Spare-part identification & aftersales parts search

A broken part on a ten-year-old machine often has no readable number, the manual is outdated and the parts catalogue is a 2,000-page PDF. Multimodal search fuses photo, nameplate OCR, free text, serial number and the machine's bill of materials (BOM) to return the right material number, stock and order link in seconds. Used by maintenance teams (internal parts) and by OEM aftersales (dealers, customers).

Problem & value

• Today: technicians browse exploded drawings, call the parts hotline or ask colleagues. Wrong parts are ordered, machines stand still, duplicate material records pile up in the ERP, and OEM parts desks spend their time on identification calls rather than quotes.
• AI value: self-service identification (fewer hotline calls), faster downtime resolution, more complete quotes (the request arrives with the part already identified), cleaner material master.
• Evidence (vendor-reported, not independently verified): Partium reports that its app DB Mat.ID at Deutsche Bahn supports 95 maintenance plants and 12,200 users and saved 46,800 person-days in one year [1]; DB Fernverkehr states the requested material is in the top 5 results with a probability of over 90% when the technician selects the series and photographs the part [1]. At machine builder ENGEL the search was embedded in the existing parts portal; reported effects are qualitative (fewer hotline calls, more self-served shop transactions, same-day quotes) [2].
• A Fraunhofer IPK / CONTACT Software survey (2024) found spare-part information to be what maintenance staff search for most [3].

Variants

• Internal MRO (maintenance, repair, operations): plant technicians search the own storerooms across sister plants; also detects duplicates and dead stock.
• OEM aftersales portal: customers and dealers identify parts from the installed base (serial number → as-built BOM → parts).
• Rail, agricultural, construction and industrial machinery: long machine lifetimes, many suppliers, superseded parts.
• Automotive/ independent aftermarket: VIN/TecDoc-style catalogue lookup extended by photo search.
• Private/DIY: "what is this screw/gasket for my dishwasher" via a retailer app (combine with KNOW-005).
• Not covered: similarity of 3D geometry in CAD (KNOW-020); manual Q&A (KNOW-004).

How it works

1. query (photo | nameplate scan | barcode/QR | text/voice | serial no.)
2. input understanding (OCR of plates/labels, VLM description of the part, language normalisation)
3. context narrowing (machine/series/serial
4. as-built BOM and assembly scope)
5. hybrid retrieval over parts master (image embeddings + text embeddings + keyword/part-no. exact match)
6. rerank / VLM verification against part images, dimensions and attributes
7. enrich (ERP: stock, location, price, supersession; supplier cross-reference)
8. answer (top-N with confidence)
9. order / requisition | low confidence
10. parts-desk hand-over

• Key components: multimodal embeddings (SigLIP 2-class image-text encoders [4], page-image retrieval à la ColPali for exploded drawings [5]), OCR for plates, structured filters from the BOM, a reranker or VLM for final verification, and a connector to ERP/CMMS.
• Hard part is data, not model: parts masters are messy (duplicates, free-text descriptions in several languages, few images). Photographing parts and structuring catalogues is the main onboarding work; vendors such as Partium offer an image-collection workflow where an agent extracts part data from photos [6].
• Human-in-the-loop: the user confirms the match against the physical part; unresolved searches go to the parts desk with the search history visible (the ENGEL pattern [2]); safety-relevant parts require confirmation by part number before ordering.

Data preparation

Data needed: material master, stock and prices (ERP/CMMS), BOMs and exploded-view drawings, parts catalogues as PDF, serial-number and installed-base records, photos of past tickets. In an SMB this typically lives in the ERP, in PDM/CAD folders, on file shares and in old catalogue PDFs, partly even as paper binders for legacy machines.

Typical gaps: scanned catalogues without text layer, drawings with part numbers only as image, several numbering schemes after superseded parts, duplicate and obsolete material records, descriptions in mixed languages, no photos per part, missing link between serial number and BOM version.

Preparation steps: export the material master and clean duplicates and supersession chains (see [PREP-005]); OCR and split catalogues into part-level chunks ([PREP-006], [DOC-014]); add tags such as machine family, assembly and synonyms from the workshop vocabulary ([PREP-010], [PREP-015]); collect 100 to 200 real photo-plus-answer pairs from the parts desk as a test set ([PREP-018]). Tools: Paperless-ngx or an OCR pipeline, Python/pandas, a vector database.

Expect roughly 50 to 60 percent of the total project effort to go into data preparation, because retrieval quality depends almost entirely on clean part masters and BOM links.

