At the beginning of this year, the US-Iran conflict drew widespread attention to Palantir's military applications. According to public reports, the US military uses its Maven system to fuse multi-source intelligence, assisting in target identification and action planning. How to transform massive amounts of data into executable decisions has become a key perspective for observing technological applications and their evolution in modern warfare, also driving discussions about Palantir's technical ecosystem.
"Ontology" is a keyword in this discussion.
In Palantir's technical architecture, ontology connects data, business objects, and actions, providing unified support for analytical judgment and business execution. As domestic interest in this approach grows, how ontologies support complex decision-making and integrate into actual business operations has become a concern for an increasing number of vendors and clients.
Utenet, which has long served defense clients, directly feels this change. Hong Wanfu, founder and chairman of Utenet, noted that since the beginning of this year, core clients have significantly increased their focus on dynamic ontologies, starting to raise requirements around specific applications and long-term construction. From early knowledge graphs to current industry models and decision platforms, Utenet is also trying to use new technical capabilities to handle continuously changing data, business relationships, and decision tasks. However, from battlefield applications gaining attention to implementation in Chinese clients' businesses, there are still a series of engineering issues. How much ontology construction work can large models undertake? How do specialized applications expand to larger business scopes? And how much of Palantir's delivery model can be referenced?
Recently, iFenxi conducted an in-depth exchange with Hong Wanfu regarding these questions.
Core Viewpoints
- Dynamic ontologies respond to the needs of multi-source heterogeneous data fusion and continuous business changes.
The changes brought by large models mainly lie in providing stronger automation capabilities for conceptual modeling, semantic alignment, and reasoning; ontologies and knowledge graphs still need to be used in combination.
- Moving from specialized applications to industry-level ontologies is a long-term project.
A business closed loop can be formed within a small scope, but as the scope expands, data standards, cross-system interfaces, permissions, and audits all need to be reorganized; coordination beyond technology is equally important.
- The difference between industry-specific agents and general-purpose agents is mainly reflected in working methods and delivery results.
Private deployment and domestic adaptation are basic requirements; more importantly, it relies on understanding industry knowledge through ontologies, following business norms, and delivering verifiable results.
- It is difficult for China to completely replicate Palantir's delivery model.
Pricing methods for software and expert services, data access conditions, and project organization mechanisms all affect whether field experience can effectively feed back into products.
The following is the transcript of this interview, slightly modified without changing the original meaning.
01、Dynamic ontologies must keep up with changes in business and data
iFenxi : How do you understand dynamic ontologies? What specific background leads clients to propose such requirements?
Hong Wanfu: From my understanding, the emphasis on the term "dynamic ontology" in China today is largely to distinguish it from relatively static ontologies of the past. Our company started with knowledge graphs early on. The common practice was to first build concept and field structures, i.e., Schema, and then process data according to this structure. Classes, attributes, and relationships were primarily defined manually beforehand. Once released, modification costs were high, especially when involving inference consistency. In recent years, this approach has struggled to meet some client needs. For example, in intelligence analysis within the defense sector, information sources are diverse; it is impossible to build a separate Schema for every source. Different sources describe the same object differently, requiring unified cognition. Furthermore, business processes themselves are not necessarily fixed; when involving multi-party collaboration, states and data update constantly, and these changes need timely feedback. Therefore, clients began raising requirements for dynamism from perspectives including modeling, semantics, data updates, events, actions, and functions. Essentially, clients hope for a relatively universal platform capable of handling heterogeneous data and adapting to continuous business changes. This "relatively universal" does not mean one system solves all problems. There is hype behind the concept, but there are indeed real needs behind it.
iFenxi : How are these changes specifically reflected in various stages?
Hong Wanfu: Taking our practice in intelligence analysis as an example. First is the change in modeling methods. Previously, we had to pre-define what information to focus on and what fields existed based on analysis goals. Now, we use accumulated cases, doctrines, and industry models to assist the system in generating preliminary ontology drafts and candidate fields; which types and relationships can enter the database still require constraint validation and expert confirmation, especially parts involving target attributes and action boundaries. When integrating new information collection methods, the system can derive which fields need supplementary collection, for analysts to confirm before inclusion. Second is semantic alignment. The same object may use codes, abbreviations, or other names in different sources. Previous knowledge graphs also performed semantic alignment, but limited by natural language processing capabilities at the time, they often required manual organization of synonyms and near-synonyms. Now, large models have stronger extraction and summarization capabilities, supporting the construction of a unified semantic layer to map different descriptions to the same object, action, or algorithm. Third is subsequent reasoning. Once information fusion and the semantic layer are well-established, they can support more complex analysis and further reasoning combined with user interaction. Ultimately, clients not only want to see a result but also know how it was derived—being able to trace back to corresponding evidence and reasoning chains. Overall, compared to the past, both automation and intelligence levels have improved.
iFenxi : Will dynamic ontologies replace knowledge graphs?
Hong Wanfu: I believe this is not a replacement relationship, but a matter of layers. In our practice, data storage and querying still primarily rely on graph databases, but what goes into the graph, how objects connect, which fields can be modified, and which conclusions must link to evidence are constrained by the ontology layer. So the more accurate statement is: Ontology defines semantics and rules, while graphs handle part of the storage and computation; the two work together. We proposed the concept of "dynamic graphs" very early on, focusing mainly on whether instance data changes could be fed back to the graph in time. Today's "dynamic ontology" pushes the change up one level: not only instance data changes, but type definitions, relationship constraints, and action boundaries adjust accordingly. Once the ontology changes, it may affect existing instances, established conclusions, and reasoning results. Therefore, governance mechanisms need to be established, including at least three things: change grading (what can be automatically published vs. what requires review); binding each reasoning result to the ontology version at that time; and impact analysis and regression verification on downstream conclusions. Without these, so-called "dynamic" just shifts manual maintenance costs online. In the past, the ontology layer relied more on manual definition or relatively fixed rules for maintenance. The changes brought by large models now are mainly reflected in the automation of three links: conceptual modeling, semantic alignment, and element updating. Client demands did not emerge today. Ten years ago, people asked me: "Can I avoid modeling? Can I extract automatically?" But at that time, NLP technology wasn't strong enough; even in domains with good training data, extraction accuracy was limited, so we dared not use automatic extraction extensively and could only adopt semi-automatic methods. Now that underlying model capabilities have enhanced, we have the conditions to further do these things.
iFenxi : Currently, how much ontology construction work can large models undertake?
Hong Wanfu: If estimating based on experience, given an industry large model that understands industry business, large models can roughly undertake 60-70% of the workload for initial ontology modeling drafts. Using only base models might reach less than half. Moreover, workload proportion does not represent quality level. Ontology is the foundation for subsequent reasoning and applications; if key relationships deviate, it may lead to consecutive errors in subsequent conclusions, and it is hard to trace where the problem lies. So what really matters is not just "how much work the model can complete," but whether errors can be detected in time. Taking the defense sector as an example again, many defense industry data and contexts are unavailable to base models. They don't understand command business or how different doctrines connect, so they can only perform tasks like text classification and information extraction. Once involving business logic, effectiveness diminishes. Even using industry models for automatic construction, extensive auditing is still required. After generating a modeling framework for targets, experts must confirm whether the overall direction is correct and if specific content meets requirements. Especially in the final decision-making phase, what we currently do is assisted decision-making; critical nodes generally involve human intervention.
02、Moving from specialized applications to industry-level ontologies is a long-term project
iFenxi : When ontologies enter real business scenarios, do clients worry that automatic updates will affect system stability?
Hong Wanfu: Definitely. Especially when states need to be written back to original business systems, complexity rises significantly. From my understanding, one advantage for Palantir is that when cooperating with clients, they can often integrate data governance, the ontology layer, and subsequent business system construction, resulting in relatively complete boundaries. In contrast, domestic vendors and systems are scattered, making similar holistic advancement much harder. Sometimes, whether state updates can be written back to the original system is not just a technical issue but involves cross-departmental coordination and organizational mechanisms, which may not be easily bridged.
iFenxi : Why are many current ontology projects still stuck in pilot or verification stages?
Hong Wanfu: When advancing ontology projects, we try to adopt a specialized approach, solving problems within clear boundaries to form a closed loop. As the construction scope expands, more things need coordination, and risks become harder to control. Basic work like data governance, standard unification, and permission division alone could last one to two years. Secondly, dynamic ontologies only gained widespread attention in the past two years; engineering still takes time. Many projects don't lack ontologies, but rather, upstream data quality, data lineage, permission control, and state write-back mechanisms aren't ready. Without these foundations, dynamic updates struggle to truly enter production systems, and forming large-scale cases takes time. Third, clients need to continuously see value. For example, after using dynamic ontologies, which judgments became faster, which associations were previously undiscoverable, can reasoning processes be traced, and why are results credible—these need to be clearly presented. Otherwise, clients only see upper-layer applications and assume these capabilities "should naturally exist," failing to understand the value of underlying semantics. Solving these problems takes time; successful pilots do not equal successful scaling. Small-scale verification can rely on concentrated pushing by project teams, but industry-level construction is a massive project involving more systems, departments, security, and audit requirements. So, we cannot judge that industry-level construction will be completed quickly just because small specialized applications are increasing.
iFenxi : Besides long construction cycles, what are other engineering difficulties for large-scale implementation?
Hong Wanfu: The biggest issue remains data governance. There are many details here, such as data standards. The defense sector doesn't lack standards; rather, there are many, with different lines focusing on different priorities, requiring extensive sorting and mapping during actual implementation. Additionally, business objectives may dynamically adjust with organizational and task changes, and models need to keep up with this change. Second is cross-system fusion. After building the ontology layer, business systems still need to be connected. Incorporating all systems into one big system may affect original business operation mechanisms, which is unrealistic. Usually, connections are made via interfaces, involving the aforementioned data write-back, state synchronization, permission control, etc. Third, architectural planning capability requirements are high. One must deeply understand the business, clearly define elements like permissions, boundaries, and audits, and consider how related information is expressed and stored in the semantic layer when an object changes. Sorting out these relationships requires strong top-level architectural capability; it's not done just by plugging in a model.
iFenxi : If ontologies start with single scenarios, how do you connect ontology capabilities across different scenarios later?
Hong Wanfu: It depends on the nature and scale of the project. If simply fusing two previously isolated scenarios, we usually set up a separate project, connecting involved ontologies and related capabilities in an application-building manner. The advantage of this method is speed, but as scenarios increase, it easily forms new silos. If it's a top-down large project plan from the start, we must think ahead about what role the platform plays and which phase it is built in. We have done such projects: Phase 1 cuts into very specific scenarios to solve users' actual problems; Phase 2 delivers the platform. But platform design and internal application of related capabilities actually began in Phase 1. In Phase 1 design, we also consider future reuse and expansion in advance, such as which objects can be shared, which semantics can be precipitated, and which capabilities can be abstracted into general modules. This way, when building the platform later, we can inherit existing scenario capabilities instead of starting over. Whether this can be done relates to project volume and overall positioning. But the core principle is the same: Single-point entry can be fast, but design must reserve space for subsequent connection and expansion.
03、Industry-specific agents must understand business and deliver reliable results
iFenxi : Since the beginning of this year, how have client requirements for agents changed? What is the relationship between these changes and ontologies?
Hong Wanfu: From our contact situation, client requirements for agents have seen two obvious upgrades this year. The first upgrade is from "answering questions" to "completing tasks." At the beginning of the year, clients focused on whether agents could assist in completing some work segments, such as writing argumentation reports or preparing topic materials, with the core demand being efficiency improvement. By the second half, clients were no longer satisfied with agents "helping a little bit," but hoped they could self-iterate, autonomously select tools, or even generate necessary tools, achieving closed-loop execution of entire tasks. The second upgrade is from "usable" to "trustworthy." Clients' requirements for agents are not just "smart," but more "reliable"; output materials must have basis, operations must have boundaries, and processes must be traceable. These two changes make the role of ontologies more prominent. Because for agents to autonomously complete tasks, they must understand what objects exist in the business, their relationships, what they can and cannot do—this is exactly what ontologies need to define. Without ontologies, agents rely more heavily on large models to "guess." And clients' requirement for "trustworthiness" essentially means agents operate within determined business rules and knowledge boundaries. Ontology is this boundary; it tells agents what the business is, what the rules are, and which actions are legal, making autonomous execution more controllable.
iFenxi : Utenet has many products. How is the overall architecture divided? What role does ontology play within it?
Hong Wanfu: It can be viewed in three layers: underlying capabilities, general platforms, and industry applications. The bottom layer consists of industry large models, specialized algorithms and solvers (computing engines), and agent frameworks. Base large models require massive computing power and data investment; leading domestic vendors are already focusing on this. We focus more on industry large models, precipitating industry data and business experience into model capabilities. Additionally, specialized algorithms and solvers in areas like operational optimization, task planning, and intelligence analysis rely on models and operators accumulated from years of project practice. The focus of the agent framework is enabling agents to understand industry knowledge, call tools, and complete complex tasks, while meeting client requirements for offline deployment, private model integration, and domestic adaptation. The middle layer is our long-accumulated data middle platform, cognitive middle platform, and decision middle platform. The overall direction hasn't changed, just adding new capabilities in the era of large models and adjusting usage methods. The top layer is applications for specific industries. In vertical fields like defense, government/enterprise, finance, and industry, we have precipitated domain ontology assets and engineering methodologies, which support the R&D of specific application products. Currently, defense-related applications dominate, such as scientific research assistants and programming assistants, and we are also exploring industrial applications. Our approach is to make solutions standard products as much as possible, but actual projects still have significant customization needs. Based on experience, about 60-70% of capabilities in a current project can reuse existing modules; the rest still require adaptive development based on site conditions. Specific proportions vary by industry and project. Ontology is not an independent module in this architecture but a unified semantic and rule foundation running through the full stack. Utenet Sati Cognitive Intelligence Platform centers on dynamic ontologies, supporting ontology construction, validation, and continuous updates, helping models, agents, and business personnel form consistent understanding of the same business objects, reducing model hallucinations, and allowing existing knowledge to be reused in different tasks. Storage and computation of related knowledge are supported by our self-developed Seraph distributed graph database, meeting domestic adaptation and large-scale data processing needs. During application, KGAG (Knowledge Graph Augmented Generation Engine) uses structures and relationships in the ontology to provide associated retrieval for large models and performs ontology validation on generated results. The two solve different problems: retrieval solves "is there a basis," and validation solves "does it comply with rules." The CMAS Multi-Agent Collaborative Decision Platform applies these business semantics and rules to task understanding and collaborative decision-making, enabling multiple agents to collaborate on a common business cognitive foundation. For tasks requiring solution generation, the decision middle platform further provides capabilities like game confrontation and operational optimization, working with ontologies and knowledge graphs to form a closed loop from "understanding business" to "generating solutions."
iFenxi : You just mentioned the scientific research assistant. Compared to general-purpose agents, what are the main differences? How does ontology integrate into specific research tasks?
Hong Wanfu: Taking defense equipment as an example, stages before finalization—including argumentation, preliminary research, and testing—involve a lot of research work. For instance, whether to develop a certain equipment, how to determine parameters, and what combat concepts and targets it faces, all need thorough study in the early stages. Daily workflows for such clients usually start with reviewing and organizing massive materials, then determining research directions, forming topics, collecting evidence, and drafting project outlines. The scientific research assistant mainly serves these tasks, with ontology running through the entire work process. In the information collection stage, general tools may not obtain relevant professional materials, and even if obtained, further organization is needed. The scientific research assistant relies on domain ontologies to hierarchically model literature, test materials, and argumentation evidence, labeling semantics so materials can be retrieved, validated, and traced back to originals. When different documents refer to the same equipment or indicator inconsistently, they are aligned via unified identifiers. Entering the research stage, tasks like parameter argumentation and technical warning have specific research paradigms and cannot rely solely on material retrieval and text generation. We precipitate these methods and domain constraints into the ontology, helping the scientific research assistant understand task intent and clarify callable analytical capabilities. For example, when drafting outlines or comparing solutions, "action contracts" specify prerequisites and evidence requirements for related operations, reducing unfounded inferences. Then there is data validation; multi-source information often conflicts. General agents usually rely on models to reconcile, easily leading to distortion. We manage conflicts through ontology, combining information accumulated by the company's think tank to distinguish viewpoints from different sources and periods, preserving historical records, and marking currently accepted conclusions. After result generation, the scientific research assistant also reviews the evidence and logic of conclusions, ensuring research results are verifiable. Domestic adaptation and intranet deployment are basic requirements. On this basis, we hope the scientific research assistant can not only retrieve materials and answer questions but also conduct research and verify results under clear business constraints. This professional task closed-loop capability supported by ontology is a significant distinction from ordinary RAG (Retrieval-Augmented Generation) applications and general-purpose agents.
iFenxi : We noticed Utenet launched the UCoder programming assistant this year. Facing defense and military scenarios, how does it differ from general programming assistants? What advantages does it have in industry ontology accumulation and private asset reuse?
Hong Wanfu: The difference between UCoder and general programming assistants on the market is quite obvious. UCoder is not simply making an "assistant that writes code," but combines our accumulated defense domain ontology knowledge with clients' private knowledge bases and R&D assets, adapting to these high-security, strongly constrained scenarios. First is understanding business. General programming assistants understand code syntax and general development needs, but understanding military projects requires corresponding business knowledge. To this end, we sort out relationships between equipment, algorithm components, interface specifications, and R&D constraints through the defense R&D domain ontology, combined with clients' private knowledge bases, helping UCoder understand the business meaning behind requirements. Second is supporting the complete R&D process. For military projects, generating code is just one segment; requirements, design, coding, and testing need to form a closed loop, with clear requirements for delivery, verification, and technology maturity. To implement these requirements in the R&D process, we split software engineering specifications, national military standards, and domestic adaptation requirements into two mechanisms: For machine-judgeable content like document structure, interface signatures, coding standards, and test case coverage, we create rules UCoder can directly validate; For content requiring human judgment like sufficiency of reviews and completeness of process records, we form checklists and evidence lists for relevant personnel to confirm. This way, UCoder has clear basis when checking designs, code, and test cases, but does not replace human compliance judgments. Third is tool integration and asset reuse. Clients' existing simulation, deduction, and testing tools, along with private components and historical code, need clear functions and calling requirements to be used in new R&D tasks. UCoder uniformly describes these assets' capabilities, inputs/outputs, and parameter requirements through ontology, performing semantic alignment and encapsulation, so R&D requirements can correspond to appropriate tools and components. Combined with permission control mechanisms, it clarifies asset access and calling boundaries, meeting management requirements for high-security scenarios.
04、Palantir's delivery model is difficult to directly replicate
iFenxi : Returning to the question we discussed at the beginning: Everyone in China is discussing Palantir, but can its implementation model be fully replicated in China? Which parts are worth referencing, and which parts must likely be localized?
Hong Wanfu: Fully replicating Palantir's model is not easy. This is not just a technical architecture issue but involves a chain of commercial and compliance mechanisms: pricing for software and expert services, data access permissions for on-site engineers, and whether projects are organized by initiation/acceptance or continuous subscription. For example, "Forward Deployed Engineers" (FDE). As I understand it, Palantir typically configures two roles on-site to cooperate: one leaning towards business consulting experts, responsible for sorting out client business needs; the other engineers, stationed directly to write code, build data pipelines, and modify ontology configurations. The advantage of this model is that experience and development achievements accumulated in projects eventually return to the product, serving not just individual projects. But the cost of long-term expert stationing is high. This model requires the market to recognize the value of software and expert services, and cooperation modes to support long-term investment—short-term prototypes can start with small amounts, and after validating effectiveness, directly convert to production subscriptions, with on-site services as upfront investment ultimately recovered by software licenses, rather than priced by implementation man-days. This differs from our budget caliber based on initiation, acceptance, and settlement milestones. We also send teams on-site, but the organization mode leans more towards project implementation. If the value of business experts is mainly understood as implementation manpower, long-term stationing is easily restricted by budgets and mechanisms. So, value recognition is a fundamental issue. The scope covered by projects and how they are organized are also affected by these conditions. We cannot look at just a technical architecture and assume the whole model can be transplanted.
