September 2026 Summaries

Neo4j’s second enterprise knowledge-layer chapter describes modeling business processes as queryable graphs rather than static BPMN diagrams, using AcmeBank’s SMB loan approval workflow to show how organizations can identify ownership, task sequences, handoffs, escalations, and automation opportunities. Process knowledge is drawn from procedure documents, workflow and ticket histories, and employee expertise, but extracted information is treated as a proposed draft with provenance until validated by human process owners. The model organizes work into BusinessProcess, BusinessActivity, and BusinessTask nodes, connecting tasks through conditional NEXT relationships and linking them to stable roles rather than individual employees, while separate ownership, handoff, and escalation relationships capture accountability and responsibility changes. Built on the organizational graph from the prior chapter, the example enables questions such as who owns a process today and historically, what its default and exception paths are, which tasks can be automated, and where cases escalate when normal decision rules do not apply. The accompanying Neo4j implementation includes schema, sample data, and Cypher queries, while the next chapter will address domain ontology to define the business terms, data fields, systems, and rules that process tasks rely on.

Contextual retrieval is an approach for AI agents that combines semantic or full-text search with graph traversal to find information connected by entities, relationships, conversation history, and prior decisions rather than relying only on similar wording. It uses context graphs as persistent memory stores for long-term enterprise knowledge, short-term task state, and reasoning traces such as tool calls and outcomes, enabling GraphRAG workflows to start from relevant nodes and follow multi-step relationships. Compared with vector-only RAG, the approach is intended to improve relevance for complex questions, support multi-hop reasoning, provide inspectable decision paths, reduce hallucinations through better grounding, and limit token use by retrieving only task-relevant context instead of replaying complete histories. The article illustrates these capabilities through customer-support and operations examples, cites studies reporting improved truthfulness and fewer hallucinations, and presents Neo4j Agent Memory as a tool for implementing context graphs in agent systems.

Neo4j’s Startup Program requires approved applicants to complete several manual steps before Aura credits are activated: creating or signing into an Aura Console account directly, locating the organization ID, adding a payment method, emailing the organization ID to Neo4j, and redeeming the emailed activation code in the console’s billing section. A payment method is necessary to release credits, though charges apply only after credits are exhausted, and applicants may use a card or AWS, Azure, or GCP Marketplace billing. Credits generally last 12 months beginning when the organization ID is submitted rather than when the application is approved, while extended inactivity may require reapplication. The initial credit amount is fixed, but startups may qualify for higher tiers through investor verification or public funding announcements. The guidance also addresses rebrands, advising companies to retain active original applicant email addresses for notifications, and duplicate applications, noting that only one application can be activated per company.

기업 AI의 성패는 최신 언어 모델이나 에이전트 도구 자체보다, 에이전트가 기업의 고유한 데이터 의미·관계·정책·업무 맥락을 이해할 수 있는지에 달려 있다는 관점을 제시한다. 이를 위해 글은 각 에이전트에 지식을 개별적으로 내장해 의미가 중복·불일치하는 방식을 피하고, 온톨로지, 엔터프라이즈 데이터, 실행 메모리로 구성된 공유형·거버넌스형 지식 계층(KL)을 구축해야 한다고 주장한다. 지식 계층은 요청 의도를 비즈니스 용어로 해석하고, 신뢰할 데이터 소스를 선택하며, 접근 정책을 실행 시점에 적용하고, 의사결정 과정을 기록해 이후 작업을 개선하는 실행 가능한 기반으로 설명된다. 기업별 지식 계층은 완제품으로 구매할 수 없으며, 그래프 기반 기술과 LLM을 활용해 실제 스키마와 문서에서 지식을 아래에서 위로 발굴하되, 사람과 도메인 팀이 정의와 표준을 승인하는 연합형 거버넌스로 지속적으로 운영해야 한다. 또한 BI 시맨틱 계층이나 단순한 맥락 검색 계층과 달리, 지식 계층은 메트릭뿐 아니라 엔터티 관계, 데이터 계보, 시스템 매핑, 정책 및 조직 책임까지 연결하며, 기업 고유의 지식이 범용화되는 AI 모델 시대의 지속 가능한 경쟁 우위가 될 수 있다고 본다.

