Skip to content

Services · Digital & automation

AI for manufacturers: agents and a documentation assistant

We build an assistant that answers questions from your plant's documentation and cites the source page, and agents that take a request for quotation, a purchase order or a breakdown report and prepare a summary, a draft or an entry for an employee to approve. The language model runs on a server at your plant or in a Polish data centre.

What we do

  • Technical documentation assistant with source citation
  • Document index: manuals, schematics, drawings, service reports
  • Data extraction from drawings, scans and purchase orders
  • Agent for requests for quotation and preliminary costing
  • Order agent: item codes, discrepancies, entry into the ERP
  • Supplier quote comparison and material certificate checks
  • Machine documentation draft and list of gaps
  • Language model at your plant or in a Polish data centre
  • Access rights, action log and rules for using AI

Example applications

The scope of documents and the course of the task depend on the application. The first two are an assistant that answers questions. The other four are agents that prepare a result for an employee to approve.

Documentation

Technical documentation assistant

Reads

  • Machine manuals, operating instructions and component data sheets
  • Hydraulic and electrical schematics, parts lists
  • Drawings and specifications from the order archive
  • Procedures and work instructions

Prepares

  • An answer to a question asked in plain language
  • The document name, revision and page the answer comes from
  • A note of the discrepancy when two documents give different values
  • A statement that the documents contain no answer

What you gain

An employee gets the tightening torque, part number or valve setting without searching binders and folders. They check the answer in the source document, which opens at the cited page.

Maintenance

Breakdown report and repair history

Reads

  • The breakdown report: description, photo, alarm code from the PLC
  • Repair history and service reports for that machine
  • The machine manual and schematics
  • Spare parts stock

Prepares

  • Similar past breakdowns and the repairs applied
  • A description of the alarm and the part of the schematic it concerns
  • A parts list with catalogue numbers and stock levels
  • A draft repair report from the technician's note

What you gain

The technician starts the repair with the machine's history and a parts list — also on a shift where nobody knows that machine. The repair report adds to the history for later reports.

Sales

Request for quotation with a drawing

Reads

  • The enquiry message and attachments: PDF, STEP, DXF
  • The drawing: title block, material, overall size, tolerances, threads, roughness
  • Earlier orders for similar parts with actual times
  • Material price list and work centre rates

Prepares

  • A summary of data read from the drawing, with uncertain items marked
  • Similar completed orders with price and lead time
  • A preliminary costing according to your company's rules
  • A list of questions for the customer and a draft quote

What you gain

The process engineer starts the quote from a ready summary and checks the items marked as uncertain. An employee sets the price and sends the quote.

Orders

Customer order and bill of materials

Reads

  • The order as a PDF, a spreadsheet or the body of a message
  • The bill of materials from an assembly drawing or from CAD
  • The item master in the ERP
  • The quote the order refers to

Prepares

  • Order lines with ERP item codes assigned
  • A proposed item code for lines with no match
  • Discrepancies between order and quote: price, quantity, date, drawing revision
  • An order for approval in the ERP

What you gain

The order is not retyped by hand. A difference in drawing revision or price is detected before production starts.

Purchasing

Supplier quotes and material certificates

Reads

  • Supplier quotes in different layouts: PDF, spreadsheet, message body
  • The purchase enquiry with its specification
  • 3.1 inspection certificates delivered with the material
  • Requirements of the order and the material standard

Prepares

  • Quotes side by side in one table: price, lead time, terms
  • Items a supplier omitted or changed against the specification
  • A comparison of the certificate with the order: grade, heat, composition, properties
  • The certificate assigned to the delivery and the order

What you gain

The buyer compares quotes in one layout. A certificate that does not match the order is detected at goods receipt, not when the documentation for the customer is being compiled.

Design

Machine documentation

Reads

  • The machine model and bill of materials
  • Component documentation from suppliers
  • Your document templates and the documentation of similar machines
  • The list of standards and requirements adopted in the project

Prepares

  • A draft operating manual and technical file to your template
  • A spare parts list and inspection tasks taken from component documentation
  • A list of gaps: supplier documents, declarations, sections to complete
  • A working translation for review

What you gain

Documentation is written alongside the project, not in the week of acceptance. The content is checked and signed by the person responsible for the machine's conformity.

From document to answer

  1. Preparation

    files and scans
    1

    Documents

    PDF files, scans, drawings and spreadsheets stay in their current folders. You specify which sets the assistant covers.

