Gemini 2.5 runnin on OCI in Frankfurt
Earlier this week (January 21), Oracle officially enabled Google Gemini 2.5 endpoints in the Frankfurt (Germany Central) region. See the announcement (https://docs.oracle.com/en-us/iaas/releasenotes/generative-ai/gemini-2-5-regions.htm) and the documentation over there: https://docs.oracle.com/en-us/iaas/Content/generative-ai/google-gemini-2-5-pro.htm
In addition, last week additional document processing functionalities have been released around Gemini on OCI: https://docs.oracle.com/en-us/iaas/releasenotes/generative-ai/gemini-2-5-document-understanding.htm
While the press release focused on the strategic partnership, for us as engineers and integrators, the news is far more specific: we finally have a compliant, native multimodal model in the EU that can handle operational data.
At Broadpin, we have already started testing the Frankfurt endpoints. The latency is minimal, but the real breakthrough is in Zero-Shot Document Processing. Here is why this release changes the architecture for inbound data automation in Oracle systems.
The End of Brittle OCR Templates?
Traditionally, automating inbound documents (invoices, delivery notes, maintenance logs) into Oracle ERP or SCM required a rigid OCR stack. You needed to define zones, train templates for every vendor layout, and handle exceptions manually. If a scan was skewed or coffee-stained, the process failed.
Gemini 2.5 is multimodal by design. It does not "read" text character-by-character; it processes the visual image of the document entirely. It understands layout, hierarchy, and context without any prior training.
Use Case: Processing "Real World" Inbound Docs
Since the Frankfurt region opened on Tuesday, we have run stress tests against the model using difficult, real-world samples common in logistics and manufacturing. The results were stark:
- Handwritten Overrides: The model successfully extracted data from printed invoices where a warehouse worker had crossed out a quantity and hand-wrote a new number (e.g., "received 4, not 10") in the margin. Standard OCR tools consistently miss this context.
- Poor Quality Scans: We fed it mobile phone photos of crumpled delivery notes taken in low light. Gemini 2.5 correctly inferred the PO_NUMBER and LINE_ITEMS despite the geometric distortion and shadowing.
- Mixed Media: It handled PDFs containing both typed text and pasted screenshots (e.g., of an email thread) without needing separate processing pipelines.
Speed to Value: Hours, Not Weeks
The most immediate impact for our clients is implementation speed.
Because Gemini 2.5 requires no model training for these tasks, the development cycle shifts from Data Science to Integration Engineering.
- Tuesday Morning: Region availability announced.
- Tuesday Afternoon: We connected an OCI Object Storage bucket to the Gemini 2.5 Frankfurt endpoint.
- Wednesday: We had a working prototype parsing unstructured PDFs into JSON payload for Oracle Integration Cloud (OIC).
There is no need to label thousands of images. You write a robust prompt, pass the image byte array, and get structured JSON back.
The Privacy & Latency Factor
Before this week, achieving this level of reasoning required sending data to US-hosted models – a non-starter for many of our EU clients. Especially due to the limitations of the latest Meta Llama models within Europe (see e.g. https://www.theverge.com/2024/7/18/24201041/meta-multimodal-llama-ai-model-launch-eu-regulations)
With the model now hosted in the European Union, the data path is strictly local.
Description on Gemini availability in Frankfurt
- GDPR Alignment: Processing occurs within the EU.
- Performance: We are seeing inference times that support near-real-time user feedback (e.g., an AP clerk dragging a file into a portal and seeing the extracted fields instantly).
While Gemini was available for quite some time through a direct contract with Google that was a an additional burden for many of our customers. That it is now possible to leverage the service directly through the Oracle OCI subscription the formal requirements to use the models got a lot easier to achieve.
What This Means for Your Stack
If you are running Oracle Fusion, NetSuite, E-Business Suite or a custom OCI app, you likely have a backlog of manual data entry tasks labeled "too complex to automate."
The barrier to automating those tasks just dropped significantly. We are now advising our clients to review their "discard pile" – the documents previously deemed too messy for automation – and run them through a proof-of-concept with this new endpoint.
I'm looking forward to the "soon to happen" release of Gemini 3 as well. Furthermore stay tuned for another blog post in the making that shows how this service can be used in #cline to speed up your development a lot.