pixelrag
Use PixelRAG for visual document RAG: screenshot capture, visual index building, FAISS/Qdrant serving, evaluation reproduction, and LoRA training/data workflows.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- pixelrag
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Use PixelRAG for visual document RAG: screenshot capture, visual index building, FAISS/Qdrant serving, evaluation reproduction, and LoRA training/data workflows.
- disable-model-invocation
- true
02 资源文件 7
- 参考文档references/agent-integration.md1.7 KB
- 参考文档references/deployment-and-operations.md1.8 KB
- 参考文档references/install-and-routing.md2.5 KB
- 参考文档references/repo-provenance.md1.6 KB
- 参考文档references/repo-routing-metadata.json517 B
- 参考文档references/troubleshooting.md2.6 KB
- 脚本scripts/pixelrag_doctor.py2.6 KB
03 SKILL.md 正文
PixelRAG
Use this skill when a task names PixelRAG, pixelshot, pixelrag, visual RAG over screenshots, document screenshot tiles, Qwen3-VL visual embeddings, FAISS/Qdrant visual indexes, PixelRAG search APIs, paper reproduction, or the PixelRAG LoRA training data pipeline.
First Checks
- Check staleness: Read repo-provenance.md before using this skill with a checkout; refresh if the commit, package version, or evidence paths changed.
- Install minimally: Base capture is
pip install pixelrag. Add only the needed extra:pixelrag[index],pixelrag[serve],pixelrag[qdrant],pixelrag[pdf], orpixelrag[dev]. - Verify commands:
python -c "import pixelrag, pixelrag_render; print('ok')" pixelshot --help pixelrag --help - Run diagnostics: Use pixelrag_doctor.py before deeper debugging.
- Read troubleshooting: Use troubleshooting.md for install, optional dependency, CUDA, Chrome, model-download, data/config, and stale-skill failures.
Route by Task
| User request | Use this route |
|---|---|
| Capture URLs, authenticated browser pages, PDFs, HTML, text, or images into screenshot tiles | render-capture |
Build a local visual index from documents, debug pixelrag.yaml, chunk/embed/build FAISS or Qdrant indexes |
index-build |
Start or query pixelrag serve, call /search, use FAISS/Qdrant backends, fetch tile images, debug departments |
serve-search |
| Reproduce paper benchmark cells, configure reader/search serves, run eval graders, debug low scores | evaluation-reproduction |
| Use the released LoRA adapter, prepare synthetic data, run or debug the separate training project | training-and-data |
Minimal Patterns
Render a page or local file:
pixelshot https://example.org --output ./tiles
pixelshot paper.pdf --output ./tiles --dpi 200
Build an index from local documents:
source:
type: local
path: ./my_docs
embed:
model: Qwen/Qwen3-VL-Embedding-2B
device: auto
output: ./my_index
pixelrag index build --config pixelrag.yaml --limit 10
Serve and search:
pixelrag serve --index-dir ./my_index --articles-json ./my_index/articles.json --tiles-dir ./my_index/tiles --port 30001
curl -s -X POST http://localhost:30001/search \
-H "Content-Type: application/json" \
-d '{"queries":[{"text":"overview diagram"}],"n_docs":5}'
References
- install-and-routing.md covers package extras, CLI dispatch, and workflow selection.
- agent-integration.md covers hosted API use, pixelbrowse, and agent/tool integration boundaries.
- deployment-and-operations.md covers production topology and safe blue-green deployment concepts.
- repo-routing-metadata.json is structured router metadata for managed repo-skill import.
Important Boundaries
- Do not treat full embedding, benchmark reproduction, or training as cheap: they may download large Qwen/HF assets, load huge FAISS indexes, start long-running services, require GPUs, and consume paid API keys.
- The
train/workflow is a separate uv project. Do not install it into the root PixelRAG package environment unless the task explicitly asks for training. - Do not run deploy scripts, systemd units, or nginx switch scripts unless the user confirms they are on the deploy host and wants operational changes.
- Do not depend on original repo examples, docs, or tests at runtime; this skill bundles distilled references and safe helper scripts.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).