Implementation paths

Which path for whom

• smb: B - the search must tap the own ERP material master, BOMs and catalogue PDFs, and no off-the-shelf product knows these; a cloud-API build on EU regions with a thin shop or ERP integration fits.
• enterprise: B - parts data, serial-number records and shop integration are core aftersales assets, so a build on cloud APIs keeps control of the data model; a commercial parts-catalogue product is an option for standard catalogues.
• developer: C - open vision-language and embedding models with a vector database allow a reusable parts-search service that runs on HW2 hardware.

Path A — Buy

• Specialised parts-search products: Partium (Find, Cockpit; offices in Vienna, Barcelona, US HQ; customers include Deutsche Bahn, ÖBB, ENGEL, Liebherr [1][2][6][7]), integrated into an existing parts portal or app.
• Adjacent platforms: aftersales/service-parts suites and PLM/SLM vendors (e.g. PTC Servigistics, Siemens Teamcenter, SAP Service Parts Management) and technical-documentation/catalogue tools (e.g. Schema ST4, Quanos) that increasingly add semantic or visual search; verify current feature status in demos.
• Check: hosting region and ISO 27001 evidence (Partium's site links a trust centre but does not state the hosting location on the product page [6]), SAP/CMMS connector, multi-language, offline use on the shop floor, ownership of the trained image data.
• Choose A when: you want results within months and have clean enough master data.

Path B — Build on cloud APIs

• Models: multimodal LLMs with EU processing (Claude on EU regions of cloud providers, Gemini and Azure OpenAI in EU regions, Mistral as EU provider) to read plates and describe parts; embedding APIs for text and images (e.g. Cohere Embed, Jina embeddings, Voyage multimodal); reranker API.
• Stack: vector/hybrid index (OpenSearch, Elasticsearch, Qdrant, Weaviate, pgvector), mobile PWA, ERP connector (SAP OData/BAPI, Dynamics), orchestration with n8n or custom service.
• Pattern: VLM extracts attributes (type, dimensions, thread, material, visible markings) → structured filter + hybrid retrieval → VLM or reranker verification. Always return the real part image for visual confirmation.
• Choose B when: you have in-house developers and need custom logic or a branded customer portal.

Path C — Self-host (open source)

• Models: Qwen3-VL (2B–32B dense, 30B-A3B MoE, Apache-2.0) [8] for plate reading and part description; SigLIP 2 encoders (open checkpoints, multilingual) for image embeddings [4]; ColPali-family models for drawing pages [5]; bge-m3 or similar for text; a bge reranker.
• Parsing: Docling (MIT, air-gap capable) [9] or PaddleOCR for catalogue PDFs and plates.
• Stack: Qdrant/pgvector/OpenSearch, vLLM or Ollama, FastAPI; works offline in the plant network.
• Hardware: HW2 (one 24–32 GB GPU) serves a ~8B VLM plus embeddings for a mid-size catalogue; the embedding index for 100k–1M parts fits on CPU/disk.
• Ops effort: moderate; the main effort is continuous data curation.

Path D — Custom training

Worthwhile when generic encoders confuse near-identical parts (bearings, seals, fasteners, connectors). Fine-tune an image-text encoder with contrastive learning on your own part photos (several images per part, different angles and wear states) plus user-confirmed search logs as hard negatives. Needs hundreds to thousands of labelled parts; production feedback (accepted/rejected matches) is the natural training stream.

State of the art (2026)

• Multimodal retrieval with strong open image-text encoders and VLMs has made photo and plate-based search practical without per-product training; page-image retrievers avoid fragile PDF parsing of exploded views [4][5].
• Agentic search: an agent asks clarifying questions ("which machine series? is there a thread?"), queries BOM, stock and supplier data, and drafts the requisition; vendors position agents beneath their search products [6].
• Typical failure modes: visually near-identical parts that differ in dimensions or material; worn, dirty or partial parts; photos without scale reference; outdated or wrong BOMs; superseded numbers; hallucinated part numbers if a generative model answers without retrieval. Mitigation: the model must only return numbers that exist in the master data, show images and key dimensions, and ask for the nameplate where possible.
• Benchmarks: no public benchmark for industrial spare-part identification; measure top-1/top-5 accuracy on your own labelled query set. Published accuracy figures are vendor claims.
• Practical advice from the field: start from the nameplate, confirm by manufacturer part number and dimensions, search by function, check sister-plant storerooms and plan for obsolescence [3].

