RAG Maintenance and SOP Assistants

What Problem Does It Solve?

When a machine stops, the clock starts. A technician needs the right page of the right manual, and often the fix someone found on a night shift two years ago. That knowledge sits in binders, shared drives, PDF manuals and free-text work order notes.

Search tools rarely help. A file search for a fault code returns every manual that mentions it, or none at all. Experienced technicians know where to look, but many are nearing retirement. Newer staff call for help or guess, and both slow the line down.

As our manufacturing and automotive page notes, much of this knowledge was never written down in one place. A RAG assistant brings together what was written down. It shows where each answer came from, so people can check it before acting.

What Questions Can It Answer?

Which Content Should It Search?

• OEM equipment manualsOperation, troubleshooting and parts sections, indexed by machine model and serial range where the manual varies.
• Standard operating proceduresReleased SOPs from your document control system, with revision level and effective date attached to every passage.
• Work instructions and setup sheetsStation-level steps, changeover checklists and quality checks for specific products.
• Troubleshooting guidesIn-house guides written by engineers and senior technicians, often the most practical source.
• Past work ordersClosed orders from the maintenance system, including fault codes, parts used and technician notes.
• Parts lists and engineering changesBills of materials, spare parts catalogs and change notices that link old and new part numbers.
• Controlled safety proceduresLockout/tagout and other safety documents, shown as approved documents rather than rewritten by the model.

How Does Hybrid RAG Work Here?

Hybrid RAG runs two searches on every question. A keyword index catches exact strings such as fault codes and part numbers. A vector index matches by meaning, so “the arm keeps stopping halfway” can find a passage about intermittent servo faults. The results are merged before the model writes an answer.

1. Released documents are pulled from document control, the maintenance system and approved manual folders. Superseded revisions are removed from the index.
2. Each document is split into chunks that follow its sections, and each chunk keeps its document number, revision, page, machine model and line.
3. An embedding model turns each chunk into an embedding, a list of numbers that captures meaning. A keyword index stores the same chunks for exact matching.
4. A technician asks a question on a tablet or workstation. The system can add context such as the selected line or machine.
5. Both searches run with the same filters, such as site, line, machine model and the user’s access rights. The ranked lists are then merged.
6. A re-ranker, a model that compares the question with each candidate passage, can reorder the top results for precision.
7. The LLM writes a short answer from the best passages, citing document, revision and page. Safety-critical procedures are linked or displayed as approved, not paraphrased.
8. If the sources do not cover the question, the assistant says so and points to the right person, such as a maintenance lead or engineer.

We start plants on this pattern because maintenance questions mix identifiers with everyday language. Meaning-based search alone tends to blur codes like “E-214” and “E-241,” while keyword search alone misses symptoms described in a technician’s own words. Hybrid search covers both with modest added work. For how it compares with the other seven patterns, see RAG architectures explained. For the building blocks, such as chunking and vector search, see our RAG architecture reference.

When Should You Add Other RAG Patterns?

Start with hybrid RAG and an evaluation set, a list of real questions with answers agreed by subject-matter experts. Add a pattern only when failed questions show a gap it fills.

Pattern: Multimodal for Diagrams and Photos

The symptom is answers that stop short, such as “see Figure 12,” when the real answer is in the figure. Wiring diagrams, exploded part views and photos inside manuals carry much of the information technicians need. Text extraction flattens them or skips them. Technicians may also want to photograph a worn part or a control panel display and ask what they are looking at.

Multimodal RAG processes those images so they can be searched. A vision model can describe diagrams, or page images can be embedded directly, and answers cite the page and figure. A vision-capable model then reads the retrieved image alongside the text. The cost is heavier processing and more testing. A misread label on a wiring diagram is a real risk, so the assistant should always show the original figure for the technician to verify.

Pattern: Graph RAG for Incident Themes

The symptom is questions no single document can answer. A safety lead may ask which equipment appears most often in near-miss reports. A reliability engineer may ask whether failures on one pump model trace to one supplier. Those answers depend on links across many reports and records.