Neo4j’s meta-knowledge-graph approach addresses the limits of storing agent memories as isolated fragments by consolidating approved learnings into reusable task skills and user-specific profiles. Task learnings are safety- and consistency-checked, grouped through shared task patterns, and periodically converted into procedures containing use cases, steps, pitfalls, and verification methods, while preserving links to their originating sessions and error records for traceability and later revision. Skills are discovered incrementally through MCP search and fetch tools or prompt-time suggestions, reducing context use by loading full instructions only when needed, and their source fragments leave ordinary recall once consolidation is complete. Separately, approved user-scoped facts are summarized into compact, versioned profiles that capture cross-project preferences without altering the shared system prompt, allowing each user to receive personalized adaptations at session start. Hook-triggered background services, thresholds, cooldowns, optional human review, and stale-data handling govern both processes, aiming to let agents turn accumulated experience into maintainable guidance while retaining provenance and the ability to correct outdated knowledge.

Dan Newland contrasts vendor-led and customer-led approaches to intelligence analysis platforms, arguing that each determines whether long-term knowledge, operational capability, and control remain with a provider or develop within the customer organization. Vendor-led implementations can accelerate complex initial deployments by supplying technology and expertise, but may create growing dependency as expanding data sources, workflows, and use cases require vendor intervention. Customer-led models involve internal data, IT, and analysis teams from the outset, allowing them to understand and adapt the platform as intelligence needs, threats, and priorities change. The article suggests organizations assess ownership by asking whether their teams could modify the system for new use cases, understand its data models and workflows without its original builders, and migrate the accumulated intelligence capability rather than only its raw data. It presents GraphAware Hume, built for Neo4j, as an open, modular platform intended to help customers retain ownership of their data, models, pipelines, workflows, and technical knowledge while receiving implementation support.

Neo4j proposes a method for massively parallel node and relationship imports that avoids deadlocks and lock contention by partitioning data according to the number of available workers rather than node-ID suffixes. Nodes are assigned deterministic hash-based partitions, while relationships are assigned source-target partition pairs; node partitions can all load concurrently, but relationship batches must prevent simultaneous access to the same partitions. For relationships between disjoint node sets, cyclic matrix diagonals provide safe parallel batches, whereas relationships within the same node set require a round-robin approach that separates forward and reverse pairs and handles self-references safely. The article also presents Neo4j Graph Data Science’s K-1 coloring algorithm as a more general way to form conflict-free batches for complex graph structures by coloring partition-pair conflicts, although it may produce less efficient schedules than the specialized Python methods.

AWS infrastructure audits often require answering relationship-based questions, such as whether security controls cover all in-scope workloads or whether systems outside a compliance boundary can reach customer data, which flat inventories cannot reliably address. The post proposes building a typed Neo4j graph from AWS Config configuration data, supplemented by Resource Explorer for inventory validation and Cloud Control only for unsupported resource types, with explicit nodes and relationships for accounts, workloads, networks, identities, policies, controls, data stores, findings, and immutable evidence snapshots. This model enables direct queries for missing GuardDuty coverage, bounded network reachability, and segmentation failures using connected-components analysis, while distinguishing network exposure from actual IAM-authorized exploitability. Vector search can help identify potentially sensitive but unclassified resources, though human review and sources such as Amazon Macie remain necessary for auditable classification. Nightly snapshots and resource fingerprints support drift detection, while an agent memory service can retain investigation history, accepted exceptions, and reusable playbooks without duplicating infrastructure data. The post emphasizes limitations including collection latency, incomplete network modeling, the need for IAM policy evaluation, and the security sensitivity of an organization-wide collector, then recommends event-driven updates, broader network coverage, snapshot metrics, and links from findings to Terraform modules and responsible teams.

Vector RAG retrieves information by ranking embedded text chunks according to semantic similarity, making it effective for straightforward questions answerable from a single relevant passage, but it cannot inherently connect facts stored across separate records. GraphRAG instead represents entities and their relationships in a knowledge or context graph, enabling multi-hop traversal that can link items such as support tickets, services, vendors, and outage reports while providing an auditable path to the supporting evidence. GraphRAG commonly combines vector or full-text search to identify a starting point with graph queries to retrieve connected context, rather than replacing vector search entirely. The post cites a 2026 NICD benchmark in which a vector-plus-graph approach achieved substantially higher truthfulness and answered more complex questions than vector-only retrieval, while noting that graph modeling, maintenance, and domain-specific queries add implementation complexity. It recommends evaluating retrieval architectures against realistic production-style test sets using measures such as context precision, recall, answer faithfulness, latency, and token cost, adopting GraphRAG when relationship-aware reasoning and explainability produce measurable benefits.

