SWE-grep and SWE-grep-mini: Fast, Agentic Models for Context Retrieval in Coding Agents

Introduction to SWE-grep and SWE-grep-mini

Modern coding agents often face a tradeoff between speed and intelligence. While advanced models can handle complex tasks, they can take minutes to retrieve the necessary context, disrupting your workflow. To address this, we’ve introduced SWE-grep and SWE-grep-mini, two fast agentic models that match the retrieval capabilities of state-of-the-art coding models but operate an order of magnitude faster. These models are now available in Windsurf’s new Fast Context subagent.

The Speed-Intelligence Tradeoff

In tools like Windsurf and Devin, we observed that agents often spend over 60% of their initial turn just retrieving context. This is a significant bottleneck. Historically, context retrieval has been approached in two main ways:

• Embedding Search (RAG): Once the codebase is indexed, queries are fast. However:

  • Results can be inaccurate for complex queries that require jumping across multiple parts of the codebase.
  • Embeddings can sometimes lead to the agent giving too much weight to irrelevant information.

• Agentic Search: The model uses CLI tools to explore the codebase, similar to a human. While flexible, it has drawbacks:

  • It’s generally slow, requiring dozens of sequential roundtrips between the user’s device and the inference endpoint.
  • The agent must sift through tens of thousands of tokens of irrelevant information, which can degrade answer quality.

Introducing SWE-grep and SWE-grep-mini

To overcome these challenges, we trained SWE-grep and SWE-grep-mini, models that combine the flexibility of agentic search with the speed of embedding search. Here’s what makes them stand out:

• Highly Parallel Context Retrieval: Both models are designed to retrieve context in parallel, significantly reducing the time required.
• Match Frontier Model Capabilities: Despite their speed, they maintain the retrieval accuracy and depth of more complex models.
• Order of Magnitude Faster: SWE-grep and SWE-grep-mini can complete tasks up to 10 times faster than traditional agentic search methods.

Implementation Details

Architecture

• SWE-grep is a larger model, optimized for handling extensive codebases with high accuracy.
• SWE-grep-mini is a smaller, more lightweight version suitable for smaller projects or environments where resource constraints are a concern.

Both models use reinforcement learning (RL) to improve their context retrieval capabilities. They learn from user interactions and feedback, continuously refining their search strategies.

Benchmarks

• Speed: In internal benchmarks, SWE-grep-mini completed context retrieval tasks in under 10 seconds for most queries, while SWE-grep achieved similar results with slightly higher accuracy.
• Accuracy: Both models maintained a high level of accuracy, matching or exceeding the performance of traditional agentic search methods.

Using Fast Context in Windsurf

We are rolling out the Fast Context subagent to Windsurf users progressively. Here’s how you can try it:

1. Update to the Latest Release: Ensure you have the latest version of Windsurf installed.
2. Use Windsurf Cascade Normally: When you make a query that requires code search, Fast Context will automatically trigger. You can also force it to trigger by submitting your query with Cmd+Enter.

Benefits of Fast Context

• Improved Workflow: By reducing the time spent on context retrieval, Fast Context helps you stay in flow and maintain productivity.
• Better Accuracy: Despite its speed, Fast Context maintains high accuracy, ensuring you get the right information quickly.

Try It Out

You can try Fast Context in our playground at https://playground.cognition.ai/ to see how it reduces the time it takes to understand large codebases.