Minish

We just released potion-code-16M-v2, our updated static embedding model for code retrieval! 🔍 Main features: - Better performance: 39.08 average NDCG@10 on the CoIR benchmark, up from 37.05 for the previous version, beating it on every single task. - Half the size: the new model is quantized by default and is just 32MB on disk. - Extremely fast: ~400x faster than CodeRankEmbed, on CPU, embedding >11k docs per second. - Better hybrid performance: the new model works well in a hybrid setup with BM25, reaching an NDCG@10 of 43.36. It’s the new default model for Semble, our code search library for coding agents. It's distilled from CodeRankEmbed using Model2Vec, then fine-tuned with Tokenlearn and contrastive training on a large set of query-document pairs from CornStack. 🔗 Links in the comments, let us know if you have any feedback!

We’ve just released Semble v0.3.0! 🚢 Biggest changes: - Automatic disk caching: indexes are now saved in the OS cache folder and reused when valid. First search builds the index; later searches are (much) faster. - Search more than code: new `--content` flag for searching docs, config, code, or all of them together. - JSON output: CLI and MCP results are now structured JSON, making Semble easier to use from agents and scripts. - Better file control: use `.sembleignore` to exclude files or force-include custom extensions. There are a few depcrecations and breaking changes, check out the release notes for a migration guide. 🔗 Links in the comments. Let us know if you have feedback or feature requests!

We’ve just released a new semble version! Biggest features:                                                                                      - Automatic reindexing: local paths are now watched for file changes. The index stays up to date without any manual steps.                                    - Token savings: run `semble savings` to see how many tokens Semble has saved across all your searches, broken down by period. See screenshot for an example.  - CLI and bash integration: use semble directly from the shell. This also adds full sub-agent support. - More benchmarks: our benchmarks now directly measure token efficiency, and we’ve added more comparisons to similar tools. 🔗 Links in the comments, let us know if you have any feedback or feature requests (and a big thanks for everyone who gave feedback already!

Today we're releasing Semble, a fast and accurate code search library built for agents 🤖! We're also releasing potion-code-16M, a small code-specialized static embedding model that powers Semble. Main features: - Fast: indexes a full codebase in ~250 ms and answers queries in ~1.5 ms, all on CPU (~200x faster indexing and ~10x faster queries than a code-specialized transformer). - Accurate: on par with code-specialized transformer models at a fraction of the size (benchmarks link in the comments). - MCP server: drop-in tool for Claude Code, Cursor, Codex, OpenCode, and any other MCP-compatible agent. Repos are cloned and indexed on demand. - Zero setup: no API keys, no GPU, no external services. Most embedding-based code search methods are either too slow to index on demand or need GPU infrastructure, while grep-style retrieval often cannot find the relevant content. Semble combines the speed and quality benefits of both, so agents waste less time and fewer tokens exploring. Under the hood, it combines static (semantic) embeddings with BM25 (lexical) search, followed by a code-optimized reranking step. We've also run extensive code-search benchmarks, which we're releasing alongside Semble. 🔗 Links to the package, benchmarks, and the model are in the comments. We'd love to hear your feedback!

Model2Vec 0.7.0 is out now, as well as a blogpost on model size reduction techniques! This release features a large number of ways to improve the distillation process. Highlights include: - 🔎 Vocabulary quantization: merge similar tokens to shrink the vocabulary and reduce model size. - ⚙️ Configurable pooling. Previously we only supported mean pooling. You can now choose between various options such as mean, first token, and last-token pooling. This leads to large performance improvements for decoder-style embedders such as Qwen3-Embedding. - 🐛 A large number of small improvements and bugfixes We also have a new blogpost on model size reduction where we showcase how to reduce model size by a factor of 15, creating a 6MB model (!) without impacting performance. Links are in the comments, as always, any feedback is appreciated!