Glossary

The 8D report is an eight-step standardised procedure for handling customer complaints, from the immediate containment action through root cause analysis to the effectiveness check. Because 8D reports have an almost identical structure over years, they lend themselves excellently to analysis: recurring causes become visible that nobody notices in an individual case.

Also searched as: 8D method · complaint handling · root cause analysis

Additive manufacturing is the layer-by-layer construction of parts, in moldmaking used above all for inserts with conformal cooling. Such inserts allow cooling channels that cannot be drilled conventionally and noticeably shorten cycle times.

Also searched as: metal 3D printing · printing conformal cooling · DMLS

An interdisciplinary field of computer science. AI systems can recognize patterns in large volumes of data and take over tasks that normally require human cognitive abilities (for example costing from historical records, or text analysis).

An AI agent is a system that combines a language model with tools and a goal, and thereby works through multi-step tasks on its own. For example: read the enquiry, search the database for similar past projects, pull material prices from the ERP, draft a costing, present it to the estimator for approval. The last step remains decisive, namely approval by a human.

Also searched as: agentic AI · autonomous AI · AI with tools

AI in plastics processing denotes the use of learning systems along the process chain from quotation through moldmaking and sampling to series production. The most viable use cases today are not autonomous process control but knowledge access, document processing and costing support. There the data exists in sufficient quantity and quality, and the benefit is immediately measurable.

Also searched as: AI injection molding · artificial intelligence plastics industry · AI use cases production

AI literacy is the ability, required by the EU AI Act, of employees to use AI systems knowledgeably and with an awareness of risk. Companies must ensure that staff working with AI understand its possibilities and its limits. In practice this means introduction, training and documented rules of use belong in the project scope.

Also searched as: Article 4 AI Act · AI training obligation · training employees

AI maturity describes how far a company is prepared for the use of AI in terms of data availability, process digitalisation, competence and governance. It is as a rule more important than the choice of model: a business with a well maintained CAQ and structured filing achieves more with a small model than a business running on scraps of paper achieves with the largest.

Also searched as: AI readiness · AI assessment · where to start with AI

AI-supported defect diagnosis is the assignment of an observed part defect to probable causes on the basis of your own defect history. The difference from a textbook: the system knows your molds, your materials and your machines and answers with reference to specific earlier cases instead of general recommendations.

Also searched as: fixing injection molding defects · finding the root cause · troubleshooting injection molding AI

AI-supported mold costing is the derivation of a costing proposal from enquiry documents and historical quotations including post-costing. The system searches for comparable past projects, shows their actual costs and proposes line items. The estimator checks and decides, but gains hours per enquiry and draws on experience that otherwise exists only in individual people’s heads.

Also searched as: automating costing · faster quotations moldmaking · AI pricing

AI-supported process optimisation is the data-based derivation of better settings from machine, quality and environmental data. Realistically it begins with analysis and proposal, not with intervention: the system shows which parameter combinations have historically led to stable quality. Responsibility for the machine stays with the setter.

Also searched as: optimising cycle time with AI · automatic parameters · analysing process data

An air-gapped system is a computer system with no physical connection whatsoever to public networks. Local AI can be run fully air-gapped: the model weights are loaded once, after which the entire operation runs offline. This is the highest achievable level of data isolation.

Also searched as: physically separated network · offline AI · isolated system

An API is a defined interface through which programs talk to one another. It is the precondition for an AI solution to reach the ERP, MES and file storage without anyone anywhere retyping data by hand.

Also searched as: interface · REST · system connection

An audit trail is the complete logging of all relevant operations in a system. For AI applications it covers the question, the sources drawn upon, the model version and the answer. It is the basis for traceability towards customers, auditors and the EU AI Act.

Also searched as: logging · traceability · audit compliance

Automated document processing is the machine extraction of structured data from records, inspection reports, delivery notes and specifications. A combination of OCR, vision model and structured output delivers fields that flow straight into the ERP. For businesses with grown paper and PDF archives this is usually the use case with the fastest result.

Also searched as: reading invoices · digitising delivery notes · capturing records automatically

Automated enquiry screening is the machine pre-assessment of incoming enquiries: drawings and specifications are read, characteristics extracted, feasibility and similarity to existing projects assessed. Businesses that leave enquiries unprocessed for lack of capacity thereby win revenue they currently simply leave on the table.

Also searched as: assessing enquiries automatically · reading drawings · speeding up the quotation process

Software for computer-aided design (CAD) and manufacturing (CAM). It is used to design the 3D models for the tools and to program the milling paths for the CNC machines.

CAQ is the software category for computer-aided quality assurance: inspection planning, measurement capture, complaint and action management. CAQ databases hold the complete defect history of a business and are therefore the single most valuable data source for root cause analysis.

Also searched as: computer-aided quality · QA software · complaint management

A cavity is the hollow space in the mold that forms the shape of the later part. Molds with several cavities produce several parts per cycle. Cavity-related data is central to quality analysis, because individual cavities can deviate systematically, a pattern often overlooked in aggregate data.

Also searched as: nest · mold cavity · multi-cavity mold · 4-cavity mold

Changeover time is the time from the last good part of one order to the first good part of the next. SMED is the methodology for shortening it systematically, essentially by moving activities into running production. With small lot sizes changeover time largely determines profitability.

Also searched as: mold change · setup · single minute exchange of die

Chunking is the splitting of long documents into meaningful sections before they are indexed. The quality of a RAG system depends heavily on it: if an inspection report is cut in the middle of a table of measured values, the search later finds nothing but fragments. Good chunking follows the structure of the document, not a fixed character count.

