AI-RAG-docuquery: A Retrieval-Augmented Document Query System
Abstract
I present AI-RAG-docuquery, an open-source system for retrieval-augmented question answering (QA) over heterogeneous document collections. The tool integrates dense vector search (FAISS), sentence-transformer embeddings, and configurable large language model (LLM) backends into a unified desktop application. It enables users to index local corpora of PDFs, DOCX, PPTX, XLSX, TXT, CSV, or Markdown files and obtain grounded, source-cited answers in response to natural-language queries. The project is released under the MIT License and is available on GitHub.
System Overview
AI-RAG-docuquery is designed as a lightweight, reproducible framework for applied Retrieval-Augmented Generation (RAG). The architecture is composed of three main layers:
Indexing and Retrieval
Documents are pre-processed via loaders for multiple formats.
Embeddings are computed using Sentence-Transformers (MiniLM or E5 family).
Vectors are stored in FAISS indexes, supporting dense retrieval; hybrid sparse/dense retrieval is optionally enabled.
Index metadata (filenames, page numbers, chunk offsets) is stored in JSONL for citation alignment.
Platform: Windows 10/11 (tested); macOS/Linux supported from source
Index storage:faiss_index/ containing FAISS binaries, JSONL metadata, and serialized configs
Build options:
--onefile PyInstaller for portable .exe
--onedir PyInstaller for robustness with shared libraries
Use Cases
Personal document QA: knowledge workers can query technical manuals, research papers, or organizational archives.
Educational support: students can index lecture slides, notes, and reference material.
Enterprise knowledge management: potential to extend with larger sharded FAISS indexes.
Availability
The full project, including source code, requirements, and build instructions, is publicly available on GitHub:
š AI-RAG-docuquery (GitHub Repository)
License: MIT
Latest release: v2.0 (2025-08-27)
Roadmap
Planned future work includes:
Multi-index management (create, merge, and query across multiple FAISS indexes).
Scalable indexing (on-disk FAISS, sharding for >10M vectors).
Additional LLM adapters (Azure-OpenAI, Anthropic, Google Gemini, Mistral).
Improved hybrid ranking and passage de-duplication.
Conclusion
With AI-RAG-docuquery I demonstrate how retrieval-augmented generation can be applied to everyday document collections in a transparent and reproducible way. By unifying vector retrieval, flexible LLM integration, and verifiable citations into a desktop-ready package, the project bridges the gap between academic RAG prototypes and practical end-user tools.
Table of contents
AI-RAG-docuquery: A Retrieval-Augmented Document Query System