0.4.2

Release Notes

• Add visiongrep index DIR, which downloads missing models, updates the index and prepares both
  models without running a query, so the first search after it only embeds the query.
• Report indexing progress as Indexed N/TOTAL images lines every 256 images when stderr is not a
  terminal, instead of printing nothing until indexing finishes.
• Document use with AI agents in the README.
• Remove the cargo binstall metadata; VisionGrep is not published to crates.io.

Install visiongrep 0.4.2

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/udayj/visiongrep/releases/download/v0.4.2/visiongrep-installer.sh | sh

Download visiongrep 0.4.2

File	Platform	Checksum
visiongrep-aarch64-apple-darwin.tar.xz	Apple Silicon macOS	checksum
visiongrep-aarch64-unknown-linux-gnu.tar.xz	ARM64 Linux	checksum
visiongrep-x86_64-unknown-linux-gnu.tar.xz	x64 Linux	checksum

0.4.1

Release Notes

• Publish prebuilt macOS and Linux binaries with each GitHub release, installable with a shell
  installer or cargo binstall visiongrep. Add crates.io package metadata for
  cargo install visiongrep.

Install visiongrep 0.4.1

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/udayj/visiongrep/releases/download/v0.4.1/visiongrep-installer.sh | sh

Download visiongrep 0.4.1

File	Platform	Checksum
visiongrep-aarch64-apple-darwin.tar.xz	Apple Silicon macOS	checksum
visiongrep-aarch64-unknown-linux-gnu.tar.xz	ARM64 Linux	checksum
visiongrep-x86_64-unknown-linux-gnu.tar.xz	x64 Linux	checksum

visiongrep 0.4.0

Image queries and re-indexing are about 2.5× faster with about 65% less memory. The vision model now uses mixed-precision int8 (dynamic quantization, with the final projection and first transformer block kept in fp32).
Measured on AWS c7a.xlarge against v0.3.0:
image query: 871 → 336 ms, peak memory 596 → 207 MiB;
search after 1% of images change: 990 → 401 ms, peak memory 603 → 234 MiB;
text queries unchanged.
Accuracy: on COCO val2017 text-to-image retrieval (5,000 images, 25,014 captions), recall changed by −0.0035, −0.0031 and −0.0044 at 1, 5 and 10.
Smaller download: the vision model is now 119 MB instead of 352 MB.
Existing indexes re-embed their images once on first use.

visiongrep 0.3.0

• Add visiongrep serve --socket DIRECTORY, an opt-in server that keeps the index and model sessions loaded. Ordinary cached searches for the same directory and index use it automatically while it is running and fall back to one-shot searches otherwise. --no-cache, --reindex, --verify-models and --timing always run one-shot.
• serve now requires exactly one of --stdio or --socket.
• Refuse --reindex while a socket server is running for the same index.
• Save optimized ONNX graphs under models/optimized/ on first use and reuse them in later processes to reduce model startup time (about 578 MiB of extra disk; falls back to the source models if a graph cannot be used or saved).
• Replace the benchmark harness with benchmarks/bench.py: prepare, paired compare and quality runs, plus aws-run, aws-fetch and aws-list for self-terminating EC2 comparisons.

Measured on AWS c7a.xlarge against v0.2.0: novel text query 6.9% faster, external image query 8.3% faster, 1% changed corpus 8.3% faster, identical results.

VisionGrep 0.2.0

Run multiple searches in one process with the new serve --stdio command:

visiongrep serve --stdio --index-path photos.db ./photos

Send one JSON request per line through stdin and receive one JSON response per line on stdout:

{"id":"search-1","query":"a dog on a beach","top":10}
{"id":"search-2","image":"./reference.jpg","top":5}

• Supports text and image searches, optional request IDs, result limits, and similarity thresholds.
• Keeps the database connection open and reuses text and vision model sessions after their first use.
• Refreshes the image index before every search to pick up added, modified, and deleted images.
• Flushes each response immediately. Invalid requests return an error response so subsequent requests can continue.

Close stdin to end the server. Existing one-shot search commands remain available.

Benchmark Foundation v1

Merge pull request #4 from udayj/codex/fix-index-ordering-top-capacity

Fix incremental indexing order and excessive --top allocation