The post GitHub Copilot Enhances Code Search with New Embedding Model appeared on BitcoinEthereumNews.com. Ted Hisokawa Sep 26, 2025 03:41 GitHub introduces a new Copilot embedding model, enhancing code search in VS Code with improved accuracy and efficiency, according to GitHub’s announcement. GitHub has announced a significant upgrade to its Copilot tool, introducing a new embedding model that promises to enhance code search within Visual Studio Code (VS Code). This development aims to make code retrieval faster, more memory-efficient, and significantly more accurate, as detailed in a recent GitHub blog post. Enhanced Code Retrieval The new Copilot embedding model brings a 37.6% improvement in retrieval quality, doubling the throughput and reducing the index size by eight times. This means developers can expect more accurate code suggestions, faster response times, and reduced memory usage in VS Code. The model effectively provides the correct code snippets needed, minimizing irrelevant results. Why the Upgrade Matters Efficient code search is crucial for a seamless AI coding experience. Embeddings, which are vector representations, play a key role in retrieving semantically relevant code and natural language content. The improved embeddings result in higher retrieval quality, thereby enhancing the overall GitHub Copilot experience. Technical Improvements GitHub has trained and deployed this new model specifically for code and documentation, enhancing context retrieval for various Copilot modes. The update has shown significant improvements, with C# developers experiencing a 110.7% increase in code acceptance ratios and Java developers seeing a 113.1% rise. Training and Evaluation The model was optimized using contrastive learning techniques, such as InfoNCE loss and Matryoshka Representation Learning, to improve retrieval quality. A key aspect of the training involved using ‘hard negatives’—code examples that appear correct but are not—helping the model distinguish between nearly correct and actually correct code snippets. Future Prospects GitHub plans to expand its training and evaluation data to include… The post GitHub Copilot Enhances Code Search with New Embedding Model appeared on BitcoinEthereumNews.com. Ted Hisokawa Sep 26, 2025 03:41 GitHub introduces a new Copilot embedding model, enhancing code search in VS Code with improved accuracy and efficiency, according to GitHub’s announcement. GitHub has announced a significant upgrade to its Copilot tool, introducing a new embedding model that promises to enhance code search within Visual Studio Code (VS Code). This development aims to make code retrieval faster, more memory-efficient, and significantly more accurate, as detailed in a recent GitHub blog post. Enhanced Code Retrieval The new Copilot embedding model brings a 37.6% improvement in retrieval quality, doubling the throughput and reducing the index size by eight times. This means developers can expect more accurate code suggestions, faster response times, and reduced memory usage in VS Code. The model effectively provides the correct code snippets needed, minimizing irrelevant results. Why the Upgrade Matters Efficient code search is crucial for a seamless AI coding experience. Embeddings, which are vector representations, play a key role in retrieving semantically relevant code and natural language content. The improved embeddings result in higher retrieval quality, thereby enhancing the overall GitHub Copilot experience. Technical Improvements GitHub has trained and deployed this new model specifically for code and documentation, enhancing context retrieval for various Copilot modes. The update has shown significant improvements, with C# developers experiencing a 110.7% increase in code acceptance ratios and Java developers seeing a 113.1% rise. Training and Evaluation The model was optimized using contrastive learning techniques, such as InfoNCE loss and Matryoshka Representation Learning, to improve retrieval quality. A key aspect of the training involved using ‘hard negatives’—code examples that appear correct but are not—helping the model distinguish between nearly correct and actually correct code snippets. Future Prospects GitHub plans to expand its training and evaluation data to include…

GitHub Copilot Enhances Code Search with New Embedding Model

2025/09/27 19:29


Ted Hisokawa
Sep 26, 2025 03:41

GitHub introduces a new Copilot embedding model, enhancing code search in VS Code with improved accuracy and efficiency, according to GitHub’s announcement.





GitHub has announced a significant upgrade to its Copilot tool, introducing a new embedding model that promises to enhance code search within Visual Studio Code (VS Code). This development aims to make code retrieval faster, more memory-efficient, and significantly more accurate, as detailed in a recent GitHub blog post.

Enhanced Code Retrieval

The new Copilot embedding model brings a 37.6% improvement in retrieval quality, doubling the throughput and reducing the index size by eight times. This means developers can expect more accurate code suggestions, faster response times, and reduced memory usage in VS Code. The model effectively provides the correct code snippets needed, minimizing irrelevant results.

Why the Upgrade Matters

Efficient code search is crucial for a seamless AI coding experience. Embeddings, which are vector representations, play a key role in retrieving semantically relevant code and natural language content. The improved embeddings result in higher retrieval quality, thereby enhancing the overall GitHub Copilot experience.

Technical Improvements

GitHub has trained and deployed this new model specifically for code and documentation, enhancing context retrieval for various Copilot modes. The update has shown significant improvements, with C# developers experiencing a 110.7% increase in code acceptance ratios and Java developers seeing a 113.1% rise.

Training and Evaluation

The model was optimized using contrastive learning techniques, such as InfoNCE loss and Matryoshka Representation Learning, to improve retrieval quality. A key aspect of the training involved using ‘hard negatives’—code examples that appear correct but are not—helping the model distinguish between nearly correct and actually correct code snippets.

Future Prospects

GitHub plans to expand its training and evaluation data to include more languages and repositories. The company is also refining its hard negative mining pipeline to enhance quality further, with goals to deploy larger, more accurate models leveraging the efficiency gains from this update.

This latest enhancement is a step towards making AI coding assistants more reliable and efficient for developers, promising a smarter and more dependable tool for everyday development.

