it to internal knowledge bases, service desks, operations workflows, education scenarios, or public-facing assistants.

Recommended entrypoint: run python app.py for the primary web workspace.

Why PolyAgent

• One workspace, four core capabilities: QA, summary, writing, and guidance.
• Multi-agent routing for compound requests such as "find the policy and draft the reply".
• RAG-ready knowledge flow with Markdown sources, vector retrieval, and traceable context.
• Demo-friendly UI assets for README screenshots, release notes, and social distribution.

Demo

Desktop home

Capability gallery

Multi-agent orchestration	Document summarization
Formal writing	Step-by-step guidance

Core Capabilities

• Answer questions from a private knowledge base with retrieval-backed context.
• Summarize uploaded files or long-form text into structured takeaways.
• Draft formal notices, applications, or structured business copy from short prompts.
• Guide users through multi-step procedures in a conversational way.
• Decompose one request into multiple subtasks and execute them sequentially.

Quick Start

Requirements

• Python 3.10+
• DEEPSEEK_API_KEY
• DASHSCOPE_API_KEY

Install

git clone https://github.com/Powfu-zwx/PolyAgent.git
cd PolyAgent

python -m venv .venv
source .venv/bin/activate
# Windows: .venv\Scripts\activate

pip install -r requirements.txt
cp .env.example .env

Default run path

python app.py

Then open http://127.0.0.1:8000.

If you want retrieval-based answers, put your Markdown documents under knowledge/data/ and build the vector index:

python -m knowledge.vectorstore build

If you are coming to the project for the first time, use the app.py route above. The ui.py entrypoint is kept only as an alternate Gradio demo interface.

How It Works

1. Prepare and compress the current conversation context.
2. Route the user request into one or more specialist intents.
3. Execute task-specific agents for QA, summary, writing, and guidance.
4. Aggregate results into one final response.

Typical Use Cases

• Internal knowledge base assistant for policy or procedure lookup
• Service desk or operations helper for multi-step internal workflows
• Document assistant for notices, summaries, and structured writing
• Scenario demos for LangGraph, RAG, and multi-agent orchestration

Repository Map

• agents/: prompt logic and task-specific agent behavior
• core/: routing, state, orchestration, and chat service helpers
• knowledge/: Markdown corpus and vector retrieval pipeline
• app.py: recommended FastAPI entrypoint for the primary web workspace
• ui.py: secondary Gradio demo entrypoint
• docs/: release notes, launch kit, and image assets

Release Readiness

• Release notes: v0.1.0
• Changelog: CHANGELOG.md
• Roadmap: ROADMAP.md
• Launch kit: docs/launch-kit.md

Current known limitations

• API keys are required for DeepSeek and DashScope-compatible model access.
• The out-of-the-box prompts and seed data are still Chinese-first.
• A hosted demo and short walkthrough video are not included yet.

Using Your Own Knowledge Base

PolyAgent expects Markdown documents under knowledge/data/. Lightweight YAML metadata such as title, category, source, and date is recommended for better retrieval quality and traceability.

After updating your documents, rebuild the vector index:

python -m knowledge.vectorstore build

Screenshot Assets

README screenshots can be regenerated with:

pip install ".[assets]"
python scripts/generate_readme_screenshots.py
python scripts/generate_social_preview.py

The screenshot script generates both English (*-en.png) and Simplified Chinese (*-zh.png) variants. The social preview script outputs docs/images/polyagent-social-preview.png.

Tech Stack

• LangGraph for orchestration
• LangChain for LLM integration
• Chroma for vector retrieval
• FastAPI plus a custom web client for the primary workspace
• Gradio as an alternate demo interface
• DeepSeek and Qwen-compatible OpenAI-style endpoints for model access
• pytest for testing

License

MIT