Also searched as: text splitting · document splitting · chunk size

Clamping force is the force with which the injection molding machine holds the mold closed during the injection phase; it is stated in kilonewtons or tons. It has to exceed the internal pressure, otherwise the mold opens slightly and flash appears. Assigning the right machine is one of the first costing decisions.

Also searched as: tonnage · machine size · calculating clamping force

Cleanroom production is manufacturing under controlled particle and microbial load according to defined cleanliness classes. In medical technology and pharmaceutical packaging it is a precondition. The accompanying documentation duty generates large volumes of structured and unstructured evidence.

Also searched as: cleanroom · ISO 14644 · medical injection molding

A cold runner is an unheated distribution system in which the sprue solidifies with every cycle and arises as waste or regrind. It is cheaper in the mold and more robust in operation, but consumes more material.

Also searched as: cold runner mold · sprue waste · sprue gate

Our term for the central, company-internal data foundation that an AI draws on. It is the collected “brain” of the company: it breaks up island systems and protects decades of employee expertise against staff turnover by making it digitally retrievable.

A confidence score is a numerical value expressing how certain a model is about a statement. In document processing it governs the division of labour: recognitions above a threshold pass through automatically, everything below lands in the manual review list.

Also searched as: confidence · certainty value · how sure is the AI

A container is a self-contained software package including all its dependencies that runs identically on any suitable server. Local AI components are practically always delivered in containers. This keeps installation, updating and rolling back manageable, even without a large IT department of your own.

Also searched as: Docker · containerisation · software rollout

The context window is the maximum amount of text a language model can take into account at once in a single request. Modern local models handle 32,000 to 128,000 tokens, which is easily a complete tool specification including its change history. A large context window is no substitute for RAG, though: it costs compute time and hit quality falls with very long inputs.

Also searched as: context length · how much text can an AI read · 128k context

Model licences govern whether and how an AI model may be used commercially. The range runs from permissive licences such as Apache 2.0 and MIT through to special licences with conditions. Checking the licence of the specific model belongs to due diligence before production use. It is quickly done, and expensive later if it was skipped.

Also searched as: LLM licence · commercial use permitted · Apache MIT Llama licence

Cycle time is the duration of a complete injection molding cycle from mold close to the next mold close. It consists mostly of cooling time and is the most direct lever on unit cost. Even one second less amounts to a considerable sum at volumes in the millions.

Also searched as: takt time · reducing cycle time · increasing output

Data exfiltration is the unintended outflow of company data. In an AI context a copy-and-paste into a browser window is enough. Since tool design and process knowledge are the real value of the firm, local operation here is not a luxury but the protection of substance.

Also searched as: data outflow · protecting trade secrets · loss of know-how

A data lake is a central store for raw data in its original form, such as PDFs, images, machine data and tables. Unlike a data warehouse, the structure is determined only on reading. That is an advantage for AI projects, because at the moment of storage it is not yet clear which question will later be asked.

Also searched as: raw data store · central data storage · data lakehouse

A data processing agreement governs the conditions under which a service provider processes personal data on your behalf. With cloud AI it is mandatory, including scrutiny of sub-processors and third-country transfers. With purely local operation there is no processing on your behalf at all, because no data leaves the company.

Also searched as: DPA · Art. 28 GDPR · cloud provider contract

A data silo is a system whose data cannot be used by other systems. A typical state of affairs in the Mittelstand: ERP, MES, CAQ, file server and individual departments’ spreadsheets know nothing of one another. The biggest lever in an AI project usually lies not in the model but in breaking open these silos.

Also searched as: island system · scattered data · system break

Design of experiments is a method for determining the influence of several parameters with as few trials as possible. In injection molding it is used to optimise melt temperature, mold temperature, injection speed and holding pressure systematically rather than intuitively. Analysing historical DoE results is an obvious AI use case.

Also searched as: DoE · trial plan · parameter optimisation

The diesel effect is a local burning of the plastic that occurs when trapped air is compressed abruptly by the incoming melt. It shows as brown to black marks, usually at the end of the flow path. The cause is almost always inadequate venting, not the material.

Also searched as: burn marks · burning of plastic · venting problem

The digital mold file is the complete, searchable consolidation of all information on a mold: design, revision levels, sampling, process parameters, faults, servicing, shot counts and complaints. It is the precondition for any meaningful AI analysis and even without AI a noticeable gain over scattered folders and email threads.

Also searched as: mold history · mold management · digital mold documentation

Digital sovereignty is a company’s ability to decide for itself about its data, its systems and how they evolve. With cloud AI the provider decides on prices, model versions, availability and shutdowns. With local AI on open-weight models the company decides. A model that runs today still runs in five years, regardless of any provider’s business strategy.

Also searched as: data sovereignty · sovereign AI · technological independence

A digital twin is the digital representation of a real object including its data and history. For an injection mold it covers design, changes, tryout results, process window, servicing and shot counts. It is the logical basis for making mold knowledge queryable instead of spreading it across folders and people’s heads.

Also searched as: digital twin · virtual representation · digital mold file

Distillation is a procedure in which a large model trains a smaller one. The result is a lean model that largely retains the capabilities of the large one for a narrow field of tasks, and runs locally on modest hardware.

Also searched as: knowledge distillation · small model from large

A draft angle is the slight inclination of vertical part surfaces that allows demoulding without damage. One to two degrees is usual, considerably more on textured surfaces. Missing draft is a classic design error that only becomes painfully apparent in the tryout.

