Haomai Technology AI Brain

From enterprise data to trustworthy answers: the SebAgent approach

Explore how Haomai Technology AI Brain brings together business systems, enterprise knowledge and equipment data through semantics, retrieval and traceable answers.

Systems integrationUpdated November 202612 min read
Manufacturing setting
Manufacturing setting

Make enterprise data open to follow-up questions

Enterprise information is spread across ERP, CRM, manufacturing execution, equipment systems and documents. Answering one business question can mean finding data across several systems and agreeing on fields, metrics and reporting periods. SebAgent is designed to add a path from a question to its supporting evidence above the systems already in use.

This article introduces the architecture of enterprise AI. It explains data connections, business semantics, analytical questions and knowledge retrieval. Examples illustrate a workflow; capabilities, deployment options and performance depend on the delivered version and project validation.

Start with the data estate

Business information includes orders, inventory, receivables and production records, as well as equipment readings, quality logs, drawings and maintenance manuals. Their structure, refresh schedules and access permissions differ. Each source needs an owner and an agreed scope of use.

In manufacturing, orders and work orders inform delivery; equipment and quality records help investigate anomalies; documents and process specifications add context. Available connections and permissions determine which sources can be combined.

Five layers from question to data

The architecture follows a layered system from user entry points to enterprise sources. The layers make it easier to discuss where data lives, who confirms a definition and how an answer can be checked.

LayerResponsibility
InteractionReceive questions and present answers in web, enterprise or collaboration applications.
Agent and task orchestrationInterpret intent, select tools and organize queries, retrieval and next steps.
Semantics and business definitionsDescribe business objects using entities, relationships, terms and governed metrics.
Data access and executionAccess authorized sources through connectors, run permitted operations and record provenance.
Enterprise data and knowledgeKeep databases, business systems, equipment data and documents within their respective management boundaries.

How a natural-language question becomes an answer

For a question such as “Which products had the highest sales last month?”, the system needs an agreed period, definition of sales, product scope and access permissions. When a source cannot support a step, it should say so rather than fill the gap with a plausible number.

  1. Identify the question and reporting period; ask for clarification when needed.
  2. Connect business terms to metric definitions and accessible sources.
  3. Plan the query or knowledge retrieval path.
  4. Check role, field, row and source permissions.
  5. Execute allowed operations and handle results or errors.
  6. Present a finding, chart or detail with units and definitions.
  7. Keep the evidence and context available for follow-up.

Sales figures lead back to order records. A discount exception leads back to the pricing policy in force. When an answer uses both, open each source to check the relevant claim.

Check the evidence from a single query

Ontology, metrics and rules create a shared language

“Sales” can mean order value, invoiced value or cash received. A semantic layer organizes business objects, relationships, synonyms and metric definitions so that people asking questions and people managing data can discuss the same meaning.

The material also describes rule expressions, version history and previews: review the conditions, impact and permissions before a rule takes effect, and retain an audit trail after changes. Specific rule features and execution boundaries need validation in the target environment.

Bring analysis back to everyday work

These questions cover manufacturing, supply-chain, equipment and knowledge scenarios. They illustrate possible lines of inquiry and contain no customer operating data.

  • Production and quality: Which changes in yield deserve attention? Can we inspect the relevant batches, machines and defect records?
  • Supply chain and operations: Which changes across orders, inventory and receivables need checking? What period and definition does each figure use?
  • Equipment and maintenance: After a recurring stoppage alert, what do the equipment records and maintenance manual each tell us?
  • Engineering knowledge: Which version of a drawing, process specification or manual is current, and what page supports the answer?

Retrieve knowledge and check its source

Database answers and document answers need different evidence. The first should identify sources, filters and aggregation; the second should identify the document, version, location and cited passage. A multimodal knowledge workflow can bring text, tables and diagrams together in a single knowledge workflow.

A manual and a live machine status are also different kinds of evidence. A document can explain an operating procedure without establishing the machine’s current state. Readers need clear notice when permissions change, documents are updated or a source is no longer available.

Move from one answer to an ongoing workflow

Agent work can be organized around skills, tools and task orchestration: select the capabilities required by a question, read authorized information, produce a result and return steps requiring confirmation to a person. Confirmed business knowledge and rules can provide context for later tasks.

Analysis, report generation, notifications and changes to business data have different risk profiles. Before executing an action, establish who can initiate it, which steps require approval and how the outcome is recorded. Available actions depend on the delivered version.

How did the outcome differ from the plan?

Suppose a team expects a new channel to bring in more orders, but by month end only inquiries have increased. For the review, record what the team expected, what happened and which queries support that assessment.

