AlphaAvatar RAG Plugin

A modular Retrieval-Augmented Generation (RAG) middleware for AlphaAvatar, enabling agents to store, search, and ground answers on user-provided or locally stored knowledge.

This plugin allows AlphaAvatar agents to work with persistent knowledge bases built from files, documents, and web content—decoupling agent reasoning from raw context windows and enabling long-term, reusable memory through retrieval.

The default backend is RAGAnything, but the plugin is designed to support multiple RAG frameworks in a pluggable way.

Features

• Persistent Knowledge Indexing Ingest and store documents (PDF / Markdown / text / HTML / web snapshots) into a searchable local index.

• Grounded Answer Generation Retrieve relevant chunks and generate answers grounded in explicit sources rather than model-only knowledge.

• Unified RAG Interface Abstracts indexing and querying behind a clean, agent-friendly API, independent of the underlying RAG engine.

• Incremental & Reusable Memory Indexes can be extended, refreshed, and reused across conversations and sessions.

• Seamless Integration with DeepResearch Designed to work naturally with DeepResearch outputs (scraped pages, downloaded PDFs) for building long-term knowledge bases.

When to Use RAG

Use indexing() when:

• The user explicitly asks to save, store, remember, or archive content (e.g. “保存一下”, “留着以后查”, “做个资料库”)
• The user provides files or URLs and implies future reuse
• Multiple documents are downloaded or collected (e.g. via DeepResearch)
• The same or similar documents are referenced repeatedly across turns

Use query() when:

• The user asks questions about previously indexed content
• Answers must be grounded in user-owned data
• Searching across a growing local knowledge base is required

Functionality

It exposes two core operations (op) that can be composed into an agent workflow:

• indexing Persist files, documents, or web content into a searchable local index. Use this when the content should be saved for future retrieval.

• query Retrieve relevant chunks from an existing index and generate grounded answers. Use this when answering questions based on previously indexed content.

Typical Workflow

1. Acquire content

   • User uploads files
   • Local file paths are provided
   • Web pages are scraped or downloaded via DeepResearch

2. Decide persistence intent

   • Explicit user intent → index directly
   • Ambiguous intent → ask a clarification question

3. Index content

   • Call indexing() to build or update the knowledge base

4. Retrieve & answer

   • Call query() to retrieve relevant chunks
   • Generate grounded, source-aware answers

Installation

pip install alpha-avatar-plugins-rag

Supported RAG Frameworks

Default: RAG-Anything

Github Website

RAGAnything is a flexible RAG framework optimized for:

• Heterogeneous document formats
• Incremental indexing
• Local-first storage and retrieval
• Agent-oriented workflows

It provides the default indexing and querying backend for the AlphaAvatar RAG plugin.

Additional RAG backends can be integrated in the future without changing agent logic.