Also searched as: taper · demoulding angle · releasability

Edge AI is the execution of AI models directly where the data arises, for example at the injection molding machine rather than in the data centre. The advantage is extremely low latency: an image check on a part has to happen at the pace of the machine, not at the pace of an internet connection.

Also searched as: edge computing · AI at the machine · inference at the network edge

The ejector system is the mechanism that pushes the solidified part out of the cavity. Pins, sleeves and stripper plates have to be placed so that no visible marks appear on visible surfaces and the part does not deform.

Also searched as: ejector · knockout · demoulding · ejector pin

EDM is a material-removing process in which material is eroded by electrical discharges. Sinker EDM reproduces complex contours with a graphite or copper electrode, wire EDM cuts with a thin wire. Both are indispensable in moldmaking for geometries no cutter can reach.

Also searched as: sinker EDM · wire EDM · spark erosion · EDM

An embedding is the translation of a text into a numerical vector that represents its meaning. Texts with similar meaning receive similar vectors. As a result a search for “part warps” also finds reports that speak of “distortion after demoulding”, even though not a single word matches.

Also searched as: vector · text embedding · semantic similarity

The central business software of a company. Resource planning, finance, HR, sales and materials management all come together here. An interface to the ERP is indispensable for efficient automation.

ETL denotes the process of extracting data from source systems, harmonising it and loading it into a target system. For local AI this is the unspectacular but decisive groundwork: without a clean supply of reports, master data and records, every model stays clueless.

Also searched as: extract transform load · data integration · data pipeline

The EU AI Act is the European regulation governing artificial intelligence. It classifies applications by risk and attaches duties of documentation, transparency and human oversight. Operational assistance systems for knowledge search and document processing usually fall into low risk classes; applications touching employees, recruitment selection for instance, are treated more strictly.

Also searched as: AI regulation · AI Act obligations · EU AI rules

Euromap 77 and 83 are industry standards built on OPC UA for connecting injection molding machines and peripherals to higher-level systems. They define how cycle times, process parameters and fault messages are provided in a uniform way, and thereby solve a good part of the classic interface problem across the machine park.

Also searched as: Euromap · connecting an injection molding machine · MES interface injection molding

Evaluation is the systematic measurement of an AI system’s answer quality against a fixed catalogue of questions with known correct answers. Public benchmarks say little about your own use case. What carries weight is a test set of 50 to 200 real questions from your own business.

Also searched as: measuring AI quality · test set · benchmark on your own data

Filling simulation is the computational prediction of filling behaviour, pressure demand, weld line position, temperature distribution and warpage before the mold is built. It moves the discovery of faults out of the tryout and into the design phase, where corrections are cheaper by orders of magnitude. Comparing simulation results with the later tryout reality systematically is a high-grade data treasure.

Also searched as: Moldflow · injection molding simulation · mold filling study · Cadmould

Fine-tuning is the targeted further training of an existing model on your own example data. It changes the behaviour of the model, meaning tone, format and technical vocabulary, but it is the wrong route for feeding in facts. Knowledge comes in through RAG, behaviour through fine-tuning. Keeping the two apart saves a great deal of money in projects.

Also searched as: retraining a model · adapting to company language · model adaptation

Flash is excess material at the mold parting line, arising when melt is pushed into the gap between the mold halves. Causes are insufficient clamping force, excessive injection pressure or wear of the parting surface. Recurring flash at the same spot is an early indicator of mold wear.

Also searched as: overflow · burr · parting line flash

FMEA is a structured method for the preventive analysis of possible failures and their consequences, rated by occurrence, severity and detection. It is mandatory in automotive supply. Existing FMEA documents are among the best structured sources of knowledge in a business and therefore ideal for RAG systems.

Also searched as: failure mode and effects analysis · process FMEA · design FMEA · RPN

Function calling is a language model’s ability to produce structured calls to registered functions instead of only producing text. This lets an AI retrieve today’s stock level from the ERP rather than estimate it. It is the bridge between language understanding and the systems already running in the business.

Also searched as: tool use · AI calling functions · connecting AI to the ERP

Gas-assisted injection molding is a special process in which nitrogen is introduced into the still liquid core of the part and forms a hollow space there. It allows thick-walled, stiff parts without sink marks while using less material, handles and frame structures for instance.

Also searched as: gas injection technology · hollow sections · GIT

A European regulation for the protection of personal data. By using local open-source technologies we ensure that operating your AI systems fully meets the strict requirements of the GDPR.

The German Trade Secrets Act protects business and trade secrets only where the company can demonstrate “appropriate confidentiality measures”. Anyone entering design data into a freely accessible cloud service risks precisely that protection. Local operation is therefore a legal argument as well as a technical one.

Also searched as: trade secret protection · appropriate confidentiality measures · GeschGehG

GGUF is a file format for quantized language models that bundles weights and metadata in a single file. It is the standard format for operation with llama.cpp and Ollama. For production use on graphics cards the safetensors format is often used instead.

Also searched as: GGUF format · local model format · safetensors

The GPU is the decisive compute unit for AI inference, VRAM its dedicated memory. Rule of thumb: the quantized model must fit entirely into VRAM, plus a reserve for the KV cache. A 4-bit quantized 14B model needs around 10 GB, a 70B model about 40 GB. If the model does not fit into VRAM, speed collapses by a large factor.