  2. passages with a source address
    2

    Index

    We split the content into passages labelled with document name, revision and page. The index is updated when a file changes.

  3. Continuous operation

    question and access rights
    3

    Question

    An employee asks a question in plain language — in a browser or on a phone on the shop floor.

  4. document passages
    4

    Retrieval

    The system selects passages from documents the person asking has access to and passes them to the language model.

  5. 5

    Answer

    The model words the answer from the passages it was given and states the document and page. The employee rates the answer.

Answers rated as wrong go back into improving the index

Operating rules for the assistant and the agent

A source with every answer

The answer includes the document name, revision and page. When the documents contain no answer, the assistant says so and does not word one from general knowledge.

Access rights as in the plant

The assistant uses only documents the person asking has access to. An operator will not get an answer based on cost calculations.

Approval by an employee

The agent prepares a quote, an order or a document. Sending it to the customer and entering it in the ERP is approved by an employee.

Action log

Every question and every step of the agent is recorded: sources, result and the person who approved it.

File search, assistant, agent

File searchDocumentation assistantAgent
What it does Returns a list of files containing the search termAnswers the question and cites the sourceCarries out the successive steps of a task and prepares the result
What it uses File names and contentsThe index of plant documentsDocuments, e-mail, the ERP and spreadsheets
Result A list of files to look throughAn answer with document name and pageA summary, a draft document or an entry for approval
Writing to systems NoneNone — read onlyAfter approval by an employee
What it requires Orderly file namesDocuments in electronic form and agreed access rightsWritten rules, a connection to the systems and an approver
When it makes sense Few documents with known namesMany documents and questions about their contentA repetitive task with a fixed course

Components of the assistant and the agent

An assistant and an agent consist of several components. The language model is one of them and is not the source of knowledge — the knowledge comes from the plant’s documents.

Document retrieval. We combine semantic search with exact search. The first finds a passage that uses different words from the question. The second is needed for part numbers, alarm codes and standard designations, where every character matters.

Language model. It words the answer only from the passages supplied by retrieval and states which document each piece of information comes from. We select a model that runs in the agreed place of data processing and handles Polish as well as the languages of supplier documentation.

Reading drawings and scans. Text in scans is read by character recognition; the title block, the bill of materials table and the annotations on a drawing are read by a model that analyses the image. Values read with low confidence are marked for checking. Reading hand-drawn drawings and low-quality scans is preceded by a trial on documents from your archive.

Agent tools. The agent performs only the operations made available to it: reading the item master, reading an order, creating a draft document. It has no direct access to the ERP database.

Control question set. Before go-live we prepare, with your staff, a set of questions with correct answers. We test the assistant on it after every change of model, index or document scope.

Where the model runs and where the data is

We run the language model on a server at your plant or in a Polish data centre. Documents stay in their current folders and systems, and the index is stored in the same place as the model. The place of data processing is agreed in writing before work starts.

Drawings and customer documentation are not passed to public chat services. For your staff we prepare rules for using AI: which documents may be processed and in which tool.

Connection to e-mail, the archive and the ERP

The agent receives messages from a designated mailbox and reads documents from network folders or from the PDM system. We connect it to the ERP through a programming interface and, where the system offers none, through a scheduled export and import of files. Entries are written to the ERP after approval by an employee.

Passing the bill of materials from CAD to the ERP is described separately: SOLIDWORKS automation.

What is needed to start

  • Documents in electronic form, or consent to scan them
  • A few dozen completed cases of the given kind: enquiries with quotes, orders, breakdown reports
  • The rules by which the task is done today: costing, item coding, document workflow
  • A way to exchange data with the ERP: an interface or a file export
  • A decision on who has access to which documents
  • An employee who rates the answers during the start-up period

AI, machine learning and monitoring

Language models work on text and documents. Numerical forecasts — operation time, scrap risk, power demand — are described separately: machine learning in manufacturing. The quotation agent uses the operation time model if the plant has one.

Machine monitoring detects a deviation in the operation of a component. The maintenance assistant attaches that machine’s repair history and the relevant part of the documentation to the alarm.

Machine documentation and conformity assessment

The agent prepares a draft of the documentation and a list of gaps. It does not replace the risk assessment, the conformity assessment or the signature on the declaration. These are done by the person responsible for the machine — at SOWMAR, the machine safety and CE marking team.

Tools we use ourselves

The 3D printing calculator on this site calculates a price from the geometry of the uploaded model. We run supplier enquiries, quotes and orders in our own system.