https://arxiv.org/abs/2606.28344. The end of web parsing. The beginning of scalable pixel-native search. link: https://pixelrag.ai/
[project]
name = "pixelrag"
version = "0.4.0"
description = "Visual Retrieval-Augmented Generation — render, embed, index, search"
requires-python = ">=3.12"
authors = [{ name = "Zhifei Li", email = "[email protected]" }]
license = "Apache-2.0"
readme = "README.md"
keywords = ["rag", "retrieval", "vision-language", "screenshots", "embeddings", "faiss"]
classifiers = [
"License :: OSI Approved :: Apache Software License",
"Programming Language :: Python :: 3.12",
"Topic :: Scientific/Engineering :: Artificial Intelligence",
]
# Core stays light — rendering / screenshots only (no torch). The `pixelshot` command and
# the `pixelrag` umbrella work on this alone. Heavy ML stages are opt-in extras below.
dependencies = [
"pillow>=10.0.0",
"websockets>=12.0",
"pymupdf>=1.27.2.3",
"pyturbojpeg>=2.2.0",
"cef-capi-py>=131.3.5",
"anthropic>=0.102.0",
# Light, pure-Python utilities the umbrella CLI and the non-torch stage code
# (chunk / config parsing / progress bars) import at module load. Kept in core
# — not the heavy `embed`/`index` extras — so `pixelrag chunk`, `pixelrag index`,
# and the torch-free test suite import cleanly without pulling torch/faiss.
"numpy>=1.26.0",
"pyyaml>=6.0",
"tqdm>=4.60.0",
]
[project.optional-dependencies]
embed = [
"torch>=2.9.0",
"torchvision>=0.24.0",
"transformers>=4.57.0",
"faiss-cpu>=1.9.0",
]
serve = [
"fastapi>=0.115.0",
"uvicorn>=0.30.0",
"faiss-cpu>=1.9.0",
"transformers>=4.57.0",
"torch>=2.9.0",
"torchvision>=0.24.0", # transformers' Qwen3-VL processor needs it; cu129 via tool.uv.sources
"qwen-vl-utils",
"pydantic>=2.0.0",
]
index = ["pixelrag[embed]", "markdown>=3.4"]
qdrant = ["qdrant-client>=1.18.0"]
all = ["pixelrag[embed,serve,index,qdrant]"]
gpu = ["faiss-gpu-cu12>=1.13.2; sys_platform == 'linux'"]
playwright = ["playwright>=1.40.0"]
pdf = ["pdf2image>=1.16.0"]
kiwix = ["libzim>=3.6.0"]
distributed = ["boto3>=1.42.0"]
eval = [
"pandas>=2.0",
"Pillow>=10.0",
"trafilatura>=1.6",
"openai>=1.0",
"aiohttp>=3.9",
"datasets>=2.14",
"huggingface-hub>=0.20",
]
# Everything tests/ needs, torch-free. The serve-side deps are listed one by one
# rather than as `pixelrag[serve]` on purpose: that extra pulls torch,
# torchvision and transformers, which the tests never touch — api.py imports them
# lazily, inside the functions that encode queries. Keeping them out is what lets
# CI run the serve and backend suites at all (see the note on core `numpy`).
dev = [
"faiss-cpu>=1.9.0",
"fastapi>=0.115.0",
"httpx>=0.27", # fastapi's TestClient
"pixelrag[qdrant]",
"pydantic>=2.0.0",
"pytest>=8.0",
]
[project.scripts]
pixelshot = "pixelrag_render.render:main"
pixelrag = "pixelrag.cli:main"
[project.urls]
Homepage = "https://pixelrag.ai"
Repository = "https://github.com/StarTrail-org/PixelRAG"
[build-system]
requires = ["hatchling"]
build-backend = "hatchling.build"
[tool.hatch.build.targets.wheel]
# One distribution bundles the umbrella + every stage module.
packages = [
"src/pixelrag",
"render/src/pixelrag_render",
"embed/src/pixelrag_embed",
"index/src/pixelrag_index",
"serve/src/pixelrag_serve",
]
[tool.hatch.build.targets.sdist]
# Repo root holds multi-GB data dirs (.venv, tiles, arxiv, …); ship only the package sources.
include = [
"/src/pixelrag",
"/render/src",
"/embed/src",
"/index/src",
"/serve/src",
"/README.md",
"/LICENSE",
]
[tool.uv]
override-dependencies = ["nvidia-cudnn-cu12==9.20.0.48; sys_platform == 'linux'"]
environments = ["sys_platform == 'linux'", "sys_platform == 'darwin'"]
[tool.uv.sources]
torch = [{ index = "pytorch-cu129", marker = "sys_platform == 'linux'" }]
torchvision = [{ index = "pytorch-cu129", marker = "sys_platform == 'linux'" }]
pixelrag = { workspace = true }
[[tool.uv.index]]
name = "pytorch-cu129"
url = "https://download.pytorch.org/whl/cu129"
explicit = true
[tool.ruff]
exclude = ["tmp", "train", "demos", "docs", "skill"]
[tool.ruff.lint]
ignore = [
"E402", # module-level import not at top — common in ML codebases
"E741", # ambiguous variable name (l, O, I)
"BLE001", # blind exception catch — pervasive in data pipelines
"EXE001", # shebang without +x
"S110", # try-except-pass
"S112", # try-except-continue
"SIM102", # collapsible if
"SIM103", # return bool directly
"SIM113", # enumerate instead of manual counter
"SIM115", # use context manager for open
"SIM117", # combine with-statements
"ASYNC210", # blocking call in async
"ASYNC220", # blocking call in async
"ASYNC230", # blocking call in async
"UP006", # deprecated typing aliases
"UP017", # deprecated datetime aliases
"UP035", # deprecated typing imports
"UP041", # timeout-error alias
"UP045", # X | None instead of Optional
"PLW1509", # subprocess preexec_fn
"PLW1510", # subprocess without check
"PLR0402", # import-as-module-alias
"PLR1730", # if-stmt-min-max
"RUF012", # mutable-class-default
"RUF013", # implicit-optional
"RUF022", # unsorted-all
"RUF059", # unused-unpacked-variable
"TRY002", # raise-vanilla-class
"PIE790", # unnecessary-placeholder
"PIE810", # multiple-starts-ends-with
"PYI034", # non-self-return-type
"PERF102", # incorrect-dict-iterator
"PERF101", # unnecessary-cast-to-int
"DTZ005", # call-datetime-now-without-tzinfo
"C401", # unnecessary-generator-set
"C408", # unnecessary-collection-call
"FURB188", # slice-to-remove-prefix-or-suffix
]
[tool.pyright]
extraPaths = ["render/src", "embed/src", "index/src", "serve/src", "train/src"]
venvPath = "."
venv = ".venv"