GraphMind Research: Building an Intelligent Research Assistant with FalkorDB
Research papers contain valuable insights, but finding connections between methods, datasets, results, and citations can be difficult. GraphMind Research combines Graph RAG, FalkorDB, semantic search, and collaborating AI agents to help researchers explore relationships and discover evidence-backed insights.
GraphMind Research: Connecting Research Papers Through Graph RAG and FalkorDB
Introduction
Research papers are a valuable source of scientific knowledge. However, understanding how different papers connect to one another can be challenging. Researchers often need to identify related methods, compare experimental results, explore citation relationships, and understand how one study builds upon another.
Traditional keyword-based search can find documents containing similar terms, but it may not fully explain the relationships between papers, authors, methods, datasets, and results.
To address this challenge, we built GraphMind Research, an AI-powered research exploration system that combines Graph Retrieval-Augmented Generation (Graph RAG), FalkorDB, semantic search, and multiple collaborating AI agents.
Our goal was to make research exploration more connected, explainable, and evidence-focused.
The Problem We Wanted to Solve
Imagine reading a research paper that introduces a new machine learning method. You might want to know:
- Which earlier papers influenced this method?
- Which datasets were used to evaluate it?
- What results did the authors report?
- Are other methods similar to it?
- Does a later paper extend or challenge its findings?
Finding these answers manually requires reading multiple papers and following their references.
A conventional search system primarily retrieves relevant documents. We wanted to go further by connecting the information inside those documents and using the relationships to guide the retrieval process.
Our Solution: GraphMind Research
GraphMind Research converts research papers into a connected knowledge graph.
Instead of treating every paper as an isolated document, we represent important entities and their relationships explicitly. This allows the system to explore connections and retrieve relevant passages from the papers associated with those connections.
The project combines two complementary approaches:
Knowledge graphs: Represent entities such as papers, authors, methods, datasets, results, and claims, along with relationships between them.
Semantic search: Retrieves relevant text passages based on their meaning rather than relying only on exact keyword matches.
By combining these approaches, GraphMind Research can use graph relationships to guide evidence retrieval and generate answers grounded in retrieved research content.
Why We Chose FalkorDB
FalkorDB is a graph database designed to store and query connected data. It is central to our architecture because the relationships between research entities are as important as the entities themselves.
Our graph contains entities such as:
- Paper
- Author
- Method
- Dataset
- Result
- Metric
- Claim
We represent relationships such as authors writing papers, papers citing other papers, methods being evaluated on datasets, and results being associated with methods and metrics.
For example, a paper may introduce a method, evaluate it on a dataset, report a metric, and cite earlier research. These connections can be queried to explore how ideas and results relate across papers.
We also use graph algorithms to analyse the research network:
- PageRank helps identify influential papers in the citation network.
- Betweenness centrality helps identify authors who connect different parts of the collaboration network.
- Weakly connected components help reveal groups of connected entities.
FalkorDB also coordinates the agent workflow through shared job nodes and a lock mechanism. This makes the graph database more than a place to store extracted research information: it also supports communication between our agents.
How Our Architecture Works
Our system combines a research-processing pipeline with a question-answering workflow.
1. PDF Parsing
The Parser agent processes research PDFs and extracts structured information, including paper details, authors, citations, and table results.
We represent extracted information in the graph and preserve research content that can later support retrieval.
2. Information Extraction
The Extractor agent identifies relevant research information and works on extracting claims and citation intent.
This stage transforms unstructured paper content into information that can be connected and queried.
3. Entity Linking
The Linker agent helps connect related information across papers. For example, it can identify different references to similar methods and create relationships between them.
This step helps reduce the problem of treating related research concepts as completely separate entities.
4. Critic and Relationship Checking
The Critic agent evaluates selected claims and candidate relationships. It uses an LLM to classify relationships and examine selected state-of-the-art claims.
This stage is important because automatically extracted information can be incomplete or incorrect. We designed the Critic to help assess these relationships, although our evaluation showed that this area still needs improvement.
5. Semantic Indexing
We use ChromaDB to store and retrieve semantically relevant text passages. The indexing process uses embeddings so that the system can retrieve passages based on meaning.
This complements FalkorDB: the graph represents relationships, while ChromaDB helps locate supporting text.
6. Research Question Answering
When a user asks a research question, a LangGraph-based workflow follows a sequence of steps:
- Understand the question.
- Plan how to investigate it.
- Execute relevant graph queries.
- Retrieve supporting passages from the selected papers.
- Check the available evidence and, when needed, perform one additional retrieval round.
- Generate an answer and record research-session information.
The aim is to make answers more useful by combining connected research knowledge with supporting passages rather than relying only on a language model's general knowledge.
