Claude Desktop Doesn't Use RAG — The New Norm: "File Search Is Enough"

Introduction: Is RAG Really Necessary?

Throughout 2023 and 2024, RAG (Retrieval-Augmented Generation) was touted as an “essential technology” for AI implementation. Vectorizing internal documents and passing relevant chunks to LLMs via similarity search became the standard pattern for enterprise AI.

However, in 2025, a shockwave ran through the AI industry. Boris Cherny, the development lead for Anthropic’s Claude Code, revealed:

Early versions of Claude Code used RAG and a local vector DB. But we quickly found that Agentic Search was vastly superior.

Claude Desktop (Cowork) inherited this philosophy. No vector DBs, no embeddings. It only uses grep (text search) and glob (filename search)—the classic “file search” tools. Anthropic’s conclusion is that this is enough.

In this article, I will explain why Claude Desktop chose this approach and what the “No RAG Needed” theory means for PMs and business leaders.

What is RAG? (A 30-Second Refresher)

Roughly speaking, RAG is a mechanism where “AI automatically searches for and references relevant internal materials before answering.” Technically, it works in these steps:

Preparation: Split internal documents into chunks, convert them into numbers (vectors) that AI can understand, and store them in a specialized database.
Search: Convert the user’s question into a vector and find “semantically similar” documents.
Generation: Pass the found documents to the LLM to generate an answer.

While it seems rational, implementation and operation come with many challenges. Designing chunking strategies, building and managing vector DBs, tuning search precision, managing data freshness—in reality, it takes more effort and cost than imagined.

Claude Desktop’s Choice: “Let the AI Search Itself”

Claude Desktop (Cowork) adopts a completely different approach called Agentic Search.

The idea is simple: “Give the AI search tools and let it find the files itself.” Specifically, Claude Desktop finds information using only basic tools:

Grep: Search text content within files (based on ripgrep)
Glob: Search by file name patterns (e.g., *.pdf, src/**/*.ts)
Read: Read the found files

This mirrors exactly how an engineer looks for code in a terminal. No special infrastructure or databases are required. Leveraging the LLM’s “ability to understand the structure of code and documents,” it reaches the necessary information by repeating searches multiple times.

Why is it Superior to RAG? — 5 Reasons

1. No Infrastructure Required

Implementing RAG requires building and maintaining multiple systems like vector DBs, embedding models, and indexing pipelines. Agentic Search works as long as you have a file system and text search tools. The hurdle for adoption is overwhelmingly lower.

2. Always Reference Fresh Data

RAG requires rebuilding the vector index whenever documents are updated. Neglecting this leads to the serious problem of “answering based on old information.” Agentic Search always reads the “current content” of files directly, so the freshness issue structurally does not occur.

3. Security and Privacy

RAG requires storing the vector index somewhere; placing it in the cloud creates data leakage risks. With Agentic Search, data never leaves the machine. It searches directly on the user’s file system, protecting privacy. Boris Cherny himself cited “data not leaving the machine” as a major benefit.

4. Transparent Search Process

RAG search is a black box of vector similarity calculations. It’s hard for humans to verify “why this document was chosen.” On the other hand, Agentic Search visualizes everything: “which keywords were searched” and “which files were read.” This makes auditing and debugging easy, fulfilling the accountability required in business settings.

5. Gets Smarter as Models Evolve

This might be the most important point. Boris Cherny stated, “Ultimately, it’s all about the model. If the model gets better, it subsumes everything else.” The RAG pipeline doesn’t fundamentally change even if the model evolves. However, with Agentic Search, as the LLM’s reasoning capability improves, its search strategy automatically improves. It’s a design where investment doesn’t easily become obsolete.

A Fair View: Knowing the Weaknesses of Agentic Search

Of course, it’s not a silver bullet.

Token costs tend to be higher than RAG. RAG retrieves relevant documents in a single search, while Agentic Search tends to increase API usage fees because it searches and reads files multiple times. Some estimates suggest that while RAG costs about $0.00008 per query, the long-context approach can cost about $0.10—a significant difference.

Also, it struggles with conceptual searches. For abstract questions like “Where is the authentication process handled?”, keyword searches with grep might struggle to reach the target code or document.

However, these weaknesses are being mitigated with model evolution. Research by Cognition has shown that using parallel tool calling can improve the speed of Agentic Search by 4x.

Business Implications: Time to Rethink AI Strategy

This “No RAG Needed” theory holds significant meaning for PMs and business leaders considering AI adoption.

Before investing heavily in building RAG, you should verify if Agentic Search is sufficient. Like Claude Desktop, there’s a possibility that existing file systems and text search alone can handle a significant amount of work.

RAG might be an over-investment, especially in cases like:

The volume of internal documents is not massive.
Target documents are updated frequently.
Security requirements are strict.
You want to start small and validate quickly.

On the other hand, RAG remains a valid option when you need to instantly pull semantically relevant items from hundreds of thousands of documents, or when cost optimization is the top priority for large-scale operations.

Conclusion: The Era Where Simplicity Wins

Claude Desktop’s design philosophy is clear: “Instead of building complex pipelines, give simple tools to a smart model.”

As of 2026, LLM capabilities are improving day by day. Search precision, depth of reasoning, and context understanding. The approach that benefits most from this evolution might not be RAG, which relies on complex infrastructure, but Agentic Search, which leverages the model’s power directly.

Of course, I’m not saying “all RAG is unnecessary.” But what Claude Desktop has proven is the fact that “you can do this much without RAG.” As a first step in AI adoption, why not start simple? It might be time to reconsider that value.