Also searched as: graphics card AI · VRAM requirement · what hardware for local AI

Groundedness is the measure of how fully an AI answer is covered by the source documents drawn upon. A well built RAG system cites the source for every statement, so the user can check the original report in two clicks. Without source references an AI answer is not usable in a technical environment.

Also searched as: source fidelity · evidence · answer with sources

Guardrails are technical safety rails that check the inputs and outputs of an AI system and block impermissible content. They include topic limits, blocks for personal data and the obligation to support answers with sources.

Also searched as: safety rails · securing AI · output filter

A hallucination is an output of a language model that is factually wrong yet convincingly phrased. It arises because the model produces the statistically most plausible continuation, not the demonstrably correct one. The remedies are RAG with source references, guardrails and the requirement to give no answer where there is no evidence. In a manufacturing setting an honest non-answer is far more valuable than an invented figure.

Also searched as: AI makes things up · wrong answers · AI reliability

Hardening is the heat treatment that raises the hardness of a steel, stated in HRC. The distortion that occurs in the process has to be planned into the manufacturing sequence, usually through finishing after heat treatment.

Also searched as: heat treatment · hardening distortion · HRC

Holding pressure is the pressure maintained after the cavity has been filled, in order to compensate for the shrinkage of the solidifying material. Its level and duration directly influence dimensional accuracy, weight and sink marks. The switchover point from injection to holding pressure is one of the most sensitive parameters in the entire process.

Also searched as: packing pressure · switchover point · volume compensation

A hot runner is a heated distribution system that carries the melt right up to the cavity, so that no sprue waste arises. It reduces material consumption and cycle time but raises mold cost and maintenance effort. Hot runner faults are a frequent topic in maintenance reports and therefore a rewarding case for a searchable knowledge base.

Also searched as: hot runner system · valve gate · runnerless molding

HSC milling is machining at very high cutting speeds with small depths of cut. In moldmaking it allows the milling of hardened steels and partly replaces EDM. The achievable surface quality noticeably reduces polishing effort.

Also searched as: high-speed cutting · hard milling · 5-axis milling

Human-in-the-loop describes processes in which the AI prepares and a human decides. In costing, complaint handling and the release of process parameters this is the only defensible arrangement: the AI supplies a proposal and the evidence behind it, responsibility stays with the expert.

Also searched as: human in the loop · AI approval process · four-eyes principle

Hybrid search combines classic keyword search with semantic vector search. In an industrial setting this is close to mandatory: part numbers, tool numbers and material designations such as “PA6-GF30” have to be found exactly, while a problem description should be found by meaning.

Also searched as: BM25 · keyword and vector · RAG search quality

IATF 16949 is the quality management standard of the automotive industry, built on ISO 9001 and extended by sector-specific requirements. It calls for extensive evidence, defect prevention and continuous improvement and generates a documentation load that is barely manageable by hand.

Also searched as: automotive quality standard · supplier certification · TS 16949

Inference is the application of an already trained AI model to a specific input, in other words the actual operation as opposed to training. In practice local AI consists almost entirely of inference: the model was trained externally but is executed in-house. Inference is orders of magnitude cheaper than training.

Also searched as: model execution · AI answer generation · running a model

An inference server is the software that loads a language model and makes it available to applications through an interface. Widely used examples are Ollama (simple setup, good for getting started and single-user work), vLLM (high throughput for many concurrent users) and llama.cpp (also runs on hardware without a dedicated graphics card).

Also searched as: model serving · AI server · inference engine

The initial sample inspection report is the document in which all drawing characteristics of an initial sample are measured and compared with the target value. It is a mandatory part of every series release. Transferring measurement records into inspection forms is one of the most rewarding automation cases in quality assurance.

Also searched as: ISIR · EMPB · initial sampling · dimensional report

Injection compression molding is a process in which the mold performs a further defined closing movement during or after injection. It produces more uniform internal stresses and is used above all for optical parts and large-area thin parts.

Also searched as: coining · compression molding · optical parts

An industrial primary shaping process in which liquid plastic is injected under pressure into a mold in order to produce plastic parts at series scale.

Insert molding is the overmoulding of inserted parts such as metal bushings, contacts or threaded inserts. It saves downstream assembly but demands reliable insert positions and usually automation. Positional deviation of the insert is a typical cause of defects in series production.

Also searched as: overmoulding · insert · overmoulding metal

An internal AI assistant is a company-wide application that answers questions on the basis of your own documents, with source references and permission checks. It replaces the round of calls asking “does anyone remember how we solved that back then?”. And it is at the same time the most effective remedy against shadow AI, because it offers the legitimate alternative that employees are looking for anyway.

Also searched as: company chatbot · ChatGPT for internal business use · AI assistant for employees

ISO/IEC 27001 is the leading international standard for information security management. Suppliers to the automotive and medical technology sectors are increasingly audited against it. AI systems fall within its scope. With local data storage the evidence is considerably easier to produce than with distributed cloud services.

Also searched as: information security certification · ISMS · customer audit

The knowledge cutoff is the point in time up to which a model’s training data reaches. The model knows nothing after that. This is precisely why RAG matters so much: your tool file from last week appears in no training data set on earth. It has to be supplied at runtime.

Also searched as: state of knowledge · training cutoff · how current is the AI

A knowledge graph is a data structure that maps objects and their relationships to one another, for instance tool → cavity → article → customer → complaint. Combined with RAG (GraphRAG) a system can then also answer questions that require several steps: “Which customers are affected if tool 4711 fails?”