Graph RAG builds a knowledge graph, a map of entities and how they relate. Here, it connects near-miss and incident reports, work orders, equipment, lines and suppliers. Safety and reliability teams can then ask about recurring patterns and follow each link back to source records. The costs are real: building the graph takes many LLM calls, extraction errors can create false links and the graph must be refreshed as reports arrive. Incident data is also sensitive, so this view usually belongs to safety and reliability teams, not every tablet on the floor.

Where Do People Stay in Control?

• Qualified people decideThe assistant finds and cites information. Technicians, supervisors and engineers decide what to do with the equipment.
• Safety procedures stay controlledLockout/tagout and other safety-critical steps are shown from the approved document, never summarized or reworded by the model.
• Engineers own the contentDocument owners approve what is indexed, and they fix gaps the assistant reveals.
• Answers can be flaggedTechnicians mark wrong or unclear answers, and a named person reviews each flag.
• Uncertainty is visibleWhen sources are missing or conflict, the assistant says so and points to a person instead of guessing.
• Changes are approvedNew document sources, prompts and patterns go through review before reaching the floor.

What Security and Compliance Controls Matter?

• Document control and revision levelsOnly released revisions are indexed. When a revision is superseded, the old version is removed so it cannot be cited.
• OT and IT separationThe assistant runs on the business network or a designated zone. Data flows respect the segmentation between plant control systems and business systems.
• Intellectual propertyDrawings, process data and pricing stay in your controlled environment, whether on-premises or in your cloud tenant.
• No training of public modelsYour documents and questions are never used to train public models.
• Role-based accessSearch results are filtered by the user’s role, site and line before the model sees any text.
• Audit logsQuestions, retrieved sources, revisions cited and answers are recorded for review.
• Customer quality requirementsThe assistant has to fit within your quality management system, such as ISO 9001 or IATF 16949 for automotive suppliers, and any customer-specific rules.

These controls support your own quality and safety programs; they do not replace them. We do not provide legal or regulatory advice. See AI governance, security and compliance for how we set up controls and oversight.

How Do You Measure Whether It Works?

• Retrieval accuracyHow often the right document, revision and page appear in the top results for evaluation questions.
• Citation accuracyWhether cited sources really support each statement in the answer.
• Exact-match performanceHow reliably fault codes, part numbers and machine models return the correct sources.
• Honest “not covered” responsesHow often the assistant admits a gap instead of inventing an answer.
• Revision correctnessWhether any answer ever cites a superseded document.
• Technician feedbackThumbs-up and flag rates, plus comments from maintenance leads.
• Time to find informationHow long technicians take to reach the right source, compared with their current method.
• Latency and cost per answerTypical and slowest response times, plus model and search costs.

How Do You Roll It Out?

1. 1

   Pick One Area

   Choose one line, machine family or maintenance team with frequent questions and well-maintained documents. Talk to technicians, operators and engineers about what they search for today and where they get stuck.

2. 2

   Clean Up the Sources

   Work with document control to confirm which revisions are released. Gather manuals, SOPs and a sample of past work orders. Remove duplicates and superseded versions before anything is indexed.

3. 3

   Build the Evaluation Set

   Collect real questions from the floor, including fault codes, part numbers and plain-language symptoms. Have subject-matter experts agree on the correct answers and sources, including questions the documents cannot answer.

4. 4

   Pilot on the Floor

   Put the hybrid RAG assistant on a few tablets or workstations. Check it with gloves, noise and weak connectivity in mind. Keep answers short, and log every question, source and flag.

5. 5

   Expand With Evidence

   Review failed questions with engineers and fix content gaps first. Add multimodal or graph RAG only when results show the need. Then extend the assistant to more lines and sites.

Frequently Asked Questions

Get Answers to Your Shop Floor

Tell us which machines cause the most searching and where your manuals, SOPs and work orders live. We will recommend a starting design that fits your quality system and your network. Then we help you build, test and run it. To talk it through, book a free AI consultation.