piFlow – A Comprehensive Local RAG Desktop Application for Enhanced Retrieval

piFlow is an innovative desktop application designed for Retrieval-Augmented Generation (RAG). This application empowers users to construct a knowledge base by importing local documents, facilitating interactions through natural language queries. Every aspect of parsing, embedding, retrieval, and reasoning operates locally on your machine or a designated server, ensuring that both documents and conversations remain under your control. Each generated answer includes source citations, allowing for easy traceability back to specific documents and paragraphs.

Key Features

• Traceable Answers: All responses come with references and direct excerpts, providing a clear path back to the source document and specific sections.
• Privacy-First Approach: All documents, vector indices, and chat history are securely stored in a local SQLite database, eliminating any reliance on cloud services.
• Directory-Level Import: Effortlessly select a directory to recursively import documents while a background task runs silently, with progress logs displayed in real-time.
• Intelligent Retrieval Pipeline (reRAG): Features an intent template routing system directing LLM-generated structured retrieval plans, dynamically merging top-K results based on intent, and supporting multi-round co-reference resolution.
• Multi-Format Parsing: Supports a variety of formats including PDF (via pdf-oxide/mupdf), Markdown, TXT, and HTML, with automatic routing of scanned PDFs to PaddleOCR.
• Pluggable Generation Backend: Offers multiple options for local GGUF inference, remote Ollama solutions, and integration with DeepSeek or Pleias-RAG-1B.
• Automatic Model Management: Contains a model registry based on a manifest; missing models can be automatically downloaded from hf-mirror.com and re-downloads can be enforced as needed.
• piFlow Agent: Serves as the main conversation interface; it can toggle between Knowledge RAG, PostgreSQL read-only access, and local file skills.

Architecture Overview

piFlow utilizes a dual-process design leveraging Tauri 2 and a Node Sidecar. The Tauri layer acts as a lightweight UI shell, while all RAG inference and native modules execute within an independent Node process. Communication between the front-end and back-end is made via localhost HTTP/SSE:

┌────────────────────────────────────────────────────┐
│         Tauri WebView (React + TypeScript)          │
│   Knowledge Base · piFlow (Main Interface)· Citation · Settings │
└──────────────────────┬─────────────────────────────┘
                       │ HTTP / SSE (localhost)
┌──────────────────────▼─────────────────────────────┐
│            Node Sidecar (apps/rag-server)           │
│                                                     │
│  RAG: Import → Parse → Chunk → BGE-M3 → Retrieve → Generate   │
│  piFlow: Pi Agent + Skills (KB / Postgres / FS)    │
│                                                     │
│  SQLite · pg-actions · Model cache · Files         │
└─────────────────────────────────────────────────────┘

Technical Stack

The technical foundation of piFlow includes:

Documentation

A detailed overview of the architecture and guidance can be found in the extensive documentation, covering:

• System Architecture - Detailed system design, module division, and platform strategy.
• piFlow Design - Insights into Agent, Skill, SSE, and source strategies.
• User Manual (Chinese) - Installation, RAG usage, and FAQs.
• Further technical details are encapsulated in various documents including guides for development, model registration, and advanced features.

Roadmap

• Phase 1 – MVP Development: Establishing the monorepo scaffold, RAG sidecar, document import functionality, answer citation, and model management.
• Phase 2 – Experience Enhancement: Incorporating advanced document handling, mixed retrieval strategies, and a visualization tool for inference pathways.
• Phase 3 – Windows Enhancements: Optimizing package size, offering optional OCR capabilities, and CUDA detection.
• Phase 4 – Advanced Features: Multiple isolated knowledge bases, management of multi-turn contexts, and pluggable document sources.