TLDR: INRAExplorer is an advanced agentic Retrieval-Augmented Generation (RAG) system that combines LLM-based agents with comprehensive knowledge graphs to overcome the limitations of traditional RAG. Designed for INRAE’s scientific data, it enables complex multi-hop reasoning, exhaustive data retrieval, and structured answers by orchestrating specialized tools interacting with a hybrid knowledge base. This system significantly enhances knowledge interaction in specialized fields, offering capabilities like identifying funding projects related to specific research topics or pinpointing experts based on their publication history and collaborations.
In the evolving landscape of artificial intelligence, Retrieval-Augmented Generation (RAG) systems have emerged as a powerful way to enhance Large Language Models (LLMs) by grounding them in specific, curated information. However, traditional RAG approaches often face significant challenges when confronted with complex queries that demand more than just simple, extractive answers. These limitations include struggling with multiple targeted retrievals, navigating intricate relationships between entities, or providing exhaustive lists of information.
Addressing these critical gaps, especially in knowledge-intensive fields, researchers have introduced INRAExplorer. This innovative system is an agentic RAG solution designed specifically for exploring the vast scientific data of INRAE, France’s National Research Institute for Agriculture, Food and Environment. INRAExplorer stands out by employing an LLM-based agent with a sophisticated multi-tool architecture that dynamically interacts with a rich knowledge base, built upon a comprehensive knowledge graph derived from open-access INRAE publications.
Beyond Traditional RAG: The INRAExplorer Approach
Classical RAG systems typically retrieve a limited number of text snippets based on semantic similarity to provide context to an LLM. While useful for basic question answering, this method falls short for queries requiring extensive lists, synthesis from multiple distinct data points, or traversing complex relational paths—such as finding all publications by an author and then identifying the funding projects for those publications.
INRAExplorer overcomes these limitations by synergizing two advanced concepts: agentic RAG and Knowledge Graph (KG)-enhanced RAG. Agentic RAG allows for dynamic reasoning, enabling the system to plan and execute a series of actions. KG-enhanced RAG provides structured, exhaustive retrieval capabilities. The core innovation lies in deeply integrating KG querying as a fundamental agentic capability, allowing INRAExplorer to move beyond the single-pass, limited-context nature of classical RAG. This fusion enables precise, relationally-aware retrieval from KGs combined with the adaptive, multi-hop reasoning of an LLM-driven agent.
The Foundation: A Hybrid Knowledge Base
The power of INRAExplorer stems from its comprehensive knowledge base, constructed from INRAE’s scientific output, specifically open-access publications from January 2019 to August 2024. This data is meticulously gathered from various sources like HAL and OpenAire, and enriched with details from BBI, ScanR, and dataset repositories. Full-text content from PDF publications is processed to extract structured information, forming meaningful chunks.
A key component is the integration of the INRAE Thesaurus into the Knowledge Graph. This thesaurus provides a controlled vocabulary, linking publications to specific ‘Concept’ and ‘Domain’ nodes, which aids in semantic exploration and understanding domain-specific queries.
The processed information is stored in a hybrid knowledge base:
-
Vector Database: Stores representations of textual chunks from publications. It uses both dense vectors for semantic similarity and sparse vectors for keyword-based matching, supporting a hybrid search approach.
-
Knowledge Graph: Models core entities and their relationships. It includes various node types such as ‘Publication’, ‘Author’, ‘Keyword’, ‘Concept’, ‘Domain’, ‘Journal’, ‘Project’, ‘Software’, ‘ResearchUnit’, ‘Dataset’, and ‘Region’. This KG, with over a million relationships, is the backbone for precise, structured queries and multi-hop reasoning.
Agent-Driven Multi-Tool Orchestration
At the heart of INRAExplorer is an LLM-based agent (utilizing the deepseek-r1-0528 model) responsible for understanding user queries, breaking them down, formulating action plans, dynamically selecting and invoking appropriate tools, and synthesizing information into coherent responses. The agent has access to a suite of specialized tools:
-
SearchGraph: The primary tool for deep interaction with the Neo4j Knowledge Graph. It allows the agent to execute Cypher queries to retrieve specific entities, traverse complex relationship paths, and gather exhaustive lists (e.g., all publications by a specific author).
-
SearchPublications: A helper tool that interfaces with the vector database for hybrid searches over publication texts. It helps identify relevant documents to initiate further exploration via the KG.
-
SearchConceptsKeywords: Another helper tool that bridges user queries with the structured vocabulary of the knowledge base by searching for relevant concepts from the thesaurus or author-declared keywords.
-
IdentifyExperts: A refined tool that encapsulates complex domain-specific logic for identifying experts on a given topic. It uses other tools like SearchPublications and SearchGraph to analyze authors, their citations, collaboration networks, and project involvement to synthesize a ranked list of experts.
Illustrative Scenarios: Multi-Hop Reasoning and Modular Expertise
INRAExplorer excels in scenarios where classical RAG systems falter. For instance, consider the query: “Find INRAE authors who have published on ‘climate change adaptation strategies’, identify the projects that funded these publications, and list other key topics these funding projects are related to.”
The INRAExplorer agent tackles this with multi-hop reasoning:
-
It first uses SearchPublications or SearchConceptsKeywords to find initial relevant publications and authors related to ‘climate change adaptation strategies’.
-
Then, using SearchGraph, it identifies ‘Project’ nodes linked to these publications via a ‘FUNDED_BY’ relationship (the first hop).
-
Subsequently, for each identified project, it again employs SearchGraph to find other ‘Concept’ nodes linked to these projects, broadening the understanding of the projects’ thematic scope (the second hop).
-
Finally, the agent synthesizes these interconnected findings into a structured answer, explicitly showing the chain: authors → publications → funding projects → other related topics.
Another powerful capability is demonstrated by the IdentifyExperts tool. For a query like “Identify leading INRAE experts on ‘zoonoses’,” the agent directly invokes this specialized tool. The IdentifyExperts tool then executes a predefined workflow: finding relevant papers, extracting authors, and calculating a composite expertise score based on various metrics like article relevance, citation counts, and activity period. This modularity ensures consistent, domain-aware execution for complex tasks.
Also Read:
- KROMA: Enhancing Ontology Matching with Context-Aware Language Models
- WebShaper: A Formalized Approach to Generating Training Data for Information-Seeking AI Agents
Looking Ahead
INRAExplorer represents a significant step forward in applying AI techniques to real-world scientific knowledge exploration. Its open design, leveraging components like Mirascope for agent orchestration, Qdrant for vector storage, and Neo4j for the knowledge graph, ensures adaptability and extensibility. Future work will focus on developing a tailored evaluation framework for such complex, multi-hop scientific queries, and exploring techniques like Reinforcement Learning from Verifiable Feedback (RLVF) to further refine the core agent model for even more robust performance.
To learn more about this innovative system, you can read the full research paper here.
