Two-way-RAG：打造本地语音交互式文档知识库系统

Two-way-RAG is a local voice-interactive Retrieval-Augmented Generation (RAG) system built with FastAPI, LangChain, and Ollama. Its core features include:

1. Privacy-first design: All data stays local (uses Ollama for local Llama3.2 model and FAISS for semantic search).
2. Voice interaction: Supports voice input (via Web Speech API) and voice output (via gTTS).
3. Private document Q&A: Enables natural dialogue with personal documents without data leakage. This system addresses the need for secure, local AI-powered document knowledge bases.

Project Background

In the era of LLM development, how to let AI understand private documents while ensuring data privacy is a key concern. Two-way-RAG solves this by providing a fully local deployable voice-interactive RAG system.

Core Concept

• Localization first: Uses Ollama to run Llama3.2 locally, FAISS for efficient semantic retrieval (all data remains on user's machine).
• Voice interaction: Allows voice input and output for natural dialogue experience.

Technical Stack

• Backend: FastAPI (high-performance async web framework).
• RAG Pipeline: LangChain (componentized design for document loading, text splitting, embedding, retrieval).
• Vector Storage: FAISS (for fast similarity search).
• Embedding Model: all-MiniLM-L6-v2 (lightweight yet effective sentence embedding).
• Voice Interaction: Web Speech API (STT) and gTTS (TTS) for bidirectional voice support.

Core Functions

1. Flexible document access: Preload from pre_trained_data directory or upload PDF/TXT dynamically.
2. Smart dialogue handling: Direct LLM response for greetings/chats; RAG for knowledge queries.
3. Session history: Auto-saved via LocalStorage.
4. Reinitialize: One-click rebuild of knowledge base after updating pre-trained docs.

Usage Scenarios

Applicable for batch initialization and incremental updates of knowledge bases.

Prerequisites

• Python 3.9+
• Ollama installed with llama3.2:latest model.

Installation Steps

1. Clone the repository.
2. Create and activate a virtual environment.
3. Install dependencies (FastAPI, LangChain, FAISS, Sentence Transformers, gTTS, etc.).

Startup

Run uvicorn main:app --reload; access via http://localhost:8000 (responsive interface with chat history sidebar).

Document Processing Phase

When starting or uploading docs: Split text into chunks → generate embeddings via all-MiniLM-L6-v2 → store in FAISS index.

Query Processing Phase

• Intent recognition: For knowledge queries, convert to vector → search FAISS for similar chunks.

Answer Generation Phase

Combine retrieved chunks with query into prompt → send to local Llama3.2 → generate answer (all local, fast response).

Application Scenarios

1. Researchers: Personal literature assistant for quick retrieval/summary of academic papers.
2. Enterprises: Internal knowledge base for employees to query company docs/manuals.
3. Developers: Learning example for RAG tech stack (clear code structure, detailed comments).

Practical Value

Represents the trend from general models to specialized systems → more accurate answers in specific domains, reduces hallucinations.

Summary

Two-way-RAG is a well-designed open-source RAG system combining voice interaction, local LLM inference, and semantic retrieval. It provides a practical private knowledge base solution with high code quality and comprehensive documentation.

Outlook

As local LLM performance and vector DB technology evolve, such systems will become more powerful and user-friendly. Ideal for users focusing on data privacy and local AI usage.