https://arxiv.org/abs/2606.28344. The end of web parsing. The beginning of scalable pixel-native search. link: https://pixelrag.ai/
docs: add collapsible PDF quickstart example to README (#58)
Yichuan Wang86e1de4f4198
diff --git a/README.md b/README.mdindex 72cacd7..650cc7f 100644--- a/README.md+++ b/README.md@@ -144,6 +144,8 @@ curl -X POST http://localhost:30001/search \### Build an index from your own documents+Works on **Linux (CUDA)** and **macOS (Apple Silicon / MPS)** — `device: auto` picks the best backend.+```bashpip install 'pixelrag[index]'@@ -155,8 +157,7 @@ source:embed:model: Qwen/Qwen3-VL-Embedding-2B- device: cuda- gpu_ids: [0]+ device: auto # cuda on Linux, mps on macOS, cpu as fallbackoutput: ./my_indexEOF@@ -166,6 +167,44 @@ pixelrag index buildpixelrag serve --index-dir ./my_index --port 30001```+<details>+<summary><strong>Try it: index a PDF and search it locally</strong></summary>++No GPU required — runs on macOS (Apple Silicon) or any machine with Python 3.10+.++```bash+pip install 'pixelrag[index]'++# 1. Grab a sample PDF (or use your own)+curl -L -o paper.pdf https://raw.githubusercontent.com/StarTrail-org/PixelRAG/main/assets/pixelrag-paper.pdf++# 2. Create config (device: auto picks MPS on Mac, CUDA on Linux)+cat > pixelrag.yaml << 'EOF'+source:+ type: local+ path: ./paper.pdf++embed:+ model: Qwen/Qwen3-VL-Embedding-2B+ device: auto++output: ./paper_index+EOF++# 3. Build the index (~3 min on Apple M-series, ~1 min on GPU)+pixelrag index build++# 4. Serve it+pixelrag serve --index-dir ./paper_index --port 30001++# 5. Search — should return page 2 (the overview diagram)+curl -X POST http://localhost:30001/search \+ -H "Content-Type: application/json" \+ -d '{"queries": [{"text": "Overview of PixelRAG and the diagram"}], "n_docs": 1}'+```++</details>+### Render a page programmatically```pythondiff --git a/embed/src/pixelrag_embed/embed.py b/embed/src/pixelrag_embed/embed.pyindex eea8ce7..e210b9a 100644--- a/embed/src/pixelrag_embed/embed.py+++ b/embed/src/pixelrag_embed/embed.py@@ -128,7 +128,9 @@ def _smart_resize_pil(img: "Image.Image", max_pixels: int) -> "Image.Image":return img.resize((new_w, new_h), Image.LANCZOS)-def _clamp_width_pil(img: "Image.Image", max_width: int = _MAX_CHUNK_WIDTH) -> "Image.Image":+def _clamp_width_pil(+ img: "Image.Image", max_width: int = _MAX_CHUNK_WIDTH+) -> "Image.Image":"""Resize image so width <= max_width, preserving aspect ratio.Dimensions are rounded to multiples of 28 (Qwen3-VL patch alignment).@@ -942,7 +944,10 @@ def _embed_tile_infos_with_engine(resized_wide[0] += 1logger.debug("Resized wide chunk %s ci=%d to %dx%d",- img_path, ci, cw, ch,+ img_path,+ ci,+ cw,+ ch,)# Dedup key: use file path for pre-chunked files (avoids# expensive tobytes + MD5 on ~2.7MB raw pixels per chunk)diff --git a/embed/src/pixelrag_embed/embed_cpu.py b/embed/src/pixelrag_embed/embed_cpu.pyindex e782e4e..3b717af 100644--- a/embed/src/pixelrag_embed/embed_cpu.py+++ b/embed/src/pixelrag_embed/embed_cpu.py@@ -174,7 +174,9 @@ def embed_items()inputs = processor(text=[text], images=[img], return_tensors="pt", padding=True)if device != "cpu":- inputs = {k: v.to(device) if hasattr(v, "to") else v for k, v in inputs.items()}+ inputs = {+ k: v.to(device) if hasattr(v, "to") else v for k, v in inputs.items()+ }with torch.no_grad():outputs = model(**inputs, output_hidden_states=True)@@ -222,7 +224,9 @@ def main():items = items[: args.limit]else:logger.info("Found %d chunks to embed", len(items))- embeddings = embed_items(items, args.model, device=args.device, instruction=args.instruction)+ embeddings = embed_items(+ items, args.model, device=args.device, instruction=args.instruction+ )output_path = Path(args.output_dir) / "shard_000.npz"np.savez(diff --git a/render/src/pixelrag_render/backends/pdf.py b/render/src/pixelrag_render/backends/pdf.pyindex 8897fcf..17f306c 100644--- a/render/src/pixelrag_render/backends/pdf.py+++ b/render/src/pixelrag_render/backends/pdf.py@@ -90,15 +90,17 @@ def render_pdf(saved_tiles.append(tile_name)w, h = img.size# Each PDF page = one chunk (no further splitting)- chunks_info.append({- "tile": tile_name,- "tile_index": idx,- "chunk_index": 0,- "file": tile_name,- "y_offset": 0,- "height": h,- "width": w,- })+ chunks_info.append(+ {+ "tile": tile_name,+ "tile_index": idx,+ "chunk_index": 0,+ "file": tile_name,+ "y_offset": 0,+ "height": h,+ "width": w,+ }+ )logger.debug(" Page %d → %s (%dx%d)", idx, tile_name, w, h)manifest = {