![\"sweet-search](\"assets/sweet-search-banner-pixelated.svg\"){kind=tracking}
\n\n\n## **🍬 Local code search for AI coding agents 🍬**\n\n\nSix fast, purpose-built tools made for *Claude Code*, *Codex* & friends 👀\n\nEvery coding agent today reaches for grep + Read by reflex. *sweet-search* challenges the narrative 😎\n\n[](https://www.npmjs.com/package/sweet-search)\n[](https://github.com/mrsladoje/sweet-search/stargazers)\n[](LICENSE)\n[](package.json)\n[](#platform-support)\n[](#the-index)\n\n
![\"Five](\"assets/code-retrieval-stats.svg\")
\n\nsealed vault · exact paired results · five representative profiles · full 11-cell matrix and held-out/OOD replication below\n\n![\"Three](\"assets/task-completion-stats.svg\")
\n\nfrozen held-out set · 200 tasks · 1 rep · 1,200 rollouts · gpt-5.6-luna · opened once\n\n![\"ss-search](\"assets/tools/ss-search.svg\")
![\"ss-grep](\"assets/tools/ss-grep.svg\")
![\"ss-find](\"assets/tools/ss-find.svg\")
![\"ss-semantic](\"assets/tools/ss-semantic.svg\")
![\"ss-trace](\"assets/tools/ss-trace.svg\")
![\"ss-read](\"assets/tools/ss-read.svg\")
![\"ss-search](\"assets/tools/ss-search-io.svg\")
\n\nA hybrid search pipeline with late interaction reranking that returns actual code blocks.\n\nSOTA on [3 of 4 academic code-search benchmarks](#bench-paper-type).\n\n```mermaid\nflowchart TD\n Q([\"🔍 natural-language query\"]) --> ROUTE{{\"🧭 WASM CatBoost router · lexical / hybrid\"}}\n\n ROUTE --> BM[\"📑 BM25F![\"ss-grep](\"assets/tools/ss-grep-io.svg\")
\n\n### ⚡ 10.2× faster than ripgrep\n> Median, end to end, on 353 real queries across 5 repos (8.5–17.7× per repo, about 1 ms per query). Same match count as ripgrep on every query.\n\n![\"For](\"assets/tools/ss-grep-ngrams.svg\")
\n\n![\"ss-find](\"assets/tools/ss-find-io.svg\")
\n\n### 🎯 Exact matches, semantically reranked\n> The idea comes from LightOn's [ColGrep](https://github.com/lightonai/next-plaid/tree/main/colgrep). We rebuilt it on the `ss-grep` index, with our own MaxSim kernel.\n\n![\"ss-semantic](\"assets/tools/ss-semantic-io.svg\")
\n\n### 📍 Ask one file a question\n> You already know the file. `ss-semantic` gives you the code spans that best match your query, each with its line range. The rest of the file stays out of your context.\n\n![\"ss-trace](\"assets/tools/ss-trace-io.svg\")
\n\n### 🕸️ Know what breaks before you edit\n> Give `ss-trace` a symbol. It returns who calls it, what it calls, and what breaks if it changes, with the code. The graph updates as you edit, so the answer matches your current code.\n\n![\"ss-read](\"assets/tools/ss-read-io.svg\")
\n\n### 📖 Read the code, see what you skipped\n> Give `ss-read` a file and a line range. It reads from disk, so the code is always current, and it names what the agent has not read yet.\n\n![\"How](\"assets/index-build.svg\")
\n\n### 🛠️ How we prepare for embedding\n\n| 🧩 **[cAST](https://arxiv.org/abs/2506.15655) chunking** | 🏷️ **Chunk enrichment, tuned per language family** |\n|:--|:--|\n| Whole functions and classes when they fit the embedder's context, split when they don't | Prepends file path · scope chain · symbol · merged siblings · imports used, built from the AST and the code graph |\n| 96 languages · 206 extensions · 18 with full tree-sitter grammars | Each language family's policy was picked by *ablation* on GenCodeSearchNet |\n\n\n### 🚀 The embedding models, and how we made them blazing fast\n#### 🤖 Two open, code-specialized models\n> - [CodeRankEmbed](https://huggingface.co/nomic-ai/CodeRankEmbed) (137M, dense) for first-stage recall.\n> - [LateOn-Code](https://huggingface.co/lightonai/LateOn-Code) (149M, late interaction) for the rerank.\n> - [LateOn-Code-edge](https://huggingface.co/lightonai/LateOn-Code-edge) (17M, late interaction) as the edge fallback, auto-selected on weaker hosts.\n\n#### ⚡ GPU-accelerated\n> - Runs on candle with Metal and CUDA, plus a Neural Engine cascade on M3+.\n> - **Hand-written fused attention kernels** for both models.\n\n| Your hardware | What runs |\n|--|--|\n| 🍏 Apple Silicon (M1+) | candle **Metal**, BF16, fused SDPA attention |\n| 🍏 Apple Silicon (M3+) | … plus a **CoreML Neural Engine cascade** |\n| 🟩 NVIDIA GPU (SM 7.0+) | candle **CUDA**; **flash-attention** on Ampere+ |\n| 💻 No accelerator | **ONNX Runtime INT8**: tuned CPU path, 132 MB model |\n\n#### 🧠 Cache-sized CPU batches\n> - Each batch fits one layer's weights and activations in the CPU cache *(which we auto-detect)*.\n> - Uses every physical core.\n> - ONNX Runtime (ORT) drives the CPU path.\n\n#### 🗜️ Two quantizations: one for speed, one for size\n> - INT8 weights make the CPU build ~2× faster.\n> - INT4 vectors make the late-interaction index ~3× smaller, with no measurable retrieval loss.\n\n