Google launches File Search for the Gemini API

Google has launched File Search, a Gemini API tool that lets developers connect their applications to their own documents without building the entire search and information retrieval system from scratch.

The tool is designed to help Gemini answer using specific data from a company, project, or collection of files. Instead of relying only on what it learned during training, the model can consult a document database and ground its answers in that information.

What File Search does

File Search integrates the process known as RAG into the Gemini API. RAG is a technique that looks for relevant information in documents before generating a response. Google handles several tasks that developers would normally have to program themselves:

• Store the files.
• Split them into useful chunks.
• Convert the content into numerical representations for meaning-based searches.
• Find information related to each query.
• Add that context to the request Gemini receives.

The search does not rely solely on exact word matches. Using the gemini-embedding-001 embedding model, it can identify that a question and a document excerpt are about the same thing even when they use different terms.

For example, a support bot could search technical manuals and respond with specific instructions. An internal assistant could look up company policies, while a creative application could find ideas or templates within a content library.

Answers that are easier to verify

One of the most relevant features is built-in citations. Responses generated by Gemini indicate which parts of the documents were used, making it possible to check where the information came from instead of accepting the answer without context.

File Search works within the existing generateContent API, so developers can add it without setting up a separate search infrastructure. It also supports common formats such as PDF, DOCX, TXT, and JSON, along with several programming languages.

Google says early users are creating support bots, internal knowledge assistants, and content discovery tools. Phaser Studio, creator of the Beam video game generation platform, says it performs thousands of daily searches across its template library. According to the company, File Search combines parallel queries in under 2 seconds, compared with a manual process that could previously take hours.

How much it costs

The pricing model eliminates two costs during normal use: storage and the generation of embeddings at query time are free.

Developers pay once when indexing files, at a fixed rate of $0.15 per million tokens, based on the applicable cost of the embedding model used. This makes spending more predictable when an application queries the same documents repeatedly.

Google has also published a demo application in Google AI Studio. You need a paid API key to try it.

For you, the important change is that a Gemini-based application can work with proprietary information without its team having to build and maintain storage, semantic search, and excerpt selection separately. What remains to watch is the actual accuracy of citations and answers as document databases grow, because connecting a model to more files does not by itself guarantee that it will always find the right context.