Shared resources & combinations

• Same building blocks (vlm + ocr-docparse + embedding-text + vector-db + reranker): KNOW-005 (e-commerce product search), VIS-022 (visual search), VIS-018 (industrial OCR for serials/labels), KNOW-020 (CAD similar-part search, shares the catalogue and embedding index), DOC-014 (document parsing for RAG), KNOW-001 (enterprise search). One HW2 server can host VLM, embeddings and reranker for all.
• Combine with: KNOW-004 (manual Q&A: "which part, then how do I replace it"), CONV-027 (live-camera assistance for field technicians), CONV-001 (customer service chatbot with part lookup), KNOW-013 (knowledge graph linking machine, assembly, part, supplier), PRED-006 (predictive maintenance triggers parts reservation), OPT-005 (inventory optimisation uses duplicate/dead-stock findings), OPT-019 (field-service dispatch with the right parts on the van).
• Data pipelines shared: material master cleansing, BOM/PLM extraction, image library of parts.

Effort, cost & KPIs

• Path A: weeks for a pilot with one product line; subscription pricing by users or parts (not publicly listed by most vendors); data onboarding is the main cost.
• Path B: 2–4 months incl. ERP integration and image collection.
• Path C: 3–6 months; one GPU server plus data engineer capacity.
• Path D: add 1–2 months for labelling and evaluation.
• KPIs: top-1 / top-5 identification accuracy, self-service rate (searches without hotline), time to identify, share of orders with wrong part (returns), hotline call volume, quote conversion, duplicate material records removed, machine downtime hours.

People & change

Involve management for budget and scope, the parts desk or aftersales head as process owner, and the ERP and master-data owners from day one, because their data defines quality ([GUIDE-003]). IT is needed for ERP and shop integration; the DPO only if customer or technician data is stored with the requests. Pilot with one product line and a small group of technicians and parts-desk staff.

Works council: usually not required for a pure parts search without user-level evaluation (no §87(1) no. 6 BetrVG trigger). It becomes relevant if usage is logged per technician or measured. Even then, inform the council early under §90(1) no. 3 BetrVG when work processes change, if one exists.

AI literacy: under AI Act Art. 4 employers must take measures to support staff AI literacy; a short hands-on session on photo quality, confidence scores and when to escalate is enough ([GUIDE-006]).

Typical concerns: "the AI orders the wrong part" and "my expertise becomes worthless". Address them with a confidence threshold, mandatory human confirmation for expensive or safety-relevant parts, and by positioning the experienced parts specialist as the owner of the knowledge base and of hard cases ([GUIDE-008]).

Risks, regulation & compliance

• EU AI Act: minimal risk — product search is not an Annex III use; AI literacy (Art. 4) applies. An embedded customer chatbot must disclose AI interaction (Art. 50). Watch for the case where the AI is a safety component of a regulated product; part identification itself normally is not. The AI Omnibus moved Annex III high-risk obligations to 2 Dec 2027 and Annex I to 2 Aug 2028 [10].
• GDPR: photos may capture workers or customer sites; crop or blur faces, restrict image retention; user and search logs are personal data (works-council agreement in Germany).
• Product safety/liability: a wrongly identified safety-relevant part (brakes, pressure, lifting equipment, rail) is a safety risk — require confirmation by part number and OEM-approved lists; keep logs of the suggestion versus the ordered part.
• IP/data: CAD drawings and BOMs are trade secrets — check training-data and data-residency terms of vendors; Partium's product page does not state hosting [6].
• Mitigations: retrieval-only answers, confidence thresholds, human hand-over, regular accuracy audits.

Getting started

1. Pick one product line and collect 200–500 real "unknown part" queries (photos, texts) with correct answers.
2. Audit the parts master: duplicates, missing images, missing BOM links; fix the top gaps.
3. Run 2–3 Path A vendors or a Path B prototype (VLM + hybrid index) against the test set; compare top-5 accuracy.
4. Integrate with ERP stock/price and add the parts-desk hand-over; roll out to a pilot group of technicians.
5. Log accepted/rejected matches and feed them back into synonyms, images and ranking.

Sources

1. Partium — Deutsche Bahn case (DB Mat.ID, vendor-reported figures) — https://www.partium.io/maintenance/deutsche-bahn-fernverkehr
2. Partium — ENGEL aftersales case — https://www.partium.io/aftersales/engel
3. Partium blog — 9 ways to find the right spare part faster (cites Fraunhofer IPK / CONTACT survey 2024) — https://www.partium.io/blog/find-the-right-spare-part
4. Tschannen et al. — SigLIP 2 — https://arxiv.org/abs/2502.14786
5. Faysse et al. — ColPali: Efficient Document Retrieval with Vision Language Models — https://arxiv.org/abs/2407.01449
6. Partium Find product page — https://www.partium.io/parts-intelligence/find
7. Partium company page — https://www.partium.io/company
8. Qwen3-VL repository (sizes, Apache-2.0) — https://github.com/QwenLM/Qwen3-VL
9. Docling project — https://github.com/docling-project/docling
10. European Commission — AI Act regulatory framework and timeline (incl. AI Omnibus) — https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai