APS Scheduling Queries

APS scheduling queries: numbers with their definition and sources

A planner spends half the day finding numbers. This week's capacity load, which grade changeovers dominate the schedule, how far a machine's actual cycle time sits from standard: the answers are spread across ERP, MES, the scheduling spreadsheet and past work orders. Answering an ad-hoc question usually means raising a ticket with IT, waiting for an export, then pivoting it, and the definition may not even match last month's table.

RatioMax shortens that loop to a single question. A dedicated data-analysis agent turns natural language into a query, runs it inside a scoped sandbox, and returns the result with the metric definition and its source attached. Definitions, process notes and changeover history live in the same knowledge base and are cited alongside the numbers. It does not reschedule production or write back to ERP; the planner and the APS still make the plan.

Where it breaks

  1. Every number needs a ticket

    An ad-hoc question spans ERP, MES and the scheduling workbook. A planner cannot assemble one table with a consistent definition on their own, so they raise a request and wait for IT. By the time the export arrives the question has usually changed.

  2. One metric, two calculations

    Capacity, load and attainment are computed differently in different reports. Two managers bring their own figures to the meeting, argue past each other, and end up re-deriving the original definition on a whiteboard.

  3. No place to look up experience

    How long cleaning takes on a grade changeover, what constrains a die swap on a given press: some of it sits in process cards, the rest only with senior staff. The planner estimates a number from memory and moves on.

  4. Rescheduling with no reference

    A rush order, a breakdown or a late material forces a reschedule. How the last comparable disruption was handled, and which order was sacrificed, is not recorded anywhere, so the same people are gathered to debate it again.

How RatioMax responds

  • Ask the numbers directly

    A dedicated data-analysis agent turns a question like which machine carried the highest load last week into a query, executes it inside a scoped sandbox and returns the result. Anything that writes or touches sensitive fields stops for human confirmation instead of changing production data.

    Platform capability: NL2SQL queries

  • Numbers carry their definition

    The answer includes the definition behind the metric and the document it came from, so the calculation and the data range can be checked. Definitions live in the knowledge base, where every change keeps a snapshot and a line-level diff.

    Platform capability: Traceable citations

  • Experience in the same library as data

    Changeover durations, cleaning requirements and equipment limits are captured as wiki pages and a knowledge graph in the same knowledge base as the data queries, so one question returns both the current figure and how it was handled before.

    Platform capability: Auto wiki

  • Planning data scoped per site

    Spaces follow plant, workshop and product line, and four-level RBAC narrows access down to a single knowledge base, so planning data opens only to the roles that need it and every access lands in a per-space audit log. The whole stack runs on your intranet.

    Platform capability: Space-level RBAC

Deployment scenarios

  • Questions before the weekly plan

    Before writing the weekly plan, ask for capacity load, open work orders and expected changeover count, so the bottlenecks that need coordinating in advance surface first.

  • Preparing a changeover

    Before a change, ask how long that grade's last switchover took and how far cleaning went, so the planner reserves a window based on record rather than on a single fixed estimate.

  • Reference for a reschedule

    When a rush order or breakdown lands, retrieve how comparable disruptions were handled and what was traded away, then decide which order goes first with that record in hand.

  • Bringing a new planner up to speed

    A new planner learns definitions and the scheduling process by asking their own questions, and each answer cites the policy document or past work order behind it instead of demanding the whole report glossary first.

