Short answer
Service manuals supply repair procedures, part names, exploded views, torque values, warnings and fault logic in a structured form. They help a model plan and ground language in a repair task. They contain no physical control data, and they are usually owned by the manufacturer, so licensing runs through OEMs and their authorized dealers, not repair shops that merely hold copies.
The buyer's problem
A team building maintenance or repair capability wants the task knowledge that technicians use: which fastener comes out first, what the torque is, which connector releases with a tab, what a fault code means. Service manuals are the primary written source. They are also hard to get in a form that can be used for training, because they come as scanned PDFs, proprietary viewers or structured publishing formats, and the owner is rarely the party that holds the copy.
This page explains what service manuals contain, what they can and cannot teach, how to evaluate their structure, and who can license them. The licensing mechanics are on a separate page, OEM service documentation: access and licensing constraints.
What a service manual contains
- Procedures: removal, installation, adjustment and test steps, usually numbered, with warnings and cautions.
- Specifications: torque values, clearances, fluid types, tool lists, part numbers.
- Illustrations: exploded views, wiring diagrams, location callouts, step figures.
- Diagnostics: symptom-to-cause tables, fault codes, decision trees. See troubleshooting guides and fault trees.
- Safety content: lockout steps, hazard statements, required protective equipment.
- Revision data: effectivity by model or serial range, service bulletins, change history.
What the data can support - and what it cannot
| Use | Supported by manuals? | Notes |
|---|---|---|
| Task decomposition and ordering | Yes | Steps, preconditions, warnings |
| Language grounding for parts and tools | Yes | Names, callouts and figure labels |
| Retrieval for planning or assistance | Yes | Needs structured chunks with figure links |
| Fault reasoning | Partly | Fault trees are idealized |
| Part and tool recognition | Partly | Diagrams differ from real appearance |
| Physical control, grasping, force | No | No motion, force or contact data |
| Real-world exceptions | No | Seized fasteners, damage, improvisation |
Manuals describe the intended method for an intact, correctly identified machine. They say nothing about how a hand holds a tool, how much force breaks a clip, or what to do when a bolt is rusted solid. For those, pair manuals with repair video and instrumented-tool data; see repair video demonstrations. The same distinction between procedures and physical demonstration is set out in what SOPs can and cannot teach a humanoid.
Format and structure
Format determines how much work stands between a licensed file and a usable dataset.
Scanned PDFs need OCR and layout analysis. Layout models are trained on document corpora; DocLayNet, for example, provides 80,863 manually annotated pages with 11 layout classes across six categories, one of which is manuals (Pfitzmann et al., 2022). That is a resource for building or testing a parser, not a guarantee that a given manual will parse cleanly. Multi-column pages, callouts and rotated tables are failure points; see parsing scanned manuals.
Born-digital PDFs and HTML carry text but may lose the link between a figure and its callouts.
Structured publications keep content as modules. S1000D is an XML-based specification in which data modules are stand-alone units of descriptive, procedural or operational data held in a common source database and identified by unique codes, used for civil and military products (ASD S1000D). Where an OEM publishes in this form, procedures arrive already segmented and cross-referenced. See S1000D and ATA documentation.
Parts catalogs and exploded views are often separate documents; see exploded views and parts catalogs.
Who holds service manuals
- OEMs own the content and decide who may license it. They are the primary rights holder.
- Authorized dealers and service networks receive manuals under agreements that usually restrict use to service work.
- Independent repair shops may hold subscription access or legacy copies, with terms that rarely permit reuse for model training.
- Fleet owners and facilities teams hold manuals for equipment they operate, again under the OEM's terms.
- Publishers and aggregators license compiled manuals and may have their own rights and restrictions.
Copyright is central here. The U.S. Copyright Office has published a pre-publication report on the use of copyrighted works in generative AI training that examines where copying occurs, fair use and licensing options (U.S. Copyright Office, 2025). It is background reading, not a conclusion about any specific manual. This is not legal advice; check what the owner's written license actually permits.
Illustrative package: structured manual extract
# Illustrative only - synthetic values, not an available dataset
package: service_docs_v0
scope: 40 procedures for one equipment family (synthetic)
text:
format: JSONL, one procedure per record
fields: procedure_id, title, applicability(model, serial_range),
preconditions, tools, steps[], warnings[], torque_specs[], revision
figures:
format: PNG/SVG with callout map
link: step_id -> figure_id -> callout_label -> part_number
diagnostics: fault_code -> symptoms -> checks -> remedy (where present)
excluded: customer data, serial-level repair histories
rights: training and internal evaluation; OEM approval documented; no redistribution
Evaluating an offer
- Open ten random procedures and confirm that steps, warnings and figures stay linked.
- Check applicability fields. A procedure that omits model or serial range is dangerous to ground a planner on.
- Check the revision date and whether superseded versions are labeled.
- Verify that torque and spec tables survived parsing with units intact.
- Ask what proportion of pages are scanned and how OCR errors were measured.
Questions to ask suppliers
- Who owns the manuals, and can the supplier show written permission to license them for AI training?
- Are the manuals the OEM's own, or a dealer's copy under an agreement that limits use?
- In what format do they exist: scanned, born-digital, or structured modules?
- Which models, serial ranges and revision dates are covered?
- Are figures and callouts separable from the text, and is the link preserved?
- Are service bulletins and superseded versions included?
- Do the documents contain customer names, serial-level histories or pricing?
- Can the supplier provide a sample in the delivery format before a full license?
Discuss your requirements
If you need repair documentation for a specific equipment family, describe the equipment, formats and permitted uses to SourceX. SourceX looks for US businesses that may be able to license suitable material, checks the data and the supplier's licensing permissions, and follows up about potential sources; nothing is contracted until a supplier agrees, and a request does not guarantee a match. Start from the humanoid training data guide or the maintenance and repair data hub, and see the catalog pages on field service maintenance records and internal documentation.
Explore sourcing and licensing options with SourceX
Sources
- DocLayNet: A Large Human-Annotated Dataset for Document-Layout Analysis — layout-analysis resource that includes manuals.
- S1000D International specification for technical publications — structured data-module publishing.
- Copyright and Artificial Intelligence, Part 3: Generative AI Training — background on copyright, fair use and licensing in AI training.