Neo4j 2026.09 makes rescored binary vector search the default for newly created vector indexes across Aura, Enterprise Edition, and Community Edition, using 1-bit binary-quantized vectors for an expanded initial HNSW search before rescoring a limited candidate set with full-precision vectors stored on disk. The approach is intended to substantially reduce RAM requirements while preserving recall, and Neo4j reports benchmark improvements of roughly fivefold in both throughput and query latency compared with version 2026.08, with detailed methodology to follow. For 768-dimensional Float32 embeddings, the company estimates that a million-vector rescored binary index requires about 224 MB for its binary vectors and HNSW graph, compared with about 896 MB for scalar-quantized indexing, because the graph itself does not shrink proportionately with vector compression. Rescored binary is recommended for most workloads, while scalar quantization may provide up to 30% lower latency when sufficient memory is available, and users can adjust search expansion and fetch counts to prioritize recall. Existing indexes should be rebuilt after upgrading to obtain the newest improvements, and Aura users are advised to use vector-optimized configurations with ample disk capacity to retain the full-precision vectors needed for rescoring.

Microsoft Fabric and Neo4j Graph Intelligence are presented as a combined approach for improving retail AI recommendations by unifying siloed sales, customer, inventory, and promotional data in OneLake Delta tables and modeling their relationships as graphs. The approach addresses limitations of relational queries and simple co-purchase analysis, which can overemphasize universally popular products such as bananas, miss subtle purchasing affinities, and struggle with novel cart combinations or complex multi-hop queries. Neo4j’s Graph Data Science capabilities, including Node Similarity algorithms, analyze product neighborhoods and overlapping relationships to identify more meaningful complements, substitutes, and cross-sell opportunities. Fabric Data Agents can translate natural-language business questions into Neo4j Cypher queries, providing Copilots with deterministic, relationship-rich context while retaining Fabric’s governance, single sign-on, and tenant security controls. Organizations can deploy the integration through Fabric’s Workload Hub or Azure Marketplace to build graph-grounded recommendation systems without separate infrastructure.

Neo4j describes a GraphRAG architecture intended to make enterprise AI applications more accurate, auditable, and trustworthy by grounding LLM responses in a structured knowledge graph and continuously evaluating results. Built on Neo4j Aura, the approach combines schema-based graph ingestion for structured data with an unstructured-data pipeline that chunks text, extracts and resolves entities, creates embeddings, validates outputs, and stores both semantic vectors and connected facts. Its GraphRAG Python package, Needle StarterKit interface, and DeepEval-based evaluation framework support modular development, graph and vector retrieval, and ongoing quality measurement. After an initial vector-RAG chatbot lacked rich graph context and automated testing, Neo4j rebuilt it with a strict domain schema, an agent that selects either vector-plus-graph retrieval or Text2Cypher queries, and an 80-question expert-curated test suite measuring correctness, relevance, faithfulness, and tool selection. Neo4j reports that the revised system improved pass rates from 68.75% to 100%, correctness from 64.75% to 97.88%, and relevance from 86% to 97.34%, while maintaining 100% faithfulness and handling 1,500 messages in its first week without manual-review escalations.

Neo4j’s Jev experiment, neo4jev, explores using TypeSafe AI’s Jev decision model to navigate knowledge graphs by selecting among actual graph relationships rather than generating text or queries. Jev is presented as a fast, low-cost probabilistic classification system that returns choices, truth-like confidence scores, and numeric scores in parallel, making it suited to repeated software decisions that would be slow or unreliable with conventional LLMs. The prototype performs beam-search traversal of Neo4j’s public Companies graph, evaluating outgoing and incoming edges at each node, checking whether a destination or goal has been reached, and recording probability distributions for every decision. It supports navigation toward a specified node, a path-shaped intent, or a free-text objective, while limiting cycles, search depth, API calls, and the number of candidate edges. The author reports that performance depends heavily on providing clean, concise graph data, schema context, balanced candidate sampling across relationship types, and meaningful goal definitions, with vague free-text goal checks remaining a limitation. The project demonstrates a broader approach in which AI systems make transparent, grounded choices from fetched options for applications such as query planning, workflow routing, tool selection, and other probabilistic decision tasks.