  • GET /api/v1/insights/decisions/{session_id}/{seq}
    Read one decision record from the specified session.
  • POST /api/v1/insights/decisions/diff-summary
    Generate a summary of the difference between expected and actual outcomes. Supplying a decision ID saves the summary to that record.

The request fields for the expected outcome, actual outcome and related queries are expected_outcome, actual_outcome and query_record_ids.

Read the public API specification

Carry permissions, security and audit through the workflow

Access control does not end at sign-in. Sources, tables, fields, rows, documents, caches and exports can all affect what a person may see. Identity changes and tenant isolation also belong in the design.

Protect connection credentials, query execution and logs separately, and retain audit records for important operations. Validate measures such as read-only access, masking, query limits and production resource isolation in the customer environment.

Integration and deployment choices

Integration options include API, SDK, command-line, MCP, single sign-on and collaboration-message integration paths, with the aim of bringing analytical and knowledge capabilities into existing workspaces. Confirm interfaces, connectors and identity mapping against the documentation and environment for the version being deployed.

Deployment planning may consider on-premises infrastructure, virtual machines, containers or enterprise cloud environments. Review networking, data location, identity, operations and recovery together. Timelines and hardware requirements need a project-specific assessment.

SebAgent and Palantir AIP: eight comparison dimensions

The source material sets out eight comparison dimensions. Competitor descriptions below reflect publicly available documentation and require scenario-specific validation.

DimensionSebAgent materialPalantir AIP / evaluation focus
Rules and inferenceThe material describes SWRL-style rules, condition combinations and instance-level inference with review and audit.Foundry Rules provides low-code rules; Functions supports TypeScript and Python. Compare semantics and execution for the intended use case. [1][4]
Natural-language ontology modelingMulti-turn clarification, drafts and confirmation help organize ontologies and rules.Palantir documents natural-language ontology creation and updates through AI FDE; availability depends on configuration. [3]
Interviews and glossaryQuestions, confidence levels and human confirmation help maintain business terminology.Compare provenance, confirmation and ongoing maintenance in both environments.
Federated accessVirtual ontology mappings and federated queries access authorized sources.Foundry supports data synchronization and virtual tables for selected sources; connector and pushdown support varies. [2]
Governed metricsPreview SQL and sample values, then use certification and versioning to review definitions.Compare metric definition, reuse, certification and audit in the actual workflow.
Natural-language questionsSource, business terms, synonyms and follow-up context form the question path.Evaluate using the same questions, permissions, data scope and reference answers.
Localization and channelsThe material lists Chinese and English interfaces and channels including WeChat, Lark and Teams; confirm version support.Compare interface languages, authorization, channels and regional delivery.
Deployment and costThe material describes private deployment and hardware tiers; actual cost depends on configuration and quotation.Calculate TCO using equivalent data scale, licenses, hardware, implementation, operations and term.

SebAgent and WorkBuddy: eight comparison dimensions

The supplied material also compares WorkBuddy across eight areas. Connectors and Skills are documented by Tencent; test each workflow with equivalent access and data.

DimensionSebAgent materialWorkBuddy / evaluation focus
Skills and agentsThe material describes creating, installing and updating skills and orchestrating agent tasks.Tencent documents Skills working with MCP tools; compare authorization, distribution and maintenance. [6]
Interactive pagesThe material describes generating interactive pages, binding enterprise data sources and setting access permissions.Test page outputs and connectors against the same scenario.
Business data questionsThe material lists multiple connectors and natural-language answers with sources; confirm version support.WorkBuddy documents MCP connectors and data queries; database access requires the right tools, permissions and configuration. [5]
Business modelsThe material lists qualitative frameworks, quantitative models and scenario combinations, with expert knowledge packaged as skills.Compare actual skills, definitions, inputs and the ability to verify outputs.
Ontology and rulesOntology semantics, rules and instance inference organize business knowledge.Compare how semantics, rules and cross-system context are configured.
Governed metricsMetric definitions, previews and certification support shared business meaning.Check definition, permissions, versions and consistency on the same metric.
Private deploymentThe material describes private deployment and federated queries; validate data boundaries in the target environment.Compare deployment choices, network boundaries, data flows and operating responsibilities.
Channels and modelsThe material lists collaboration channels and configurable model endpoints; confirm version support.WorkBuddy offers tools and connectors; model and channel choices depend on version and configuration. [5][7]

Start with a question that can be checked

Choose one business question and document its data sources, expected definition, reference answer and access roles. Then check source traceability, failure messages and whether permissions behave as intended.

After the first evaluation, expand the data sources, scenarios and collaboration workflows. That gives product teams, business owners and implementers a shared body of evidence for the next decision.

Further reading

Read the full SebAgent, Palantir AIP and WorkBuddy comparison

Back to all articles

Try “permissions,” “data source,” “manufacturing” or “API.”