The tool enables developers to utilize their reliable data as the basis for Gemini.
Important details:
– Google is introducing the File Search Tool, a user-friendly RAG system integrated within the Gemini API.
– The File Search Tool can utilize documents and databases to provide Gemini responses with factual context.
– Responses from the Gemini API that utilize the File Search Tool include citations that allow users to verify the model’s output.
Google is incorporating a native retrieval-augmented generation (RAG) system into the Gemini API to facilitate developers in reinforcing their AI tools with substantial data. This feature, known as the File Search Tool, is described by the company as a “straightforward, embedded, and scalable approach to grounding Gemini with your data.” It produces outputs from the Gemini API that are “more precise, pertinent, and verifiable,” according to the announcement blog post.
The File Search Tool utilizes the Gemini Embedding model but relieves users of the substantial effort needed to create a custom RAG system. Instead, the tool manages file storage, chunking, embeddings, and dynamic context injection automatically. Following this, the File Search Tool employs vector search to interpret a provided prompt and determine related information and data from the supplied documents.
Google states that the File Search Tool can reveal pertinent information from submitted documents even if the precise words or phrases are not present. Furthermore, the model’s responses are straightforward to verify, as citations indicate exactly where the File Search Tool and Gemini API derived an answer. Users can subsequently validate the model’s output to confirm accuracy.
The File Search Tool accommodates a broad range of application and text-based file formats, detailed here. It encompasses most major file formats and programming types, including PDF, DOCX, TXT, and JSON. This facilitates the process of transferring your existing database to the Gemini API through the File Search Tool.
Requirements to try the File Search Tool:
A demonstration application is available in the Google AI Studio, allowing developers to see how the File Search Tool can be integrated into their business applications. The tool is paid, yet developers only need to pay a fixed fee of $0.15 per 1 million tokens for the initial embedding and indexing of their files. Following this, storage and embedding generation are free for each individual query.
Google mentions that this payment structure “makes the File Search Tool not only significantly easier but also very economical to develop and expand.” It is now accessible in the Gemini API for those interested in testing it.
This marks the second announcement related to the Gemini API this week, as Google previously revealed support for JSON Schema, simplifying the API’s use in multi-agent workflows.