iFenxi : Given this situation, how does Utenet involve business experts in delivery?
Hong Wanfu: Delivery in China inherently involves solutions, project managers, implementation teams, and requirement analysts cooperating at different stages to help clients sort out needs and build systems well. Differences between companies lie more in at which node business experts intervene and how deep they go. Key nodes like detailed requirement reviews, contract reviews, and solution confirmations require judgments combining business and technology. Utenet emphasizes product leadership; although business experts do not need to be stationed long-term, they must participate in these nodes to ensure the implementation process conforms to prior planning and delivery results align with the product roadmap.
iFenxi : How do business experts feed field experience back into products?
Hong Wanfu: The biggest benefit of experts being on-site long-term is accessing the most authentic frontline needs and directly precipitating them into products during the process. In the defense sector, due to actual business condition restrictions, we cannot directly adopt the long-term expert stationing model; we rely more on internal mechanisms to precipitate and transmit needs—relevant teams sort out need backgrounds and industry experience, then synchronize them to product teams through internal exchanges. Compared to direct feedback from on-site experts, this process requires communication and organizational costs, and continuous exchanges to confirm, minimizing understanding deviations. But in other government/enterprise projects, we are more flexible. On-site implementation teams and solution managers can directly feedback client issues to the product side, and product managers regularly go to the frontline for face-to-face exchanges with clients. This "going out" approach allows us to capture common needs faster and judge which are individual cases and which deserve becoming general capabilities. So essentially, whether through stationing or internal communication mechanisms, the key is letting product teams stay close to real scenarios, forming a closed loop: field needs drive product iteration, and iteration effects return to the field for verification.
iFenxi : How do you view the long-term implementation of ontologies in China?
Hong Wanfu: Whether ontologies can be implemented long-term depends on whether they can systematically help clients solve problems and whether input-output ratios are reasonable; these need industry validation. Promoting knowledge graphs was also hard before; many people didn't understand why to do it. Later, it gradually became a foundational capability; now doing retrieval-augmented generation or dynamic ontologies also uses graphs. Whether technology stays ultimately depends on what problems it solves in business. But truly mastering ontology for an industry is not easy. Industry business logic and data standards must be clearly sorted out, requiring corresponding organizations and systems to push it forward. Industries with better informatization foundations and existing data standards may be relatively advantageous; however, differences between industries are large and cannot be simply analogized. At least from the defense sector we deeply cultivate, this road is still long. I think the heat around ontologies won't fade quickly in the short term because it indeed touches clients' actual pain points. But whether it can develop long-term depends mainly on three aspects: Can the same ontology sustain the second and third scenarios? Can the impact on downstream systems be effectively controlled when ontology changes occur? And when reasoning errors happen, can they be traced to specific bases and error links? If these three trends move in a positive direction, ontologies have a basis for sustained implementation; otherwise, no matter how hot, it's hard to truly stay.
iFenxi : Are clients willing to open up truly valuable business scenarios?
Hong Wanfu: From our contact situation, client needs are clear. Although cross-organization and cross-system collaboration mechanisms are not easy to bridge, many people who truly do business understand these technologies and are willing to bring out important scenarios to discuss whether technology can solve actual problems. Of course, among some ontology application needs we have encountered, there are indeed scenarios with low relevance to core business; even if constructed, how much effect they can exert is questionable. So we cannot just look at whether there are projects or whether the technology is used. We must return to the scenario itself to see what value is generated for clients after the problem is solved.