Why Google’s File Search could displace DIY RAG stacks in the enterprise
Companies now understand that Retrieval Augmented Generation (RAG) enables applications and agents to find the best, most informed information for queries. However, typical RAG setups can pose a technical challenge also exhibit undesirable properties.
To help resolve this, Googling has released the File Search Tool on the Gemini API, a fully managed RAG system “that abstracts away the collection pipeline.” File Search removes much of the collection of tools and applications involved in setting up RAG pipelines, so engineers don’t have to cobble together things like storage solutions and creator embedding.
This tool competes directly with RAG business products Open AI, AWS And Microsoftwhich also aim to simplify the RAG architecture. However, Google claims its offering requires less orchestration and is more self-contained.
“File Search provides a simple, integrated, and scalable way to base Gemini on your data, delivering answers that are more accurate, relevant, and verifiable,” Google said in a blog post.
Businesses can access certain File Search features for free while searching, such as storage generation and file embedding. Users will start paying for embeds when these files are indexed at a flat rate of $0.15 per 1 million tokens.
Google’s Gemini Embedding model, which ultimately… top embedding model on the Massive Text Embedding Benchmark, enables file searching.
File search and integrated experiences
Google said File Search works “by handling the complexity of RAG for you.”
File Search manages file storage, chunking strategies and embeddings. Developers can call File Search within the existing generatorContent API, which Google says makes the tool easier to use.
File Search uses vector searches to “understand the meaning and context of a user’s search query.” Ideally, it will find the relevant information to answer a question from documents, even if the prompt contains incorrect words.
The feature has built-in citations that point to the specific parts of a document used to generate answers, and also supports various file formats. These include PDF, Docx, txt, JSON and “many common file types in programming languages,” says Google.
Continuous RAG experiments
Companies may have already started building out a RAG pipeline as they lay the foundation for their AI agents to actually use the right data and make informed decisions.
Because RAG represents an important part of how companies maintain accuracy and leverage insights about their business, organizations need to quickly gain insight into this pipeline. RAG can be technically challenging because orchestrating multiple tools together can become complicated.
Building ‘traditional’ RAG pipelines means organizations must build and refine a file ingestion and parsing program, including chunking, embed generation, and updates. They then need to contract a vector database, such as Pine conedetermine its retrieval logic and place it all within the context window of a model. In addition, they can add source references if desired.
File Search aims to streamline all that, although competing platforms offer similar features. OpenAIs Assistants API allows developers to use a file search function, which directs an agent to relevant documents for answers. Bedrock from AWS unveiled a managed data automation service in December.
While File Search works similarly to these other platforms, Google’s offering abstracts all, rather than just some, elements of the RAG pipeline creation.
Phaser Studio, the maker of the AI-powered game generation platform Beam, said in Google’s blog that it was using File Search to search its library of 3,000 files.
“With File Search, we can instantly surface the right material, whether that’s a bulleted code snippet, genre templates, or architectural guidelines from our Phaser brain corpus,” said Phaser CTO Richard Davey. “The result is that ideas that once took days to prototype are now playable in minutes.”
Since the announcement, several users have expressed interest in using the feature.