Image source: Shutterstock


Source: https://blockchain.news/news/github-copilot-enhances-code-search-with-new-embedding-model

Disclaimer: The articles reposted on this site are sourced from public platforms and are provided for informational purposes only. They do not necessarily reflect the views of MEXC. All rights remain with the original authors. If you believe any content infringes on third-party rights, please contact service@support.mexc.com for removal. MEXC makes no guarantees regarding the accuracy, completeness, or timeliness of the content and is not responsible for any actions taken based on the information provided. The content does not constitute financial, legal, or other professional advice, nor should it be considered a recommendation or endorsement by MEXC.

You May Also Like

Volante Technologies Customers Successfully Navigate Critical Regulatory Deadlines for EU SEPA Instant and Global SWIFT Cross-Border Payments

Volante Technologies Customers Successfully Navigate Critical Regulatory Deadlines for EU SEPA Instant and Global SWIFT Cross-Border Payments

PaaS leader ensures seamless migrations and uninterrupted payment operations LONDON–(BUSINESS WIRE)–Volante Technologies, the global leader in Payments as a Service
Share
AI Journal2025/12/16 17:16
Fed Acts on Economic Signals with Rate Cut

Fed Acts on Economic Signals with Rate Cut

In a significant pivot, the Federal Reserve reduced its benchmark interest rate following a prolonged ten-month hiatus. This decision, reflecting a strategic response to the current economic climate, has captured attention across financial sectors, with both market participants and policymakers keenly evaluating its potential impact.Continue Reading:Fed Acts on Economic Signals with Rate Cut
Share
Coinstats2025/09/18 02:28
Google's AP2 protocol has been released. Does encrypted AI still have a chance?

Google's AP2 protocol has been released. Does encrypted AI still have a chance?

Following the MCP and A2A protocols, the AI Agent market has seen another blockbuster arrival: the Agent Payments Protocol (AP2), developed by Google. This will clearly further enhance AI Agents' autonomous multi-tasking capabilities, but the unfortunate reality is that it has little to do with web3AI. Let's take a closer look: What problem does AP2 solve? Simply put, the MCP protocol is like a universal hook, enabling AI agents to connect to various external tools and data sources; A2A is a team collaboration communication protocol that allows multiple AI agents to cooperate with each other to complete complex tasks; AP2 completes the last piece of the puzzle - payment capability. In other words, MCP opens up connectivity, A2A promotes collaboration efficiency, and AP2 achieves value exchange. The arrival of AP2 truly injects "soul" into the autonomous collaboration and task execution of Multi-Agents. Imagine AI Agents connecting Qunar, Meituan, and Didi to complete the booking of flights, hotels, and car rentals, but then getting stuck at the point of "self-payment." What's the point of all that multitasking? So, remember this: AP2 is an extension of MCP+A2A, solving the last mile problem of AI Agent automated execution. What are the technical highlights of AP2? The core innovation of AP2 is the Mandates mechanism, which is divided into real-time authorization mode and delegated authorization mode. Real-time authorization is easy to understand. The AI Agent finds the product and shows it to you. The operation can only be performed after the user signs. Delegated authorization requires the user to set rules in advance, such as only buying the iPhone 17 when the price drops to 5,000. The AI Agent monitors the trigger conditions and executes automatically. The implementation logic is cryptographically signed using Verifiable Credentials (VCs). Users can set complex commission conditions, including price ranges, time limits, and payment method priorities, forming a tamper-proof digital contract. Once signed, the AI Agent executes according to the conditions, with VCs ensuring auditability and security at every step. Of particular note is the "A2A x402" extension, a technical component developed by Google specifically for crypto payments, developed in collaboration with Coinbase and the Ethereum Foundation. This extension enables AI Agents to seamlessly process stablecoins, ETH, and other blockchain assets, supporting native payment scenarios within the Web3 ecosystem. What kind of imagination space can AP2 bring? After analyzing the technical principles, do you think that's it? Yes, in fact, the AP2 is boring when it is disassembled alone. Its real charm lies in connecting and opening up the "MCP+A2A+AP2" technology stack, completely opening up the complete link of AI Agent's autonomous analysis+execution+payment. From now on, AI Agents can open up many application scenarios. For example, AI Agents for stock investment and financial management can help us monitor the market 24/7 and conduct independent transactions. Enterprise procurement AI Agents can automatically replenish and renew without human intervention. AP2's complementary payment capabilities will further expand the penetration of the Agent-to-Agent economy into more scenarios. Google obviously understands that after the technical framework is established, the ecological implementation must be relied upon, so it has brought in more than 60 partners to develop it, almost covering the entire payment and business ecosystem. Interestingly, it also involves major Crypto players such as Ethereum, Coinbase, MetaMask, and Sui. Combined with the current trend of currency and stock integration, the imagination space has been doubled. Is web3 AI really dead? Not entirely. Google's AP2 looks complete, but it only achieves technical compatibility with Crypto payments. It can only be regarded as an extension of the traditional authorization framework and belongs to the category of automated execution. There is a "paradigm" difference between it and the autonomous asset management pursued by pure Crypto native solutions. The Crypto-native solutions under exploration are taking the "decentralized custody + on-chain verification" route, including AI Agent autonomous asset management, AI Agent autonomous transactions (DeFAI), AI Agent digital identity and on-chain reputation system (ERC-8004...), AI Agent on-chain governance DAO framework, AI Agent NPC and digital avatars, and many other interesting and fun directions. Ultimately, once users get used to AI Agent payments in traditional fields, their acceptance of AI Agents autonomously owning digital assets will also increase. And for those scenarios that AP2 cannot reach, such as anonymous transactions, censorship-resistant payments, and decentralized asset management, there will always be a time for crypto-native solutions to show their strength? The two are more likely to be complementary rather than competitive, but to be honest, the key technological advancements behind AI Agents currently all come from web2AI, and web3AI still needs to keep up the good work!
Share
PANews2025/09/18 07:00