Google expands Gemini API
Paul
Author
File Search with multimodal RAG support
Google has expanded the Gemini API File Search tool with three useful upgrades aimed at making retrieval-augmented generation (RAG) systems more practical and easier to trust.
The headline change is multimodal support, which means File Search can now work across both text and images. Instead of only searching written content, developers can build systems that understand visual material as well. In practice, that means an application could search for an image based on its style, tone or subject matter, rather than relying purely on filenames or tags.
Google has also added custom metadata filtering. This lets developers attach structured labels to otherwise unstructured content, such as department, document status or other key-value fields. At query time, those filters can narrow the search space, which should improve both speed and accuracy by cutting out irrelevant material.
The third update is page-level citations. When a model generates an answer from a source such as a large PDF, File Search can now point back to the exact page where the supporting information came from. That is a small but important improvement, because it makes RAG outputs easier to verify and more useful in settings where trust and traceability matter.
Taken together, these changes show Google pushing Gemini’s developer tooling in a more practical direction. The focus here is less on flashy demos and more on making AI systems better at working with messy real-world data while still showing their sources.
Why it matters
- Multimodal File Search means RAG systems can retrieve from images as well as text.
- Metadata filtering helps reduce noise and improves relevance at scale.
- Page-level citations make responses more transparent and easier to fact-check.
- The update moves Gemini’s tooling towards more verifiable, production-friendly AI workflows.