Also searched as: knowledge graph · relationship knowledge · GraphRAG

Knowledge retention denotes the systematic transfer of experience into a retrievable form. In moldmaking it is existential: a setter with thirty years of experience takes thousands of unwritten decision rules with him when he leaves. A local RAG system on tryout reports, 8D reports and maintenance notes makes this knowledge searchable, without anyone ever having had to write it down.

Also searched as: securing experience · knowledge transfer retirement · loss of know-how in production

Kubernetes is a platform for the automated management of many containers across several servers. It pays off from a certain system size onwards; for a single AI server in a mid-sized company it is usually overhead. Honest advice here means: use it only where the complexity actually brings benefit.

Also searched as: K8s · orchestration · managing containers

The KV cache is a buffer in which a language model stores the processing of the text so far, so that it does not have to be recomputed for every new token. It grows with context length and user count and is in practice often the reason why graphics memory runs short, not the model itself.

Also searched as: key-value cache · VRAM usage · context memory demand

llama.cpp is a lean inference implementation that also runs language models on ordinary processors without a dedicated graphics card. It is the technical basis of many local AI tools and the reason why smaller models even run on a well equipped workstation.

Also searched as: llama.cpp CPU · LLM without GPU · run GGUF

A large language model such as ChatGPT or Llama 3. It has been trained on enormous amounts of text data and can process natural language, summarize texts or answer complex questions. We use exclusively open-source LLMs that run locally.

Local AI means running AI models on your own hardware inside the company network, without any data being transferred to external cloud services. Prompt, document and answer never leave your own infrastructure. For businesses with process knowledge worth protecting, such as tool designs, costing bases or customer drawings, this is often the only legally and contractually clean way to use AI at all.

Also searched as: on-premise AI · AI in-house · private AI · AI without cloud

LoRA is a technique that adapts a model through small add-on modules instead of retraining all weights. QLoRA combines this with quantization. Fine-tuning on a single graphics card thus becomes feasible, and the resulting adapters are a few hundred megabytes rather than many gigabytes.

Also searched as: low-rank adaptation · cheap fine-tuning · adapters

The machine hourly rate is the cost of a machine allocated to one operating hour, including depreciation, energy, floor space, maintenance and a share of personnel cost. It is the basis of every part costing and in many businesses has not been updated for years.

Also searched as: hourly machine rate · cost centre accounting · manufacturing cost

Masterbatch is a highly concentrated colour or additive granulate that is metered into the base material in small proportions. Deviations in dosing lead to colour variation and altered processing properties, a frequent reason for complaints on visible parts.

Also searched as: colour batch · additive batch · colouring plastics

MES is the software layer between ERP and machine that schedules, controls and reports back on production orders in detail. It supplies real-time data on orders, states and faults, the pulse of production.

Also searched as: manufacturing execution system · shop floor control · detailed scheduling

The Model Context Protocol is an open standard through which AI applications access data sources and tools in a uniform way. Instead of building a bespoke connection for every system, one MCP server is provided per source, for ERP, MES and file storage. This lowers integration effort considerably and works entirely inside the local network.

Also searched as: MCP · AI interface standard · connecting AI to systems

A model family is a series of language models from the same maker in different sizes. Relevant for local operation are Llama (Meta), Mistral and Mixtral (Mistral AI, France), Qwen (Alibaba) and Gemma (Google). They differ in licence, quality in German and hardware appetite. Picking one is a project decision, not a matter of taste.

Also searched as: Llama · Mistral · Qwen · Gemma · which local model

Mold costing is the advance determination of the manufacturing cost of an injection mold from part geometry, cavity count, material, quantity and quality requirements. In many businesses it is done from experience by a handful of people. Precisely that dependence makes it the most valuable use case for a system that learns from past quotations and actual post-costings.

Also searched as: calculating mold price · quotation costing moldmaking · mold cost estimate

Mold service life is the number of shots a mold achieves before it needs overhauling. It depends on steel grade, material (especially with glass-fibre reinforced types), process management and maintenance discipline. A dependable shot count per mold is the precondition for predictive maintenance.

Also searched as: shot count · mold lifetime · mold wear

Mold temperature control is the deliberate management of mold temperature through media channels. It determines cooling time, warpage and surface quality. Conformal cooling, usually additively manufactured, follows the part geometry and often shortens cycle times by double-digit percentages.

Also searched as: mold cooling · temperature control unit · conformal cooling

Multi-component molding is the production of a part from two or more plastics in one process, for example for hard-soft combinations with TPE sealing lips. Molds with a rotary platen, index plate or transfer technology are elaborate, but they save entire assembly steps.

Also searched as: 2K molding · two-shot molding · hard-soft combination

A multimodal model processes images alongside text, sometimes audio or video as well. In plastics processing this is the key to technical drawings, defect photos of parts and photographed handwritten notes from the tryout. A vision-language model (VLM) can, for example, describe which surface defect is visible in a photo.

Also searched as: vision model · VLM · AI understanding images · read a drawing

NIS2 is the EU directive raising the level of cybersecurity, and it covers considerably more companies than its predecessor, among them many suppliers to manufacturing industries. It requires risk management, reporting channels and evidence. Locally operated systems simplify that evidence, because the attack surface is smaller and data storage clearly attributable.

Also searched as: NIS2 directive · cybersecurity obligation · affected companies

An NPU is a processor component specialised in AI computation, increasingly built into workstations and industrial PCs. Its performance is stated in TOPS (trillions of operations per second). NPUs suit small models and image analysis at the machine, not large language models.

