This repository contains the persistent knowledge base used in the MIRRI/TUCC case study.
The project explores a multi-agent, knowledge-based approach for working with specialized scientific metadata. Instead of asking a language model to interpret the full raw dataset at every interaction, the information is organized in advance into a navigable set of Markdown pages. Agents can then retrieve and combine only the parts of the knowledge base that are relevant to a specific question.
The goal is to support more reliable, consistent, and traceable answers over a domain that is unlikely to be fully represented in the training data of general-purpose LLMs.
The repository is centered on two main areas:
wiki/ The knowledge base used by agents for retrieval and question answering.
private/ Validation material used to evaluate agent answers and model behavior.
AGENTS.md contains the operational instructions for agents working on this repository, including how they should query the knowledge base, preserve traceability, and avoid unsupported inferences.
The wiki/ directory is the curated knowledge base. It contains interconnected pages derived from the microbiological metadata of the Turin University Culture Collection (TUCC).
The knowledge base includes information about:
- microbial strains and accession identifiers;
- taxonomy and organism type;
- habitats, substrates, localities, and geographic origins;
- cultivation conditions and growth media;
- physiological and biotechnological characteristics;
- datasets, publications, and source traceability;
- people and roles associated with collection, isolation, or deposit;
- availability, risk group, GMO, type strain, Nagoya, quarantine, and use flags.
The main entry point is wiki/index.md, which links to the major sections of the knowledge base.
The private/ directory contains validation resources used to test and compare agent behavior. These materials are separate from the knowledge base itself.
In particular, validation tasks are intended to check whether different agents can retrieve, combine, and justify answers using the same persistent knowledge base, without relying on hidden data or unsupported assumptions.
This project is licensed under the GNU General Public License v3.0. See the LICENSE file for details.