This repository contains the implementation of the Smart Book Seeker: Agent-Augmented Retrieval System for Metadata-Sparse Libraries (SBS-AARS), designed to enhance book retrieval in modern libraries facing challenges due to sparse metadata.
Workflow of the Smart Book Seeker Agent-Augmented Retrieval System for Metadata-Sparse Libraries (SBS-AARS). The system operates through four coordinated steps:
- The User-Needs Agent clarifies ambiguous search requirements through interactive dialogue with the user.
- A structured user profile is constructed from the clarified needs.
- The User Agent and Librarian Agent engage in multi-turn collaborative discussions, where the Librarian Agent generates and executes keyword-based search queries against the library collection, and the User Agent evaluates retrieved books by accepting or rejecting candidates based on their alignment with user requirements.
- The final curated set of books is delivered to the user.
The dashed box highlights the core Librarian Agent-User Agent Discussion Mechanism, which enables iterative refinement of retrieval strategies through reasoning-based dialogue.
Before running the SBS-AARS (including experiments), ensure you have installed ElasticSearch for simulating realistic search scenarios and Analysis-IK for chinese text tokenization. Make sure to set up your .env file with the necessary configuration parameters before proceeding.
src/smart_book_seeker/: Main source code for the Smart Book Seeker system.src/smart_book_seeker/agents/: Contains the implementation of various LLM agents including User-Needs Agent, User Agent, and Librarian Agent.data/: Contains books for creating simulated real library scenarios.notebooks/: Scripts and configurations for running experiments.results/: Stores the results of experiments.
This project is licensed under the Apache 2.0 License. See the LICENSE file for details.