Neo4j’s weekly update highlights preparations for NODES 2026, a free 24-hour virtual conference on November 12 featuring more than 100 speakers and preceded by October workshops on GraphRAG, agentic systems, reasoning, memory, and production graph-and-GenAI practices. Featured content examines a self-learning multi-agent architecture that uses a semantic layer and reviewed, contradiction-checked memory; NICD research suggesting lightweight graphs can substantially improve multi-hop RAG answers and reduce hallucinations compared with vector-only approaches; and an enterprise knowledge-layer model that treats organizational roles as core nodes for governed AI agents. The post also profiles Luis Latorre Jose’s AgentMemory for .NET, a graph-native memory implementation for Microsoft Agent Framework, and Hexology’s geospatial digital-media work documenting Ukrainian cultural heritage. Additional resources include GraphAcademy training, webinars, workshops, upcoming livestreams, conferences, community meetups, and an example of querying an icon graph database through an MCP server using deterministic and natural-language access.

Neo4j announced updates aimed at helping enterprises apply graph technology to industry-specific AI challenges and scale production deployments. Its new GraphAware Financial Crime Intelligence solution uses predefined graph-based detection rules and an entity-resolved knowledge layer to identify connected fraud patterns across accounts, identities, devices, and organizations, while preserving context and provenance for explainable investigations. To support expanding AI workloads, Neo4j Virtual Graph, currently in preview, provides zero-copy graph capabilities for data that remains in external systems, while AuraDB adds multi-database support, automated API provisioning, and larger configurations with up to 2TB of RAM and 5TB of storage. Neo4j also introduced generally available cross-cluster database replication for Enterprise Edition to improve resilience, with AuraDB support planned, and previewed Neo4j Graph Analytics for self-managed customers to run analytics independently from operational databases.

Neo4j outlines a set of capabilities intended to help organizations scale graph-powered AI applications across large datasets, multiple workloads, and production availability requirements. Virtual Graph, in public preview for Aura customers, enables zero-copy graph querying over Snowflake, Databricks, and BigQuery by translating Cypher into SQL while keeping data under its existing governance, making it suitable for warehouse-latency GraphRAG, exploration, and enrichment workloads. Multiple Databases on AuraDB Business Critical and Virtual Dedicated Cloud allows isolated tenant, environment, project, or workload databases to share an instance’s compute and storage, reducing the operational burden of managing separate instances. AuraDB also offers High Memory configurations of up to 2 TB of RAM and 5 TB of storage for large graph workloads, while Cross-Cluster Database Replication provides active-passive replication across independent clusters for disaster recovery, compliance, migrations, and geographically local reads. Finally, Neo4j Graph Analytics, offered through an early access program for self-managed deployments, separates analytics algorithms from transactional database resources through dedicated, scalable compute, helping prevent intensive graph analysis from affecting production operations.

Security operations centers face cybersecurity fatigue as analysts contend with constant alerts, false positives, and compliance demands, and although agentic AI can automate tasks, it often lacks the tacit knowledge and business context held by experienced staff. The post argues that embedding policies and definitions separately in individual AI agents creates duplicated, inconsistent knowledge, causing agents to make conflicting assessments of concepts such as critical assets, acceptable risk, or inactive accounts. It proposes a centralized Cybersecurity Knowledge Layer that combines a security ontology defining assets, controls, policies, and operational constraints; current enterprise data such as telemetry, inventories, and threat intelligence; and contextual memory containing past decisions, false positives, and remediation outcomes. Neo4j is presented as infrastructure for this layer because graph-based data modeling supports multi-hop attack-path analysis, GraphRAG combines semantic retrieval with network topology, temporal memory preserves changing states over time, and Model Context Protocol integrations can connect specialized agents to shared governed context. By separating agents’ reasoning and action capabilities from the organization’s central knowledge, the approach aims to make automation more consistent, explainable, adaptable, and capable of preserving institutional learning across changing AI models and security tools.

Graph visualization converts connected data into interactive diagrams of nodes, relationships, and properties, helping users explore dependencies, paths, clusters, and patterns that tables, charts, and dashboards may obscure. It is particularly useful for relationship-focused applications such as inspecting AI agent memory and reasoning context, identifying fraud rings, tracing supply-chain disruptions, mapping IT dependencies, validating recommendations, and understanding customer journeys. While conventional charts remain better suited to summarizing metrics and trends, graph tools support exploration of complex, multi-hop connections and can be combined with graph algorithms to reveal less obvious structures. Neo4j highlights Bloom for no-code graph exploration and editing, Dashboards for combining graph views with charts and metrics, and the TypeScript-based Visualization Library, along with Python support, for embedding customizable graph interfaces in applications.

Neo4j GraphAware Financial Crime Intelligence is presented as a graph-native platform for banking and insurance teams that aims to address financial crime investigations hindered by fragmented data, siloed tools, and isolated alerts. It creates a persistent, entity-resolved knowledge layer that combines internal systems and external intelligence, connecting customers, accounts, transactions, devices, cases, watchlists, corporate data, and other relationships. The platform supports detection of connected risk patterns such as identity reuse, shared infrastructure, circular fund flows, money mule networks, and organized insurance fraud, while enriching alerts with context for investigation. Investigators can trace relationships across multiple degrees of separation, incorporate case-specific information, use graph analytics and AI-assisted workflows, and preserve evidence, provenance, and decision rationale for review and regulatory defensibility. Designed for fraud, anti-money-laundering, KYC, sanctions screening, transaction monitoring, and claims investigations, it is intended to complement rather than replace existing detection engines, case-management tools, and data infrastructure, allowing organizations to start with defined use cases and extend models, workflows, and data sources over time.

A Neo4j experiment argues that financial-crime AI systems benefit more from enriched semantic graph layers than from larger language-model context windows alone. Using a synthetic dataset of 30,000 accounts and 5 million transactions containing hidden mule rings and synthetic identities, it compared an agent working from raw graph relationships with one supplied precomputed graph metrics such as community clusters, centrality, similarity, and path distances. The enriched approach more reliably identified clusters of accounts sharing personally identifiable information and detected structural bridge accounts connecting suspicious communities, while the raw approach was better able to trace specific multi-hop circular transaction loops that enriched aggregate properties did not expose. The results suggest that graph algorithms can convert complex transaction networks into more usable structural signals for AI investigation, but that the most effective fraud-detection workflow combines enriched graph context to prioritize risks with raw transaction traversal to investigate individual money flows.

Graphs model data as interconnected nodes and relationships, helping organizations identify patterns in areas such as fraud, recommendations, supply chains, cybersecurity, and customer analysis that may be difficult to see in traditional tables. Knowledge graphs extend this model by adding shared definitions, taxonomies, ontologies, and business rules, creating a governed view that clarifies what entities and relationships mean and supports discovery, compliance, and AI grounding. Context graphs add enterprise knowledge, conversational history, and prior decision traces to provide the most relevant information for a specific task or moment, particularly for AI agents that require accurate, explainable decisions and persistent memory. GraphRAG is presented as a retrieval method that uses graph relationships to supply connected context to generative AI rather than isolated documents, while knowledge and context graphs together can form a knowledge layer linking enterprise data, models, workflows, and actions.

Decision traces are structured, persistent records that explain how AI agents reached decisions by linking outcomes to reasoning steps, tool calls, policies, conversations, and other relevant context. Stored as reasoning memory within a context graph alongside long-term enterprise knowledge and short-term conversation history, they can support explainability, debugging, compliance audits, consistency, learning, and collaboration among multiple agents. Unlike application logs, which track events, or LLM traces, which detail individual model runs, decision traces focus on why a decision was made and can be retrieved as precedents for future similar tasks. Neo4j Agent Memory provides SDKs, framework integrations, and APIs for explicitly capturing, completing, inspecting, and searching these traces in a graph, enabling teams to review agent behavior and give agents reusable decision history.

Neo4j Labs has introduced the experimental `@neo4j-labs/nams-ai-provider`, a community provider for the Vercel AI SDK that gives AI agents persistent, cross-session memory through the hosted Neo4j Agent Memory Service. The package automatically retrieves relevant memories before a model call and stores new information afterward, using a graph-based model of entities, preferences, decisions, histories, and confidence rather than only vector-searchable chat text. It supports provider, middleware, and tool-based integrations, allowing developers to choose invisible automatic memory, visible model-controlled memory, or a hybrid approach, while hooks can enforce that memory is queried and conversation turns are persisted. Retrieval searches several memory sources in parallel, limits prompt injection to avoid overwhelming user requests, and includes fallback searches for semantically mismatched wording. Optional extraction converts stored content into typed entities that can be queried, corrected, merged, and traced across conversations, though hosted API support for graph relationship edges is not yet available. Designed to run through HTTPS without requiring a Neo4j deployment or vector database setup, the package aims to fail safely when memory services are unavailable, but it remains actively maintained experimental software with no service-level or backward-compatibility guarantees.

A Neo4j-based Neocarta semantic layer is presented as a metadata-only graph that helps Text-to-SQL agents discover relevant tables, columns, and join paths across large, messy data catalogs without repeatedly scanning warehouse schemas. Unlike metric-focused semantic layers that define certified measures over curated data, this approach emphasizes cross-platform schema discovery by ingesting warehouse metadata, descriptions, foreign-key relationships, sample values, and embeddings into a graph while leaving actual data in the warehouse. In benchmarks on a 278-table BigQuery Census catalog and a 264-table Databricks legacy lakehouse, graph-based retrieval reportedly improved reliability for smaller models and reduced token use, latency, and warehouse catalog queries for larger models; compact retrieval on Census reduced tokens by 45–81%, while Databricks tests showed 19–54% token reductions across eight Anthropic models. The approach uses semantic search to locate an anchor table and graph traversal to supply related tables and join details, allowing agents to generate SQL after limited retrieval rather than brute-force exploration. The post argues that such a graph can complement rather than replace existing governed metric layers, potentially expanding into an enterprise knowledge layer containing glossary terms, ownership, governance metadata, usage-derived relationships, and agent memory.

Neo4j describes a no-code, multi-agent integration in which Salesforce Agentforce uses the Model Context Protocol (MCP) to send natural-language business questions to Neo4j Aura Agent, which handles graph-specific reasoning and returns grounded recommendations. Using a Northwind retail example, the Aura Agent can identify a customer, analyze purchase history, determine an unavailable frequently ordered product, find suitable in-stock alternatives in the same category, rank them, and provide supporting evidence through deterministic Cypher Templates. Agentforce remains responsible for employee interaction, intent routing, Salesforce permissions, and potential CRM workflows, while Aura Agent owns graph traversal, domain rules, and retrieval logic. Unlike a prompt template, Aura Agent includes planning, tool selection, graph retrieval, and publishing through secured REST or MCP endpoints. The architecture uses machine-to-machine authentication and read-only Aura operations, while keeping graph schemas, ranking rules, privacy constraints, and Cypher logic on the Neo4j side so Salesforce can continue submitting stable business-level requests.

Using a small Neo4j test graph of three learners with different levels of course activity, including one user with no relationships, the post demonstrates how mandatory Cypher relationship patterns can unintentionally exclude unconnected nodes from results. It explains that `MERGE` patterns should contain only stable identity properties, such as email addresses, while mutable details belong in `SET` to avoid duplicate nodes. A query matching users through enrolments and courses returns only users who satisfy every relationship requirement, causing new or inactive users to disappear despite existing in the database. To preserve all users, the recommended approach is to anchor the query on the primary nodes and retrieve optional related information with list comprehensions, `COUNT {}` patterns, or `COLLECT {}` subqueries, which can yield empty lists or zero counts without removing rows. Although `OPTIONAL MATCH` can also include missing relationships, chaining several optional matches can multiply rows and create unnecessary work, whereas separate list expressions avoid duplication. The broader lesson is that dashboards and reports can overlook precisely the unconnected customers, employees, or learners who may need attention, so queries should be reviewed for implicit requirements before concluding that data is absent.

Neo4j’s Document Intelligence can be adapted from its intended GraphRAG and document-analysis use cases to create an AI-supported homework quiz system from text-based PDF study materials. Using a free Aura instance, users define a simple graph model containing Question and Answer nodes, import a document to generate question-and-answer pairs linked to source chunks, and remove pairs that do not contain exactly one answer. A free Aura Agents approach can then randomly retrieve questions, compare student responses with graph-provided textbook answers and context, and assign correctness and quality scores, although it cannot retain progress or prevent repeated questions. For more control, a small Java application and Cypher queries can track already asked questions, use an external LLM API such as OpenAI to evaluate answers on a 0–3 scale, provide explanations and correct answers, and calculate an overall result. The demonstration uses material on the 1657–1658 Swedish-Danish war and illustrates both the potential of structured graphs for tutoring and practical limitations such as single-user state tracking, API costs, and the need to respect copyright when uploading documents.