Show HN: anygrep – semantic search over notes, PDFs and images, on the CPU

Anygrep, a new MIT-licensed terminal tool from developer kuroko1t, performs semantic search over notes, PDFs and images using Google's EmbeddingGemma 2 model converted to int8 ONNX and run entirely on the CPU, keeping a single local index. On a Ryzen 7 7700 with the default 8 threads, anygrep indexes text at about 1,100 tokens/s and processes an image in about 1.2 s, while a search command takes about 1.5 s, mostly spent loading the model. The tool requires Python 3.10+, downloads a roughly 350 MB model on first use (200 MB more for image indexing), and is installed via pipx or uv tool install.

anygrep searches your notes, PDFs and screenshots by meaning, from the terminal. It runs one model EmbeddingGemma 2 https://huggingface.co/google/embeddinggemma-2 , int8 ONNX on the CPU and keeps one local index. The demo searches a folder that was already indexed, and the wait for the model to load is sped up. pipx install anygrep or: uv tool install anygrep uvx anygrep "when is the tax return due" ~/notes or try it without installing Requires Python 3.10+. Tested on Linux; CI also runs the tests on Windows and macOS Apple silicon . The model is downloaded on first use: about 350 MB, and 200 MB more once you index images. See docs/usage.md https://github.com/kuroko1t/anygrep/blob/main/docs/usage.md installing for Intel Macs, Windows on ARM and uninstalling. anygrep "wifi password reset" search here new files are indexed first anygrep "when is the tax return due" ~/notes search one folder anygrep --image whiteboard.jpg ~/Screenshots find similar pictures anygrep -l "invoice from acme" file names only; piped output is path:line:text anygrep -Q ~/notes interactive: the model stays loaded anygrep index --background ~/Pictures big folder: index in a low-priority process Results are up to 10 files, closest in meaning first -m N for more , not every match; queries are not regular expressions. A file whose text or name contains a query of 3 or more characters as typed an invoice number, an error code comes first, marked exact ; words found in many files are not boosted this way details https://github.com/kuroko1t/anygrep/blob/main/docs/usage.md exact-matches . Queries and documents can be in different languages. Weak matches are hidden --weak shows them and near-duplicate files are shown once. See anygrep -h and docs/usage.md https://github.com/kuroko1t/anygrep/blob/main/docs/usage.md . On a terminal, the first search in a big folder indexes the new files first, with a progress bar; Ctrl-C keeps what is done and the next run continues. index --background does it while you work your home folder is never indexed automatically, so name a folder inside it . In a script or cron job, a search indexes new files only if that takes under about 30 seconds; run anygrep index beforehand for a big folder details https://github.com/kuroko1t/anygrep/blob/main/docs/usage.md indexing-before-a-search . It reads text files .md .txt .html and more , the text layer of PDFs, and .jpg .png .webp .bmp .gif images. Hidden, ignored and binary files are skipped. python import anygrep with anygrep.Index as idx: the index the command uses idx.update "~/notes" index new and changed files for hit in idx.search "when is the tax return due", k=5 : print hit.score, hit.path, hit.snippet Install it with pip in your project's environment. It uses the same index, scores and filtering as the command and prints nothing. The API may change before 1.0; see docs/python-api.md https://github.com/kuroko1t/anygrep/blob/main/docs/python-api.md . On a Ryzen 7 7700 with the default 8 threads, text indexes at about 1,100 tokens/s and an image takes about 1.2 s. A search command takes about 1.5 s, mostly loading the model. More in docs/benchmarks.md https://github.com/kuroko1t/anygrep/blob/main/docs/benchmarks.md . See also how it works https://github.com/kuroko1t/anygrep/blob/main/docs/how-it-works.md , relevance https://github.com/kuroko1t/anygrep/blob/main/docs/relevance.md and comparison with other tools https://github.com/kuroko1t/anygrep/blob/main/docs/comparison.md . - No OCR: small text inside images is often missed, and scanned PDFs are skipped. - No audio, video or Office documents yet anygrep says how many Office files it skipped , and no file watching run anygrep index from cron . MIT. The model is downloaded separately: an int8 ONNX conversion https://huggingface.co/onnx-community/embeddinggemma-2-ONNX of Google's EmbeddingGemma 2, released under the Apache License 2.0 its model card also points to the Gemma Prohibited Use Policy . anygrep is not affiliated with Google.