Make past quality investigations easier to find: test photo search with batch context

Your quality engineers often already hold the answer to a new defect: a past investigation that found the cause, the batch and the disposition. A defect photograph can take them straight to that record, provided every match arrives with the product, revision, batch and process it belongs to.

Photo search is now practical because open multimodal embedding models place images and text in one search space. Google released EmbeddingGemma 2 on October 6, 2026, and describes text or image queries that find the closest matches in a media library. Pair that retrieval with an ontology of your quality records, and the engineer sees the investigations behind similar pictures, with the context needed to judge them.

An ontology is a shared model of your business objects, their properties, their relationships and the actions people may take on them, as set out in this operational ontology overview. In quality work, those objects are photographs, incidents, products, revisions, batches, processes and dispositions.

What does photo search with batch context give your quality team?

Photo search with batch context removes the assembly work between a new defect and a sound judgement. Three benefits follow from the mechanism:

• Less chasing of records. Each photograph links to its incident, and each incident to its batch, product revision and process. The engineer opens one result instead of cross-checking folders, the ERP and old reports.
• Faster, safer comparison. A lookalike from another revision or process arrives with that difference shown, so the engineer can set it aside quickly instead of reading resemblance as a shared cause.
• More value from records you already keep. Connectors read existing photographs and incident records where they sit. The same definitions of product, batch and incident then serve the next workflow, such as supplier claims or corrective-action follow-up.

FAIS builds that foundational business model and the automation that connects retrieval to it. The photograph becomes evidence attached to an incident; the recorded disposition stays a decision from that investigation rather than an instruction for today's batch. The quality engineer decides what the evidence means.

Why EmbeddingGemma 2 fits a factory archive

EmbeddingGemma 2 is designed for local hardware, which suits defect photos and part records that should stay inside your network. The model card lists 740M parameters under an Apache 2.0 licence: a 270M text model plus a 170M vision encoder and a 300M audio encoder that load only when needed. With quantization, Google reports about 567MB of active RAM for the full multimodal model on a Pixel 11 Pro (announcement).

One input can combine text and images within a shared 8,192-token context, roughly 29 images at the default image budget (model card). An investigation's summary and its photographs can therefore be indexed together and searched with a new photo.

Google's evaluations give a general reference point: 57.28 mean Hit@1 on MMEB v2 Image for the full-precision model at 768 dimensions (model card). Your own surface marks set the real bar, which is why the pilot below measures them directly.

Before and after: a surface mark on a machined housing

Consider a Taiwan manufacturer investigating a surface mark on a machined housing. Today an engineer searches folder names and free-text descriptions, opens photographs one by one, then checks separately which revision, batch and process each old case involved.

With photo search and batch context, the engineer submits the current photograph with its part number and batch. Retrieval proposes past incidents; the connected workflow shows each with its revision, batch, process, inspection conditions and disposition, and flags fields that were never recorded. A close match from an earlier revision is marked as such. The engineer reads the relevant cases and reaches a judgement sooner.

How should you test whether photo search works?

A successful pilot shows your team gathering relevant evidence faster without more misleading matches. Agree criteria with quality reviewers first, then run the current search, photo-only retrieval and photo retrieval with batch context on the same past cases:

• Relevance: reviewers mark useful investigations and misleading lookalikes independently of the ranking.
• Effort: time to assemble a usable evidence set, including opening source records.
• Context errors: wrong-product, wrong-revision and wrong-process matches, plus missing links.
• Robustness: small defects and changes in lighting, magnification and camera angle.

Test the configuration you intend to run. The model card reports near-lossless quality down to 256 dimensions, while 128 dimensions lowers the MIEB lite image score from 64.64 to 59.06. A larger image token budget improves quality at the cost of latency.

Where to start this quarter

Start with one product family, one quality reviewer and one question: do its photographs reliably link to incident and batch records? Where they do, retrieval is the experiment. Where they do not, connecting them is the first job, and that work also serves every later automation touching those products.

That connection between your quality archive and your batch records is the work FAIS does: modelling the products, incidents and decisions your engineers rely on, then connecting retrieval and agents to that model. If past investigations are hard to find in your plant, the useful conversation starts with how your records connect today.