From Documents to a Knowledge Graph: Organizing an AI Expert Team with WorkBuddy

By Quwei Ye (Tree / Shuye) October 11, 2026

Our folders keep filling up with documents, yet when we need to answer a specific question, we often have to search through them all over again.

While preparing my Introduction to Artificial Intelligence course, I already had the course plan, lesson plans, slides, and project materials. But which activities support a particular learning objective? What knowledge should students review to complete a hotel management project? Do different documents describe the same requirement consistently? The answers are often scattered across separate files.

I wanted to try organizing these materials into a knowledge graph, then have a “Knowledge Graph Expert Team” in WorkBuddy use tools to investigate questions, trace relationships, check evidence, and, where appropriate, perform computation and reasoning.

This article records a practice being built step by step. According to the records as of October 2, the expert team had been installed and was available for selection, and the task templates and rules for handling course materials had been established. Full extraction, database ingestion, and evaluation across the entire course remained unfinished. I want to explain this approach clearly while identifying what still needs to be tested.

01 Begin with a Question Worth Pursuing

A knowledge graph organizes knowledge through entities and relationships. People, organizations, events, and concepts can become nodes. In a course, class sessions, learning objectives, knowledge points, assignments, and assessment criteria can also be nodes. Relationships between nodes describe how they are connected.

For example, helping students prepare for a hotel review analysis project requires us to know: What knowledge does the project involve? In which class sessions is that knowledge introduced? What exercises are available? And what criteria are used to assess the results?

Clarify the questions you want to answer before deciding what belongs in the graph. A graph is useful to the extent that it helps us find these connections and return to the source materials to verify them.

The course materials for this practice comprised one course plan, 16 lesson plans, 16 slide decks, and four project documents. I established an important principle: use the course plan and lesson plans as the primary sources, with slides serving only as supplementary references.

The course plan sets out the overall arrangement, lesson plans describe how each class session unfolds, and slides help explain terminology, diagrams, and cases. The appearance of a knowledge point in a slide deck does not by itself make it a formal learning requirement or assessment requirement. When primary sources conflict, the issue should be listed for further checking.

02 How Skills Become an Expert Team

The WorkBuddy Open Platform offers capabilities including skills, experts, expert teams, connectors, and Buddy applications. These names can seem confusing at first, but their respective roles become clearer within a concrete task. [1–5]

A skill preserves a method for getting work done. How to design nodes and relationships, extract information from documents, or check query statements can all be captured in reusable instructions, scripts, and templates. Organizing skills means accumulating working methods; it does not mean retraining a large language model. [1]

The seven Neo4j skills we assembled cover graph modeling, document extraction, structured data import, Cypher queries, Python access, GDS graph computation, and MCP tool integration. They provide guidance for executing the different stages.

Experts and expert teams organize roles and collaboration. This project configured five roles: Gu Zhixing coordinates delivery; Lin Zhiwei handles modeling and extraction; Cheng Guiyi integrates data and loads it into the database; Shen Mingli handles retrieval, computation, and reasoning; and Yan Suyuan reviews the evidence. These names express the division of responsibilities within this project. [2–3]

Handoffs are central to collaboration. Extracted results must include their locations in the original sources; entity merges must retain their justification; analyses must include the queries or algorithm parameters used; and issues found during review must be returned for revision. Simple tasks may need fewer roles. Having more roles does not automatically make an answer correct.

A connector provides the actual connection to tools and data. It enables experts to call a database or an external service. After installing a skill that describes how to connect, we still need to configure the connection and permissions and check that the tools work properly. [4]

A project, meanwhile, retains ongoing context, such as the scope of the materials, source priorities, and delivery requirements. Each new task only needs to specify what is to be done on that occasion. In this way, skills, roles, tools, and project context gradually form a reusable workflow.

03 Graph Relationships Must Lead Back to Their Sources

Giving documents to AI does not make every extracted relationship reliable. Building a graph involves taking an inventory of the materials, designing the schema, extracting, validating, merging, and loading the data. Versions and update records must then be maintained.

“A lesson plan mentions clustering” and “clustering is an explicit prerequisite for a project” are two different claims. The appearance of two terms on the same page does not directly establish a prerequisite relationship between them.

Evidence should therefore accompany each specific claim: which document, paragraph, or slide it comes from; what the original wording says; and which version it applies to. Relationships that cannot yet be confirmed can remain items awaiting verification.

An early check of the materials uncovered just such a situation: the course’s parent directory was labeled “26–27,” while the text of the materials stated “Semester 1, 2025–2026.” The system should retain this discrepancy rather than overwrite the text with the folder name. In another example, a lesson plan stated that the corresponding slide deck had not been provided, but a new archive already contained a file for that class session. Its contents still needed to be checked.

Duplicate entities must also be merged carefully. The Chinese and English names and abbreviations for the same concept can be consolidated after checking, while retaining their aliases. Entities with different meanings but similar names should be kept separate. Repeated imports should not keep adding more copies of the same concept to the graph.

This course also uses a separate dataset, ai_intro_submitted_v1, kept apart from demonstration materials and other graphs. That boundary must apply to imports, queries, updates, exports, and graph computation. When several people use the system, appropriate access controls are needed in addition to the dataset identifier.

04 Retrieval, Computation, and Reasoning Address Different Questions

Once a graph is built, the first question is what it can answer. Retrieval, graph computation, and rule-based reasoning have different uses and should be validated separately.

Knowledge retrieval follows relationships to find answers with identifiable sources. For example, starting from a course project, it can find related knowledge points, class sessions, activities, and assessment requirements while returning their original source locations. In the future, full-text or vector retrieval could also be combined with the graph to form a graph-based retrieval-augmented generation (GraphRAG) workflow. This part still requires development and validation.

Graph computation examines the structure of a knowledge network. Neo4j’s Graph Data Science (GDS) tools provide algorithms for centrality, community detection, similarity, and pathfinding, among others. Before selecting an algorithm, we should clarify which nodes and relationships we want to study, then determine direction, weights, and parameters. [6]

A knowledge point connected to many things may deserve attention, but it is not necessarily the most difficult or most important. Communities identified by an algorithm do not necessarily correspond to teaching units, and the shortest path may not be the best learning sequence. For a small graph, an ordinary query may already be sufficient.

Rule-based reasoning checks or derives conclusions from explicit premises. Suppose we use a checking rule that “learning objectives should have evidence of activities and assessment.” If an objective is linked only to activities, the system can flag: “No assessment evidence was retrieved from the current graph; verification is required.” It cannot directly conclude that the teacher has made no assessment arrangements.

Derived results need to state the rules, premises, and scope used. The current approach mainly uses explicit rules and queries; this does not mean a complete formal ontology reasoning system has been deployed.

Throughout this process, an agent interprets the task, chooses tools, observes the results, and organizes the deliverables. Neo4j stores knowledge relationships, tools execute queries and computations, and the language model helps explain the results. Only when every stage leaves a record can we continue asking how an answer was formed.

05 How a Personal Account Can Proceed When Native App Creation Is Restricted

I initially hoped to turn this workflow into a native Buddy application called “Knowledge Graph Workbench,” where users could open it, choose a mode, enter their materials, and ask questions. Actual use revealed a condition that required an adjustment.

During my trial on October 2, the Open Platform informed me that my personal developer account needed enterprise verification to create a Buddy application. This message concerned eligibility to create an application; it does not establish that all Buddy applications can be used only by enterprise users. The platform’s instructions at the time of use remain the basis for determining eligibility.

For personal experimentation, we can first select an installed expert team within a WorkBuddy project, call the skills and local Neo4j tools, and work through a specific task. The existing file “使用入口.html” helps users enter a mode, material paths, a dataset, and a question, then generates task text to copy into the conversation. At present, it serves as a task preparation aid.

Once the relevant developer qualifications are in place, native application creation, configuration, preview, review, and publication can be considered. The official documentation describes these stages, but “the expert team is available for selection,” “application configuration is complete,” and “the entire workflow has been validated” still need to be confirmed separately. [5]

We have prepared four workbench modes: building and updating a graph, evidence-based question answering, graph computation, and rule-based reasoning. We have also designed six scenario entry points: building a graph from materials, adding materials, tracing relationship chains, examining important nodes and communities, making rule-based deductions, and checking quality. Each needs to be tested through real tasks.

