AquiLLM: An Architecture for Supporting Tacit Knowledge Capture in Research Groups

Imagine a group of scientists working together in a lab, like a band jamming in a garage. They have their sheet music (published papers) and their instruments (data), but the real magic happens in the messy, unrecorded moments: the inside jokes, the "don't touch that knob" warnings, and the clever shortcuts they invent on the fly. This is called tacit knowledge—the stuff you know because you've done it, not because you read it. Usually, when a band member leaves, that secret knowledge walks out the door with them.

To solve this, scientists are trying to use Artificial Intelligence (AI) to act as a super-memory for these groups. Think of AI as a very fast, very smart librarian who can read every book in the library instantly. However, most of these "librarians" are owned by big companies. This creates a problem: if you ask a company's librarian a question, they might keep a copy of your question to train their next librarian, or they might not show you exactly how they found the answer. For scientists, who need to be able to check their work and keep their data private, this is a no-go. They need a librarian they can build themselves, keep in their own basement, and trust completely.

This is where AquiLLM comes in. It's a new, open-source tool designed specifically for research groups to capture that messy, unspoken knowledge. The paper describes how the team upgraded AquiLLM to be more like a helpful, local research assistant that lives on the group's own computers. Instead of sending data to the cloud, it uses Retrieval-Augmented Generation (RAG). Imagine RAG as a system where the AI doesn't just guess answers from its own brain; instead, it first runs to a specific bookshelf, grabs the exact pages it needs, and then reads them out loud to answer your question. This ensures the AI sticks to the facts the group actually has, rather than making things up.

The authors of this paper didn't just build a basic version; they gave it a major makeover based on what real scientists (like astrophysicists) actually needed. They added features that let the system "see" images and charts, not just read text, and gave it a "memory" that remembers who you are and what you've talked about before, without mixing up different projects. They also made sure the system is modular, meaning you can swap out parts like the brain or the eyes without rebuilding the whole robot.

The paper suggests that these upgrades make the system much better at handling the complex, visual, and collaborative nature of real science. However, the authors are careful to note that while the system works well for finding facts in documents, it still struggles a bit when asked to compare information across many different sources or when the "memory" gets too full. They haven't proven it's perfect yet; they've just shown that it's a promising step toward AI that respects privacy and helps teams keep their collective wisdom safe.

The Big Upgrade: From a Clunky Robot to a Smart Lab Partner

So, what exactly did the team do to AquiLLM? Think of the original system as a helpful but slightly clumsy robot. It could talk, but it was slow, it couldn't see pictures, and it sometimes forgot what you told it five minutes ago. The new version is like giving that robot a brain transplant, new eyes, and a personal notebook.

1. The "Local Brain" and the "Local Library"
The biggest change is that everything happens right on the group's own computers. The team replaced the need for expensive, external commercial services with open-weight models. Imagine instead of calling a famous, expensive consultant to solve a problem, you have a brilliant local tutor who lives in your house. This tutor (the AI model) is free to use, and you own all the notes they take. The team specifically chose models like Qwen3.6-27B for chatting and Qwen3-VL for seeing images. They also swapped out external search tools for local ones, meaning the "search engine" part of the AI also lives in the group's basement, not in a corporate cloud. This keeps sensitive data, like unpublished research or private notes, completely off the internet.

2. Giving the AI "Eyes" (Multimodal Capabilities)
Science isn't just about words; it's about graphs, photos of stars, and diagrams. The old system struggled with these. The new AquiLLM can now "see." It uses a multimodal model that can look at an image of a graph, understand what the lines mean, and talk about it just like a human would. It's like upgrading from a robot that can only read a menu to one that can look at the food on the plate and tell you how it tastes. This is huge for researchers who need to discuss figures from their papers without having to describe every single pixel in text.

3. The "Personal Memory" vs. The "Shared Notebook"
One of the coolest new features is a memory layer. Imagine you have a conversation with a friend. You don't want to re-explain your entire life story every time you say "hello." The new system has two types of memory:

• Stable Facts: This is like a permanent ID card. It remembers your name, your favorite research topic, and your tone.
• Episodic Memory: This is like a diary of your recent chats. If you asked about a specific star cluster yesterday, the AI remembers that context today.
  Crucially, this memory is private to the user and the project. It's not mixed up with everyone else's data. The paper notes that this memory is optional and must be turned on, because storing too much can get messy. If the memory fails, the system just falls back to the documents, so it doesn't break the whole conversation.

4. "Skills" and "Skill Packs"
This is a clever way to teach the AI how a specific team works. Imagine a new intern joining a lab. They need to know the specific rules: "Always plot data in blue," or "Check this checklist before running the experiment." In the new system, these rules are bundled into Skills. When a researcher starts a chat about a specific project, the AI automatically loads the "Skill Pack" for that project. It's like the AI puts on a different hat depending on which team it's talking to. A spectroscopy team gets one set of instructions, and a photometry team gets another, all without the AI getting confused.

5. The "Smart Librarian" (Context Packing)
AI has a limit on how much it can read at once (its "context window"). If you give it a whole library, it might get overwhelmed. The team added a salience-aware context packing system. Think of this as a super-efficient librarian who knows exactly which pages of a book are most important for the question you're asking. Instead of just cutting off the end of the book randomly, the system ranks the information: "This sentence is critical," "This image is relevant," "This old chat message can be summarized." It uses a tool called LLMLingua-2 to shrink the text without losing the meaning, allowing the AI to fit more important info into its brain without crashing.

The Reality Check: What's Still Tricky?

The authors are very honest about what this system can't do yet. They don't claim it's a magic wand.

• It's a Hybrid: The system is still a mix of old and new code. Some parts work perfectly on local computers, while others are still set up for the cloud. This can cause some weird glitches, like the system behaving differently depending on which "mode" it's in.
• Memory Lag: Because the memory updates happen in the background (to keep the chat fast), there's a tiny delay. If you ask a question right after you told the AI something new, it might not have "heard" you yet.
• The "Tacit" Test: The team tested the system on formal documents like astronomy papers, and it did a good job. But they admit they haven't fully tested it on the messier stuff yet, like Slack messages, meeting minutes, or code comments. They suggest that capturing this kind of informal, "tacit" knowledge is the next big challenge.
• Speed vs. Smarts: The system is fast, but sometimes it has to choose between speed and accuracy. If the computer runs out of memory (VRAM), it has to slow down or drop some information.

The Bottom Line

AquiLLM is a significant step forward for scientific research. It moves away from the idea that AI must be a black box owned by a corporation and brings it back into the hands of the researchers. By making the system local, multimodal, and memory-aware, it respects the privacy and workflow of science. The paper suggests that while it's not perfect, it's a powerful tool for keeping a research group's collective brain intact, even as members come and go. It's not a solved problem, but it's a very promising prototype for the future of collaborative science.