Also searched as: neural processing unit · AI accelerator · TOPS

Automatic optical character recognition. It makes it possible to read printed or scanned documents (old delivery notes, invoices, reports) by machine and convert them into digitally usable text.

OEE is overall equipment effectiveness, the product of availability, performance and quality rate. It condenses plant performance into a single figure. Its value only emerges through detailed analysis of the losses. That, however, often founders on unstructured, handwritten fault notes.

Also searched as: overall equipment effectiveness · calculating OEE · plant utilisation

Ollama is an open-source runtime that starts language models locally with a single command. It encapsulates model handling and provides an interface that your own applications can dock onto. For pilot projects Ollama is often the fastest route from zero to the first locally answered question.

Also searched as: install Ollama · Ollama for business · run a local LLM

Unlike cloud solutions, on-premise systems run directly on the company’s own server hardware. The data therefore stays in the company’s own hands and is not passed on to external service providers.

OPC UA is the vendor-independent communication standard for machine and plant data in industry. It supplies not only values but also their meaning in a uniform information model. For data-driven analysis in injection molding, OPC UA is the preferred route from the machine into the IT system.

Also searched as: OPC Unified Architecture · machine data standard · Industry 4.0 interface

An open-weight model is an AI model whose trained weights may be downloaded freely and operated locally. The distinction matters: “open weights” means the weights are available; “open source” in the strict sense would additionally require open training data and methods. For local operation only the licence of the weights is decisive. It governs whether commercial use is permitted.

Also searched as: open model weights · open source LLM · free AI models

Parameters are the trained numerical values inside an AI model; their count is stated in billions (7B = 7 billion). More parameters usually mean better answer quality, but also higher memory and compute demand. For clearly bounded tasks such as document extraction, models of 7 to 14 billion parameters are often enough; open-ended technical advice benefits from larger ones.

Also searched as: 7B · 70B · model size · how big does an LLM need to be

Abbreviations for production data acquisition and machine data acquisition. These systems collect data from the shop floor, such as machine runtimes, unit counts or fault messages.

The plasticising unit is the assembly of barrel, screw and nozzle that melts, homogenises and injects the granulate. Screw geometry and speed influence how gently the material is treated and how homogeneous it becomes, a critical point with shear-sensitive materials.

Also searched as: screw · injection unit · melting granulate

Post-costing is the comparison of planned with actually incurred costs of an order. Without it every advance costing remains a guess. The systematic plan-versus-actual comparison over years is the training and reference basis for any data-supported quotation support.

Also searched as: plan versus actual · actual cost · costing quality

PPAP (per AIAG) and PPF (per VDA) are the standardised procedures for the customer approval of series parts. They cover the initial sample inspection report, process evidence, FMEA, gauge capability and further documents. Assembling this documentation ties up considerable time of qualified staff.

Also searched as: automotive sampling · production part approval · initial sample

Predictive maintenance is condition-based, forward-looking maintenance on the basis of operating data. Instead of fixed intervals, shot counts, pressure curves, temperatures and fault patterns determine the moment of servicing. The precondition is a dependable data history per mold and machine.

Also searched as: condition-based maintenance · planning maintenance · predicting failure

Process capability is the statistical measure of how reliably a process stays within the tolerance limits. cp describes the scatter, cpk additionally the position relative to the centre of the tolerance. In automotive supply, cpk values of 1.33 or 1.67 are commonly required thresholds.

Also searched as: cpk value · capability index · capability study

The process window is the range of settings within which a process reliably delivers good parts. The wider the window, the more robust the production run against variation in material, environment and machine. It is determined systematically, for instance through design of experiments.

Also searched as: process limits · parameter window · robust process

A prompt is the instruction that gets a language model to perform a task. Prompt engineering is the systematic design of those instructions. In enterprise solutions prompts are not the user’s job but part of the application. The employee asks the question in plain language, the fine control sits in the system.

Also searched as: input instruction · prompt engineering · how to ask an AI

Prompt injection is an attack in which instructions hidden inside processed content redirect the behaviour of an AI system. A prepared supplier PDF can carry an instruction that moves the AI to unwanted output. Protection comes from strict separation of instruction and content, output filters and limited tool permissions.

Also searched as: attack on AI · manipulated documents · AI vulnerability

A proof of concept is a tightly bounded undertaking that demonstrates the feasibility of an approach on real data. What works well is a use case with a measurable result and a manageable data set, for example the automatic analysis of two years of tryout reports. Whatever does not work here will certainly not work at scale.

Also searched as: PoC · pilot project · AI feasibility study

Quantization is the reduction of a model’s numerical precision in order to lower memory demand and compute effort. A model quantized to 4 bit occupies roughly a quarter of the original memory, usually with only a small loss of quality. Quantization is what makes it possible to run capable models on affordable hardware in day-to-day operation.

Also searched as: LLM quantization · 4-bit · Q4_K_M · shrink a model

An advanced AI method. It connects a language model (LLM) with your own knowledge base (a vector database). When you ask the model a question, it first searches your own tryout reports and records for relevant information and only then generates the answer. This is how the system becomes your COMPBRAIN.

Rapid tooling is the accelerated production of molds for prototypes and small series, often from aluminium or with a simplified structure. It shortens the time to the first series parts considerably, at limited service life.