06 Start with One Source and One Question

If you would like to try this approach, begin with a clearly bounded set of materials, such as one teaching unit within a course. Create a project in WorkBuddy, select the “Knowledge Graph Expert Team,” and confirm that the current computer can access the chosen materials and database tools.

For the first task, specify four items clearly: choose “Build and Update a Graph” as the mode; provide the actual folder path on your own computer; use a stable, separate dataset identifier; and phrase the question as a specific request whose answer can be checked.

For a course, the task could be phrased like this:

Using the course plan and lesson plans as the primary sources and the slides as references, organize the relationships among learning objectives, knowledge points, activities, assignments, and assessment requirements that are supported by the original text. Preserve source references, merge clearly duplicated entities, and list conflicts and cases with insufficient evidence. After validation, load the data into the database, answer the specified questions, and deliver a graph page where relationships and evidence can be viewed.

In the first round, check a few concrete things: Can key relationships be traced accurately to their original sources? Do repeated imports create unnecessary entities? Do queries inadvertently include other materials? Can the system acknowledge that an answer is unknown when the materials do not provide it? These questions reveal more about reliability than the number of nodes does.

When sharing results, a self-contained HTML file can provide a graph snapshot for others to browse; live queries and updates require an accessible backend. If the page also needs to open a full report, package the report with it, use relative paths, and ask recipients to extract the archive first.

This also aligns with the six-stage framework I continue to use: record the materials and their sources; reflect on relationships; ask questions that can lead to action; carry out a small-scale practice; review the results; and transfer reliable methods to the next set of materials. Each step leaves a basis for the next judgment.

I hope this way of working will eventually make lesson preparation, research, and knowledge organization easier. At present, however, there is not enough evidence to claim that it has improved teaching outcomes. The next step is to complete a full task, compare the previous approach with the new workflow, examine what it saves and what additional verification work it creates, and then decide how to extend its use.

If, the next time we open our materials, we can find the relationships we need more quickly and explain where those relationships came from, this practice will have earned further development.

About the Author: Quwei Ye

Quwei Ye (Tree / Shuye) is a professor whose work focuses on the ability to ask questions and on learning innovation in the AI era. He has spent more than forty years in educational practice, with extensive experience in higher education teaching, management, and educational evaluation. In recent years, his interests have turned to generative AI and AI agents. His six-stage framework—Record, Reflect, Question, Act, Review, and Transform—connects learning, action, and transfer.

Personal website: www.treeye.cn

WeChat: tree_ye1961

Email: shuye_1961@qq.com

Meet at JinShang

Founded in 2016 by interior designer Lilian Tian, JinShang is a community-oriented space brand that brings together work, events, dining, art, and lifestyle.

JinShang Chengxi is located at No. 34 Xinjiekou Outer Street, Xicheng District, Beijing. Chengxi Courtyard is suited to small meetings and gatherings over tea; the Xiyan Reading Space brings books and conversation together; and Chengxi Theatre hosts lectures, drama, and public cultural events. Together, these offer different spaces for reading, exchange, reflection, and creation.

Discussions about AI also need a space for thoughtful exchange. I hope people will bring their own materials and questions to discuss what they discover in actual use: where time is saved, where human judgment is still needed, and which lessons can inform the next action.

Notes on Sources and Practice

The status of this practice follows the original records as of October 2. All seven technical diagrams and platform screenshots have been retained unchanged. The course structures and agent workflows shown in them do not imply that the entire process has been validated.

[1] WorkBuddy Official Documentation: Skills

[2] WorkBuddy Official Documentation: Experts

[3] WorkBuddy Official Documentation: Expert Teams

[4] WorkBuddy Official Documentation: Connectors

[5] WorkBuddy Official Documentation: Buddy Applications

[6] Neo4j Official Documentation: GDS Graph Algorithms

Read the Chinese edition