#!/usr/bin/env python3 """Build the public release tree from data/ artifacts. Reads: ../data/embeddings.npy float32 (N, 384) ../data/arxiv_ids.npy object (N,) ../data/meta.parquet paper_id, arxiv_id, title, primary_category, categories, n_chunks ../data/chunk_counts.npy int32 (N,) ../data/{dire,umap}_layout_n{n}_d{d}.npy float32 (N, d) for d in {2,3} Writes: metadata.parquet arxiv_id, title, primary_category, categories, n_chunks embeddings/part-NNNNN-of-NNNNN.parquet layouts/layouts_2d.parquet arxiv_id, method, n_neighbors, x, y layouts/layouts_3d.parquet arxiv_id, method, n_neighbors, x, y, z provenance/embedding_config.json provenance/software_versions.json provenance/file_hashes.json This is the single source of truth for what is in the public release. Re-running it on the same inputs is reproducible. """ from __future__ import annotations import argparse import hashlib import json import platform import sys import time from datetime import datetime, timezone from pathlib import Path import numpy as np import pandas as pd import pyarrow as pa import pyarrow.parquet as pq EMBED_DIM = 384 EMBED_MODEL = "BAAI/bge-small-en-v1.5" EMBED_POOLING = "mean" EMBED_NORMALIZATION = "L2 (chunks were unit-norm; pooled vectors re-normalized to unit L2 before downstream use)" EMBED_DTYPE_RELEASE = "float32" # --------------------------------------------------------------------------- # Helpers # --------------------------------------------------------------------------- def parse_categories(raw): """Categories column is JSON-string text from postgres jsonb. Parse to list.""" if raw is None or raw == "": return [] if isinstance(raw, list): return [str(x) for x in raw] if isinstance(raw, str): try: v = json.loads(raw) except json.JSONDecodeError: return [s for s in raw.split() if s] if isinstance(v, list): return [str(x) for x in v] if isinstance(v, str): return [v] if v else [] return [] def sha256_file(path: Path, chunk: int = 1024 * 1024) -> str: h = hashlib.sha256() with open(path, "rb") as f: while True: buf = f.read(chunk) if not buf: break h.update(buf) return h.hexdigest() def write_metadata(meta_in: pd.DataFrame, out: Path) -> Path: """Write public metadata.parquet — drops internal paper_id, parses categories list.""" df = pd.DataFrame({ "arxiv_id": meta_in["arxiv_id"].astype(str), "title": meta_in["title"].astype(str), "primary_category": meta_in["primary_category"].astype(str), "categories": [parse_categories(c) for c in meta_in["categories"]], "n_chunks": meta_in["n_chunks"].astype("int32"), }) schema = pa.schema([ ("arxiv_id", pa.string()), ("title", pa.string()), ("primary_category", pa.string()), ("categories", pa.list_(pa.string())), ("n_chunks", pa.int32()), ]) table = pa.Table.from_pandas(df, schema=schema, preserve_index=False) path = out / "metadata.parquet" pq.write_table(table, path, compression="zstd") return path def write_embeddings(arxiv_ids: np.ndarray, emb: np.ndarray, out: Path, n_shards: int) -> list[Path]: """Write sharded embedding parquets with fixed_size_list.""" assert emb.dtype == np.float32, f"expected float32, got {emb.dtype}" assert emb.shape[1] == EMBED_DIM assert emb.shape[0] == arxiv_ids.shape[0] embed_dir = out / "embeddings" embed_dir.mkdir(exist_ok=True) schema = pa.schema([ ("arxiv_id", pa.string()), ("embedding", pa.list_(pa.float32(), EMBED_DIM)), # fixed_size_list ]) n = emb.shape[0] bounds = np.linspace(0, n, n_shards + 1, dtype=int) paths: list[Path] = [] width = max(5, len(str(n_shards))) for i in range(n_shards): a, b = int(bounds[i]), int(bounds[i + 1]) ids_slice = arxiv_ids[a:b].astype(str) emb_slice = emb[a:b] # Build the fixed-size-list array: backing values are the flat float32 buffer. flat = pa.array(emb_slice.reshape(-1), type=pa.float32()) embed_arr = pa.FixedSizeListArray.from_arrays(flat, EMBED_DIM) ids_arr = pa.array(ids_slice, type=pa.string()) table = pa.Table.from_arrays([ids_arr, embed_arr], schema=schema) fname = f"part-{i:0{width}d}-of-{n_shards:0{width}d}.parquet" path = embed_dir / fname pq.write_table(table, path, compression="zstd", use_dictionary=False) # embeddings don't dict-compress paths.append(path) print(f" wrote {path.name}: rows {a}..{b} ({b - a:,}), " f"{path.stat().st_size / 1e6:.1f} MB") return paths def write_layouts(arxiv_ids: np.ndarray, layout_specs: list[dict], out: Path) -> list[Path]: """Long-format layouts: arxiv_id, method, n_neighbors, x, y[, z]. layout_specs: list of {"path": Path, "method": str, "n": int, "d": int}. """ layouts_dir = out / "layouts" layouts_dir.mkdir(exist_ok=True) n_papers = arxiv_ids.shape[0] paths: list[Path] = [] for d, fname in [(2, "layouts_2d.parquet"), (3, "layouts_3d.parquet")]: rows: list[pa.Table] = [] for spec in layout_specs: if spec["d"] != d: continue xy = np.load(spec["path"]) assert xy.shape == (n_papers, d), ( f"layout {spec['path']} shape {xy.shape} != ({n_papers}, {d})") cols: dict[str, pa.Array] = { "arxiv_id": pa.array(arxiv_ids.astype(str), type=pa.string()), "method": pa.array([spec["method"]] * n_papers, type=pa.string()), "n_neighbors": pa.array(np.full(n_papers, spec["n"], dtype=np.int32)), "x": pa.array(xy[:, 0].astype(np.float32)), "y": pa.array(xy[:, 1].astype(np.float32)), } if d == 3: cols["z"] = pa.array(xy[:, 2].astype(np.float32)) tbl = pa.table(cols) rows.append(tbl) print(f" layout: {spec['method']} n={spec['n']} d={d} " f"({spec['path'].name})") if not rows: continue big = pa.concat_tables(rows) path = layouts_dir / fname pq.write_table(big, path, compression="zstd") paths.append(path) print(f" wrote {path.name}: {big.num_rows:,} rows, " f"{path.stat().st_size / 1e6:.1f} MB") return paths def write_provenance(out: Path, data_dir: Path) -> list[Path]: prov = out / "provenance" prov.mkdir(exist_ok=True) embedding_cfg = { "model": EMBED_MODEL, "model_dim": EMBED_DIM, "chunk_embedding_normalization": "L2 unit norm (BGE default)", "pooling": EMBED_POOLING, "post_pooling_normalization": ( "L2 unit norm applied at use-time by run_reducer.py; " "embeddings stored here are the *un-normalized* mean-pooled vectors " "(norms ~0.83-1.0)" ), "released_dtype": EMBED_DTYPE_RELEASE, "n_papers": int(np.load(data_dir / "embeddings.npy", mmap_mode="r").shape[0]), "embedding_dim": EMBED_DIM, "source_text": ( "LaTeX-derived chunk text from arXiv source files; chunks are author-" "level paper text. Raw text/PDF/source files are NOT redistributed in " "this release. Only the numerical mean-pooled embedding per paper is " "released, alongside arXiv descriptive metadata." ), } p1 = prov / "embedding_config.json" p1.write_text(json.dumps(embedding_cfg, indent=2) + "\n") versions = { "build_timestamp_utc": datetime.now(timezone.utc).isoformat(), "python": sys.version, "platform": platform.platform(), "numpy": np.__version__, "pandas": pd.