Also searched as: prototype mold · aluminium mold · pre-series mold

Recyclate is reprocessed plastic from production waste (post-industrial) or used products (post-consumer, PCR). Its use is increasingly demanded by regulation and by customers, but it brings fluctuating properties with it. Mastering exactly those fluctuations calls for closer process monitoring.

Also searched as: regrind · recycled content · PCR

Reranking is a downstream step in which a specialised model sorts the retrieved passages once more by how well they actually fit the question. It often lifts the answer quality of a RAG system more than a larger language model would, at considerably lower cost.

Also searched as: cross-encoder · result quality · sorting results

Residual moisture is the water content in the granulate before processing. Hygroscopic materials such as PA, PC, PET and ABS have to be pre-dried, otherwise streaks, bubbles and losses of strength occur. Drying records are an often underrated data source in the root cause analysis of batch problems.

Also searched as: material drying · hygroscopic · drying time PA

A role and permission model governs which user may access which data. In RAG systems it is security-critical: the permission check has to take effect at retrieval, otherwise the AI quotes personnel file content to a shop-floor employee. The authorisation model belongs in the search layer, not in the prompt.

Also searched as: permissions · RBAC · who may see what

The runner system is the channel through which the melt travels from the machine nozzle into the cavity. Position and size of the gate largely determine the filling pattern, weld line position, warpage and appearance. Changes to the runner are therefore among the most consequential interventions in a mold optimisation.

Also searched as: sprue · gate · manifold · gate position

Semantic search finds content by its meaning rather than by exact character strings. In filing structures that have grown over time this is decisive, because the same thing has been named differently over the years. “Sink mark”, “sink” and “dent” all refer to the same phenomenon in many reports.

Also searched as: meaning-based search · search by sense · intelligent search

Shadow AI is the uncontrolled use of external AI services by employees without management approval. It rarely arises from ill will but from time pressure, and it results in costings, drawings and customer data ending up in other people’s data centres. An approved internal tool is more effective than any ban.

Also searched as: shadow AI · ChatGPT banned at work · unauthorised AI use

Shrinkage is the reduction in volume of the plastic as it cools and solidifies. It is allowed for in mold design through an oversize and depends on material and process. With semi-crystalline materials such as PA or POM it is considerably higher than with amorphous ones such as ABS or PC.

Also searched as: shrinkage allowance · processing shrinkage · post-shrinkage

A sink mark is a shallow depression on the part surface that appears when material shrinks internally. It occurs preferentially at accumulations of material, for example over ribs and bosses. Countermeasures are wall thickness optimisation, higher holding pressure or a changed gate position.

Also searched as: sink · dent · surface defect injection molding

A slide is a moving mold element that makes undercuts on the part possible by retracting sideways before ejection. Slides raise mold cost, maintenance effort and susceptibility to faults considerably. Avoiding them through part design is a substantial cost lever.

Also searched as: side action · undercut · core pull

A small language model is a language model with typically fewer than ten billion parameters that runs on commodity hardware. For narrowly defined tasks such as classifying fault messages or extracting fields from delivery notes, SLMs reach the quality of large models at a fraction of the cost. The art lies in choosing the smallest model that still solves the task reliably.

Also searched as: SLM · small language model · lean AI

SPC is statistical process control, in which measured values are recorded continuously and monitored on control charts. The aim is to spot trends before scrap arises. The resulting time series lend themselves excellently to pattern recognition.

Also searched as: statistical process control · control chart · quality control loop

Speech-to-text is the automatic conversion of spoken language into text. Locally operable models such as Whisper achieve usable results even with shop-floor noise. The practical benefit: the setter dictates an observation at the machine instead of typing it up later from memory, or not recording it at all.

Also searched as: speech recognition · Whisper local · dictation on the shop floor

Standard mold components are standardised parts such as mold plates, guides, ejectors and cooling connections. They shorten design and procurement considerably. Using standard components instead of bespoke design is one of the most effective cost levers in moldmaking and is historically well documented in costings.

Also searched as: Hasco · Meusburger · standard parts moldmaking · mold base

Start-up scrap is the scrap that arises when ramping up production until stable conditions are reached. It scales with the frequency of changeovers and, with small lot sizes, forms a considerable block of cost that many businesses do not record cleanly.

Also searched as: setup scrap · start-up parts · scrap at ramp-up

Structural foam molding is a process in which a blowing agent creates a fine-celled structure inside the part. The results are lower weight, less warpage and lower cavity pressure, though with limited surface quality.

Also searched as: MuCell · physical foaming · lightweight plastics

Structured output means forcing a model to answer in a fixed, machine-readable format. Only this turns AI into a reliable component of a process chain: a scanned initial sample inspection report yields defined fields that software can process directly, without interpreting free text.

Also searched as: JSON output · structured extraction · fixed response format

Surface texture is the deliberately produced finish of the cavity surface, which is reproduced on the part. Designation systems include VDI 3400 for eroded textures and Mold-Tech for grains. High-gloss surfaces for optical parts require elaborate manual polishing, an often underestimated block of cost.

Also searched as: graining · eroded texture · VDI 3400 · Mold-Tech · high-gloss polishing

The system prompt is the basic instruction hidden from the user that fixes the role, tone, language and limits of an AI system. This is where it is anchored that answers are given solely on the basis of the stored company documents and that gaps in knowledge are stated openly instead of guessed.

Also searched as: system instruction · set the AI role · controlling behaviour

Temperature is a parameter that controls the randomness of a model’s output. A low value produces reproducible, factual answers; a high value produces variation. For costing, extraction and technical information a very low value is chosen deliberately, because the same question should yield the same answer.