Terms and knowledge entities

Advanced planning and scheduling
A system that sequences production and sets start times under finite capacity and material constraints, producing an executable schedule. It is not the same layer as the execution-focused MES.
Weekly schedule
The execution-level plan that breaks a monthly target down to week, machine and shift, satisfying due dates, changeovers and equipment availability at the same time.
Capacity load
Occupied hours against available hours for a machine or operation in a period. The ratio depends on its definition, and two figures computed differently cannot be compared.
Changeover time
The downtime from the end of one batch to the first conforming output of the next, covering cleaning, die change, adjustment and first-article confirmation. It constrains every schedule.
Grade changeover
Moving production from one material grade to another in compounding. Each switch costs cleaning losses and first-article scrap, and the number of switches directly affects output.
Work order
The unit that carries one production task, recording product, quantity, routing and planned start and finish. Scheduling and reporting both point at it.
Scheduling constraints
The limits a schedule must satisfy together: machine capability, tooling availability, material arrival, due dates and shift patterns. Ignore any one and the plan cannot be executed.
NL2SQL
Translating a natural-language question into a structured query. Whether the answer is right depends on how well table structure, field meaning and metric definitions were understood.
Metric definition
How a metric is computed and what it covers: numerator, denominator, time window and exclusions. Two figures built on different definitions mean nothing when placed side by side.

These definitions are also exposed structurally so search engines and AI engines can read and cite the concepts accurately.

Related deployment

Guangdong Zhongsu New Materials Co., Ltd.

Stock code 301686 · National “Little Giant” specialised enterprise

Founded in 2009, focused on R&D, production and sales of modified engineering plastics, with three production bases in Dongguan Chang’an, Ji’an (Jiangxi) and Dongguan Shuixiang.

Solution: In the Hongmei plant’s AI-powered APS scheduling project, RatioMax delivers the enterprise knowledge base and APS data Q&A, fusing RAG with NL2SQL.

  • ≤30 min Planning time (from 4–8 h per run)
  • 80%+ Faster planning-data queries
  • 30–45% Fewer grade changeovers
  • 10% Higher overall plant output

Results shown are project targets

Solution questions

Does this system schedule production automatically?
No. It answers questions and holds knowledge: it translates a planner's question into a query, returns the result with its definition, and keeps changeover and disruption experience retrievable. Scheduling stays with the planner and the APS, and the system neither writes to ERP nor trades off due dates for you.
Can an NL2SQL query return the wrong number?
It can, which is why every step is checkable: the query runs inside a scoped sandbox, and the result carries its metric definition and source document for review. Definitions are versioned in the knowledge base with snapshots and diffs. A planner still confirms the number before it drives a decision, and speed does not remove that step.
Our planning data sits on an isolated network. Can it run there?
Yes. Full on-premise deployment is supported on Docker Compose or Kubernetes, with four core services that deploy and scale independently, so data never leaves your network. Each knowledge base can bind a private OSS or S3 bucket, credentials are encrypted with AES-256-GCM, and outbound requests pass SSRF protection.
Does it connect to the ERP and MES we already run?
Through the API and the MCP service. The platform exposes around 360 API endpoints, and the MCP server offers 29 standard tools over stdio, SSE and HTTP with OAuth2 authorisation. Feishu, Notion, Yuque, RSS and the CLI also work as sources. The system reads for question answering and does not take on transactional writes.
How do we settle conflicting metric definitions?
Write the definition down as a knowledge base entry: who set it, how it is computed, what it excludes, with a version snapshot. Auto-generated wiki pages then link related definitions together. Every subsequent query cites the same entry, so the argument moves from which number is right to which definition is right, and changes leave a diff.
Are the Zhongsu case results already achieved?
No. The planning time, query efficiency and changeover reduction published on the site are project targets for that engagement, not verified outcomes. The case page carries the same note: the figures are the targets the project set, and we do not present targets as delivered results.
Who can see which planning data?
Spaces separate plants and four-level RBAC narrows access down to a single knowledge base, so planning, the shop floor, purchasing and finance each see their own scope, with every access written to a per-space audit log. OIDC handles single sign-on, two-tier API keys scope integrations, and external systems receive only the authorised slice.
Can a query result be traced afterwards?
Yes. Each answer records the passages it matched, the query it generated and the scope it ran against, and Langfuse plus W3C cross-service tracing keeps the whole chain for replay. A later review can see which definition version and which data were behind the number.

Other solutions