Sweet-search is a code-search tool that indexes the working tree (including uncommitted edits) and offers ss-* commands for fast, cheap code search and navigation; follow precise search practices, use broad-to-narrow strategy, and verify sources by tracing symbols, with instructions about autogenerated files and output formatting.
The paper formalizes bidirectional function association between source and decompiled code, trains a Qwen3-Embedding model with contrastive learning, and shows superior generalization and performance on baseline tasks.
Leviathan indexes records in a single SQLite file so agents can ask plain-word questions and get short, cited answers. The author reports that answers on one million records used a median of 436 tokens, compared with more than 100,000 tokens using grep.
MAST is a code-search tool for TypeScript/JavaScript that indexes AST-based chunks in SQLite and uses a fusion of BM25 and a declaration-exact ranker to return only the relevant code declarations. It operates as both an MCP server and a CLI, relies on a symbol graph for fast structural answers, and requires local processing with specific development toolchains. New files require reindexing, and results include token-use statistics for efficiency.
Microsoft is updating Outlook search to return seven ranked results and to improve relevance, using feedback, better stop-word handling, and a distilled language model; improvements include offline indexing for poor connectivity and faster latency for slow searches.
Translated title. Original: 知識グラフ化ツールGraphifyでgrep/Readを置き換えたら、トークンは何倍減るのか実測した
This article reviews Graphify-Labs/graphify, a Python CLI that transforms code, docs, schemas, and configs into a knowledge graph, enabling offline analysis with tree-sitter and a --code-only mode. It discusses token reduction when AI agents query the graph instead of grep or Read, and reports on an actual test of the ai-100days repository using the tool in offline mode.
DeepSeek releases a faster, cheaper AI model (V4.1-Flash-1017) and opens its web search APIs to developers, boasting improved reasoning and shorter responses. It claims strong benchmark performance against competitors and updates its Harness tool to support the new model while lowering runtime and costs.
Chroma is described as open-source AI search infrastructure, with more than 26,000 GitHub stars, use in over 90,000 codebases, and over 11 million monthly downloads. The text promotes the control and flexibility of open-source databases and links to Chroma’s GitHub and Discord.
The document explains semantic search by contrasting it with keyword search and outlining key concepts—embeddings, vectors, and similarity measures. It includes a basic semantic-search pipeline and sample JavaScript code for computing vector magnitude and Euclidean distance.
The post argues that AI coding agents need both repository information and evidence from build, security, deployment, and runtime systems. It reviews Augment’s reported retrieval tests and Harness’s Code Context Engine and Software Delivery Knowledge Graph, while noting limits in the available benchmarks.