Also searched as: creativity setting · deterministic answers · sampling temperature

Thermoplastics are plastics that soften repeatedly on heating and are therefore injection mouldable and recyclable; thermosets cross-link irreversibly; elastomers are rubber-elastic. Within thermoplastics a distinction is drawn between amorphous types (ABS, PC, PS) and semi-crystalline ones (PP, PA, POM, PBT). That difference has direct consequences for shrinkage, warpage and process management.

Also searched as: types of plastic · PP ABS PA POM · semi-crystalline amorphous

Thin-wall molding denotes the injection molding of parts with very small wall thicknesses, typical in packaging. It requires high injection speeds and pressures as well as short cycle times and leaves hardly any process latitude. Close process monitoring pays off particularly well here.

Also searched as: thin wall molding · thin-walled packaging · high injection speed

Time to first token is the interval between sending a request and the first visible character of the answer. It largely determines whether a system is perceived as fast. In RAG systems the time for the document search is added on top.

Also searched as: TTFT · response delay · AI latency

A token is the smallest processing unit of a language model, roughly a fragment of a word. In German 1,000 tokens correspond to about 700 words. Tokens are the unit of account for speed and, with cloud services, for cost. With local AI token costs disappear entirely; only hardware utilisation still counts.

Also searched as: LLM token · token count · word pieces AI

Tokens per second is the measure of a language model’s output speed. From roughly 15 to 20 tokens per second an answer feels fluid to a human, because it appears faster than it can be read. For sizing hardware this figure is far more concrete than any manufacturer specification.

Also searched as: t/s · local LLM speed · how fast is local AI

DIN 16742 is the standard for tolerance specification on plastic molded parts. It takes account of the fact that plastic dimensions scatter depending on material, mold and process. Applying it cleanly prevents a large share of the disputes between customer and supplier during sampling.

Also searched as: plastic part tolerances · dimensional tolerance injection molding · tolerance group

Tool steel is the material group for mold inserts and plates, selected by hardness, toughness, polishability and corrosion resistance. Common grades are 1.2311, 1.2312, 1.2343 and the corrosion-resistant 1.2083. With glass-fibre reinforced plastics the choice of steel largely determines service life.

Also searched as: 1.2343 · 1.2311 · mold steel · hot-work steel

The highly specialized industrial discipline that designs, mills and assembles the custom molds (tools) for injection molding and other forming processes. It demands the highest precision and a great deal of expert knowledge.

TCO is the total cost of a system over its life, that is hardware, electricity, maintenance, support and further development. Local AI has higher initial and lower running costs than cloud AI. Experience puts the tipping point at steady use by several departments. From there, your own hardware pays for itself within a few years.

Also searched as: total cost · cost of local AI · cloud vs on-premise cost

Traceability is the ability to trace a part back to batch, machine, mold, cavity and time of manufacture. In medical technology and automotive it is mandatory. In the event of a complaint it decides whether a recall covers a hundred or a hundred thousand parts.

Also searched as: batch tracking · cavity identification · lot traceability

The test run for a newly manufactured mold. It checks whether the produced part meets the quality requirements. The findings from tryouts (sources of defects, pressures, temperatures) are valuable historical data that belong in COMPBRAIN.

A vector database is a store that holds embeddings and finds the entries closest in meaning to a query at high speed. It is the memory of a RAG system. Locally operable examples include Qdrant, Weaviate, Milvus and pgvector as an extension for PostgreSQL.

Also searched as: vector store · Qdrant · semantic index · AI knowledge base

Vendor lock-in is dependence on one provider from which switching is possible only at disproportionate expense. With cloud AI it creeps in through proprietary interfaces and data formats. Open models, open formats and keeping your own data hold the door open.

Also searched as: provider dependence · switching costs · supplier tie-in

Venting denotes the deliberate gaps and channels in the mold through which displaced air escapes during filling. They are typically a few hundredths of a millimetre deep. Clogged vents are one of the most frequent and most quickly remedied causes of filling problems.

Also searched as: vent channel · air trap · mold vacuum system

Visual quality inspection with AI is the automatic detection of part defects on camera images directly at the line. It detects flash, sink marks, discolouration and incomplete filling in cycle. Local operation is technically imperative here. The latency of a cloud connection does not fit into a cycle time of a few seconds.

Also searched as: optical inspection AI · camera inspection · detecting surface defects

vLLM is an inference server for productive multi-user operation of language models. Thanks to efficient memory management (PagedAttention) the same graphics card serves considerably more concurrent requests than under naive execution. For a plant with 50 users on a shared model this is the difference between usable and unusable.

Also searched as: vLLM throughput · LLM in production · PagedAttention

Warpage is the unwanted deviation in shape of a part caused by uneven shrinkage. Causes are differing wall thicknesses, uneven temperature control, fibre orientation or demoulding too early. Warpage is the classic case in which root cause analysis depends heavily on the experience of individual employees.

Also searched as: part warpage · deformation · dimensional deviation injection molding

A weld line forms where two melt fronts meet and no longer fuse completely. It is visible on the surface and mechanically a weak point. Its position can be predicted in filling simulation and shifted deliberately into uncritical areas through the choice of gate.

Also searched as: knit line · flow line · meld line

Zero trust is a security model in which no access is trusted by default, not even from the internal network. Every request is authenticated and authorised. For AI systems that reach across all departments to access documents, this is the appropriate basic stance.

Also searched as: zero-trust architecture · network security · access model