AquiLLM

An open source RAG-LLM for preserving tacit knowledge in research groups

AquiLLM (pronounced ah-quill-em) helps researchers manage, search, and interact with their research materials. The goal of AquiLLM is to enable teams to access and preserve their collective knowledge, and to enable new group members to get up to speed quickly.

View on GitHub Learn More

About AquiLLM

  • Research groups struggle to capture and retrieve knowledge distributed across team members.
  • Much of this knowledge is informal—emails, notes, and conversations—and remains fragmented or undocumented.
  • This constitutes tacit knowledge: experience-based expertise central to research practice.
  • Accessing it is time-consuming and requires contextual familiarity.
  • Existing RAG systems often do not address the necessary privacy of internal research materials.
  • AquiLLM is an open-source, modular RAG-LLM system designed for research groups.
  • It supports diverse data types and configurable privacy for both formal and informal knowledge.

Designed for Open-Weight Models and Local Systems

Supports local deployment using open-weight models, avoiding dependence on external APIs. This allows research groups to control data, model selection, and inference within their own computing environments.

System Architecture

AquiLLM presents a modular architecture designed to address fundamental challenges in research knowledge management. Our system integrates document processing pipelines, information retrieval mechanisms, contextual memory systems, and locally-deployed language models to support scholarly inquiry and collaborative research workflows.

System architecture diagram illustrating the modular components of AquiLLM: researcher interfaces, web-based user interface, Django backend with WebSocket channels, document ingestion and indexing pipeline, retrieval and reranking subsystems, memory management, shared data repositories, and local vLLM inference services.
Figure 1. AquiLLM system architecture demonstrating the integration of core research support components: document ingestion and semantic indexing, query-time retrieval with reranking mechanisms, persistent memory for conversational context, and local language model inference. The modular design enables researchers to maintain data sovereignty while leveraging advanced natural language processing capabilities for knowledge discovery and synthesis across heterogeneous information sources. (Figure created by Jack Stark)

Research Publication

The methodological foundations, system implementation, and empirical evaluation of AquiLLM are detailed in our peer-reviewed research:

Campbell, C., Boscoe, B., & Do, T. (2024). AquiLLM: a RAG Tool for Capturing Tacit Knowledge in Research Groups. Proceedings of the US Research Software Engineer Association Conference (US-RSE 2025). arXiv:2508.05648

This work presents our approach to addressing knowledge management challenges in collaborative research environments through the development of a privacy-aware, locally-deployable retrieval-augmented generation system.

Contributors and Collaborators

Principal Investigators

  • Prof. Bernie Boscoe (Southern Oregon University)
  • Prof. Tuan Do (UCLA)

Developers

  • Chandler Campbell
  • Jack Stark

Contributors

  • Amy Cheatle (Cornell University)
  • Zhuo Chen (University of Washington)
  • Andrew Lizarraga (UCLA)
  • Jonathan Soriano (UCLA)
  • Morgan Himes (UCLA)
  • Srinath Saikrishnan (UCLA)
  • Jacob Nowack (Southern Oregon University)
  • Jackson Godsey (Southern Oregon University)
  • Tee Grant (Southern Oregon University)

Former Contributors

  • Skyler Acosta (Southern Oregon University)
  • Kevin Donlon (Southern Oregon University)
  • Elyjah Kiehne (Southern Oregon University)

About the Name

AquiLLM Logo

AquiLLM is a combination of the words "Aquila," the constellation, and "LLM," which stands for Large Language Model. Aquila is one of the most prominent constellations in the northern sky. The name reflects our group's history in working with Astronomy and Astronomers.

Our work is partially supported by:
Alfred P. Sloan Foundation Alfred P. Sloan Foundation
National Science Foundation National Science Foundation
We are grateful for the support of our funding agencies.
Cloud credits for research and development have been generously provided by NSF NAIRR Pilot, ACCESS CI, Gemma Academic Program, and CloudBank.
Special thanks to folks at Jetstream2 for their computational support.