Technology Stack
We selected the following technologies for different parts of the system:
- FalkorDB: Knowledge graph storage, graph queries, agent coordination, and graph analysis.
- Python: Core application logic and research-processing workflows.
- FastAPI: Backend API.
- LangGraph: Orchestration of the research question-answering workflow.
- ChromaDB: Semantic indexing and text retrieval.
- Ollama: Local execution of language and embedding models.
- React and Vite: Frontend application.
- Tailwind CSS: User interface styling.
- React Flow: Graph-oriented visualisation.
This combination allowed us to build a system that connects structured relationships, semantic retrieval, AI workflows, and an interactive frontend.
Evaluation: What Did We Learn?
We evaluated the system using a hand-built ground-truth benchmark containing 12 synthetic research papers.
The results helped us understand which parts of the pipeline were working well and which required more development.
Results from our evaluation
- Citations found: 49 out of 49.
- Table results extracted: 70 out of 70.
- EXTENDS relationships: 4 out of 4.
- SIMILAR_TO relationships: 1 out of 1.
- Citation-intent accuracy: 0.918 using graph rules and cue words.
For citation-intent classification, our combined graph-rule and cue-word approach performed better than the alternatives we tested. In our evaluation, the LLM-only approach achieved 0.735 accuracy, while the cue-word-only approach achieved 0.551.
However, the evaluation also revealed important weaknesses.
Claim extraction achieved a precision of 0.364 and a recall of 0.615. For claim relationships, only 1 of 9 expected relationships was found, and the predicted relationships had precision below 0.07.
These results show that extracting straightforward structural information, such as citations and table results, was more reliable than identifying complex scientific claims and their relationships.
The benchmark is small and synthetic, so these results should not be interpreted as proof of production-level accuracy. Testing on a larger and more diverse collection of real research papers would be an important next step.
Challenges We Faced
Extracting reliable scientific claims
Research claims are often expressed in complex language and depend on context. Identifying the main claim is not the same as extracting a sentence, and identifying whether one paper supports, contradicts, or extends another requires careful reasoning.
Our evaluation showed that claim extraction and relationship classification need further improvement.
Connecting similar entities
Different papers may refer to the same method using different names or descriptions. Entity linking must distinguish genuine matches from concepts that are merely related.
Balancing graph structure and semantic retrieval
A graph can identify which papers and entities are connected, but it does not automatically provide all the text needed to answer a question. Semantic retrieval supplies that additional evidence.
Combining both approaches requires careful retrieval planning and evidence checking.
Coordinating multiple agents
The Parser, Extractor, Linker, and Critic agents have different responsibilities. Coordinating their work through shared graph state requires a clear data model and a reliable workflow.
What We Learned
Building GraphMind Research taught us several important lessons.
First, graph structure adds context. Relationships between papers, methods, datasets, and results can support exploration that is difficult with isolated documents alone.
Second, Graph RAG is more than adding a graph to a chatbot. The graph must actively help determine what information to retrieve and how that information supports an answer.
Third, evaluation matters. A system may appear to work on a few examples while still producing unreliable relationships. Comparing outputs against ground truth helped us identify weaknesses that were not obvious from demonstrations alone.
Finally, specialised agents need measurable responsibilities. Separating parsing, extraction, linking, and checking makes the workflow easier to reason about, but the overall quality still depends on the accuracy of each stage.
Future Improvements
There are several areas we would like to improve:
- Increase the accuracy of scientific claim extraction.
- Improve the precision of relationships between claims across papers.
- Evaluate the system on a larger and more diverse research-paper collection.
- Expand the evidence-verification process.
- Improve entity resolution for methods and datasets.
- Measure retrieval quality and answer faithfulness with a broader evaluation suite.
These improvements would help us assess how well the system generalises beyond our current benchmark.
Conclusion
GraphMind Research explores how knowledge graphs and semantic retrieval can work together to make research papers easier to investigate.
By using FalkorDB to represent and query relationships, ChromaDB to retrieve relevant passages, and collaborating AI agents to process research information, we built a prototype that connects information across papers and supports evidence-focused question answering.
Our evaluation demonstrated promising results for citation and table-result extraction, while also revealing clear limitations in claim extraction and relationship classification.
For us, the most important takeaway is that useful research AI should do more than generate plausible answers. It should help users discover meaningful connections and make those answers traceable to supporting evidence.
Team
GraphMind Research was built by:
- Pooja V
- Gomathi V
- Niveditha S
AI Assistance Disclosure
AI coding assistants were used during development to support design, planning, and implementation. Claude was used for design guidance and step-by-step instructions, and Claude Code assisted with applying those instructions. The project was reviewed and tested by the team.
We believe transparent disclosure of AI assistance is important when sharing technical work and documenting how a project was built.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article