Most Retrieval-Augmented Generation (RAG) systems promise to make language models smarter by grounding them in facts. But ask them to do anything complex—like trace research funding chains or identify thematic overlaps across domains—and they break down into isolated snippets.

INRAExplorer, a project out of Ekimetrics for INRAE, dares to change that. By merging agentic RAG with knowledge graph reasoning, it offers a glimpse into the next generation of AI: systems that don’t just retrieve answers—they reason.

The RAG Plateau: Why Snippets Don’t Scale

Most RAG systems are built around one trick: fetch the top-k semantically relevant chunks using dense vectors. That works for questions like “What is photosynthesis?” But what about:

“List all INRAE projects that funded research on climate change adaptation, and the other research topics they’re involved in.”

That’s not a question—it’s a path. And classical RAG can’t walk it.

Instead, INRAExplorer shows that real-world queries require:

  • Multi-hop retrieval: chaining from authors → papers → projects → topics.
  • Exhaustiveness: not just one or two examples, but all matching entities.
  • Structured relationships: you don’t just want results—you want to know how they’re connected.

Inside INRAExplorer: Tools, Graphs, and Agentic Choreography

At the heart of INRAExplorer is a language model agent (using deepseek-r1-0528) that doesn’t act alone—it plans, chooses tools, and builds reasoning chains.

Here’s its toolkit:

Tool Role
SearchGraph Query INRAE’s Neo4j-based knowledge graph with Cypher for structured, multi-entity paths
SearchPublications Hybrid semantic + keyword search to find entry points (uses both Jina embeddings and BM25)
SearchConceptsKeywords Bridges free text queries to thesaurus terms and keywords
IdentifyExperts A high-level tool that combines sub-queries to rank domain experts on a topic

Each tool has a clear boundary and purpose, letting the agent modularize its behavior. For example, instead of “guessing” expert names from papers, the IdentifyExperts tool:

  1. Searches for top papers.
  2. Extracts authors.
  3. Scores them based on citation count, topic relevance, collaboration networks, and recency.

This layered, auditable design avoids the common trap of LLM hallucinations.

Multi-Hop in Motion: Walking the Graph

Let’s break down one of INRAExplorer’s showcased scenarios:

Query: Who at INRAE is publishing on climate change adaptation? What projects fund them? What else are those projects involved in?

The agent executes:

  1. SearchPublications to find papers on “climate change adaptation”.
  2. SearchGraph to find Author and Project nodes linked to those papers.
  3. Another SearchGraph hop from Project to other Concept nodes (via DESCRIBES) to list additional topics.

What comes back isn’t a paragraph—it’s a structured chain of reasoning. That’s not just impressive. It’s necessary for domains like science, law, or finance.

Real Deployment, Real Lessons

INRAExplorer isn’t vaporware. It runs on open infrastructure:

  • Neo4j for the KG.
  • Qdrant for hybrid search.
  • GROBID for structured PDF parsing.
  • Mirascope for agent orchestration.

And its knowledge base? Over 400K nodes and a million edges, curated from INRAE’s open-access corpus and thesauri.

But the real win is architectural:

  • Agentic planning + KG traversal gives depth.
  • Modular tools provide control and reproducibility.
  • Hybrid search offers flexibility.

This isn’t just one-off engineering—it’s a blueprint for anyone building domain-specific copilots.

Toward Evaluation and Specialization

Future work for INRAExplorer focuses on what all serious systems must face: evaluation.

  • Existing benchmarks don’t test multi-hop, structured reasoning.
  • INRAExplorer proposes domain-specific, co-designed evaluations with experts.

Another frontier: model specialization.

  • Ideas like Reinforcement Learning from Verifiable Feedback (RLVF) may help fine-tune smaller agents for graph-centric workflows.

Final Thoughts: From Agents to Analysts

INRAExplorer shows what’s possible when you treat AI like a research assistant, not a search engine. It’s not here to fetch snippets. It’s here to trace paths, identify connections, and tell you not just what—but why.

If classical RAG is about relevance, then agentic RAG is about reasoning.

And in knowledge-heavy domains, that makes all the difference.

Cognaptus: Automate the Present, Incubate the Future.