Original Summary

I’ve been working on sift-light , a local search tool for AI agents. Between 1.1.0-1 and 1.1.1-7 , the work has focused on a problem that sounds simple until you try to solve it: An agent finds a relevant passage, gives you a real file path and line number, and then draws a conclusion the source doesn’t actually support. Sometimes the file changed after the search. Sometimes the result is related to the question but doesn’t answer it. Sometimes the answer depends on an exception sitting just outside the excerpt. Here’s a small example: “Refunds are allowed within 30 days.” A few lines later: “Gift cards are excluded.” Ask whether a ten-day-old gift card purchase is refundable, and the first sentence alone looks convincing. The citation can be accurate while the answer is wrong. That’s the kind of gap these releases address. They also add image and scanned-PDF retrieval, because useful project information doesn’t always live in searchable text. I’m the maintainer. Here’s what has shipped, what we tested, and where the limits are. 1. Finding relevant material when the wording is different Exact search is useful when you know the identifier, filename, or phrase. It gets harder when you know what something does but have no idea what it’s called. From 1.1.0-1 , sift-light adds an independent native text retrieval lane alongside the existing text retrieval path, with bounded rank fusion. In practical terms: You can search for a concept using ordinary language. Both retrieval lanes remain available independently. The combined ranking doesn’t remove access to the individual results. Results still lead back to source locations you can inspect and validate. That last part matters to me. A semantic match is a useful lead; I still want to see the actual source before trusting a conclusion. 2. Searching images and scanned PDF pages A diagram in a project folder or a scanned page can contain exactly what you need, while ordinary text search has very little to work with. The multimodal path added in 1.1.0-1 supports: Text → image: describe what you’re looking for and retrieve matching images. Image → image: use an image to find visually related material. Scanned PDF pages: search rendered pages, including image-only and mixed PDFs, with original page numbers and previews. For example, you can look for “a diagram showing a queue between two services” without knowing the image’s filename. These features are optional and require configuration. The bundled native media path requires Apple Silicon and separately configured workers and model files. Ordinary exact text search remains model-free. 3. Checking whether the source has changed A result can be correct when retrieved and become outdated before the agent uses it. In 1.1.1-1 , retained ordinary search results were bound to verified content identities. Source inspection, navigation, and saved-evidence validation recheck that identity. If the file changed or disappeared, the tool reports that state explicitly. It

  • 情报分类：商业与市场研究
  • 分类依据：内容涉及商业、投资或市场动态
  • 信息来源：Reddit · SideProject
  • 发布时间：2026/10/14 10:53:57