__version__, "pyarrow": pa.__version__, } # Best-effort versions for the analysis stack for mod in ("dire", "cuml", "ripser", "fastdtw", "umap"): try: m = __import__(mod) versions[mod] = getattr(m, "__version__", "unknown") except ImportError: pass p2 = prov / "software_versions.json" p2.write_text(json.dumps(versions, indent=2) + "\n") return [p1, p2] def write_hashes(out: Path) -> Path: """Walk release tree, write SHA256 of every released file (excl. file_hashes.json itself).""" hashes: dict[str, str] = {} for f in sorted(out.rglob("*")): if not f.is_file(): continue rel = f.relative_to(out).as_posix() if rel == "provenance/file_hashes.json": continue hashes[rel] = sha256_file(f) p = out / "provenance" / "file_hashes.json" p.write_text(json.dumps(hashes, indent=2, sort_keys=True) + "\n") return p # --------------------------------------------------------------------------- # Main # --------------------------------------------------------------------------- def main(): ap = argparse.ArgumentParser() here = Path(__file__).resolve().parent ap.add_argument("--data", type=Path, default=here.parent / "data", help="source data directory") ap.add_argument("--out", type=Path, default=here, help="release tree destination") ap.add_argument("--n-shards", type=int, default=8, help="number of embedding parquet shards") args = ap.parse_args() t0 = time.time() print(f"data: {args.data}") print(f"out: {args.out}") # Load source artifacts print("\n[1/6] Loading source artifacts...") meta = pd.read_parquet(args.data / "meta.parquet") arxiv_ids_npy = np.load(args.data / "arxiv_ids.npy", allow_pickle=True) emb = np.load(args.data / "embeddings.npy") # full load: ~1.1 GB # Alignment check: meta.parquet rows must correspond 1-1 with embeddings.npy rows assert len(meta) == len(arxiv_ids_npy) == emb.shape[0], ( f"row-count mismatch: meta={len(meta)}, arxiv_ids={len(arxiv_ids_npy)}, " f"emb={emb.shape[0]}") if not (meta["arxiv_id"].astype(str).values == arxiv_ids_npy.astype(str)).all(): raise SystemExit("arxiv_id ordering in meta.parquet != arxiv_ids.npy — " "release would be misaligned.") print(f" N = {len(meta):,} papers, dim = {emb.shape[1]}, dtype = {emb.dtype}") # 1. Metadata print("\n[2/6] Writing metadata.parquet...") meta_path = write_metadata(meta, args.out) print(f" wrote {meta_path.name}: " f"{meta_path.stat().st_size / 1e6:.1f} MB") # 2. Embeddings (sharded) print(f"\n[3/6] Writing {args.n_shards} embedding shards...") write_embeddings(arxiv_ids_npy, emb, args.out, args.n_shards) # 3. Layouts print("\n[4/6] Writing layouts...") layout_specs = [] for n in [8, 16, 32, 64, 128]: for method in ["dire", "umap"]: p = args.data / f"{method}_layout_n{n}_d2.npy" if p.exists(): layout_specs.append({"path": p, "method": method, "n": n, "d": 2}) # 3-d only at n=16 in current data for method in ["dire", "umap"]: p3 = args.data / f"{method}_layout_n{n}_d3.npy" if p3.exists(): layout_specs.append({"path": p3, "method": method, "n": n, "d": 3}) write_layouts(arxiv_ids_npy, layout_specs, args.out) # 4. Provenance print("\n[5/6] Writing provenance...") write_provenance(args.out, args.data) # 5. Hashes (last — must hash everything written above) print("\n[6/6] Computing SHA256 hashes...") hp = write_hashes(args.out) print(f" wrote {hp.relative_to(args.out)}") print(f"\nDone in {time.time() - t0:.1f}s.") print(f"Release tree: {args.out}") if __name__ == "__main__": main()