Tablinum-2: the Lite-Beam engram, release 2 (card, code, recipes, 9 released tables)
901f964 verified | """medrag_chunk.py IN.jsonl OUT.parquet [SOURCE] - MedRAG corpus snippets (textbooks / pubmed, {id, title, content}) -> engram | |
| passages, same schema as chunk.py (pid = file_idx << 32 | n, page_id, title, section, text) (10-20, health table). Text is | |
| "Title: content", split into WORDS-word windows when longer, so the re-ranker's 256-token window sees the whole passage. | |
| file_idx comes from OUT's name AAA_BB (BB + 100*AAA), as engram.Passages expects. Textbook titles are file names | |
| (Anatomy_Gray) -> "Anatomy Gray".""" | |
| import sys, re, json, os | |
| import pyarrow as pa, pyarrow.parquet as pq | |
| WORDS = 180 | |
| src, dst = sys.argv[1], sys.argv[2]; source = sys.argv[3] if len(sys.argv) > 3 else os.path.basename(src)[:-6] | |
| a, b = re.search(r"(\d+)_(\d+)\.parquet$", dst).groups(); fidx = int(b) + 100 * int(a) | |
| cols = {"pid": [], "page_id": [], "title": [], "section": [], "text": []}; n = 0 | |
| for line in open(src): | |
| r = json.loads(line); title = r["title"].strip() | |
| if source == "textbooks": | |
| title = title.replace("_", " ") | |
| w = r["content"].split() | |
| for i in range(0, max(len(w), 1), WORDS): | |
| part = w[i:i + WORDS + 30] if len(w) - i <= WORDS + 30 else w[i:i + WORDS] | |
| if len(part) < 8: | |
| break | |
| cols["pid"].append((fidx << 32) | n); cols["page_id"].append(int(r.get("PMID") or 0)); cols["title"].append(title) | |
| cols["section"].append(source); cols["text"].append(f"{title}: " + " ".join(part)); n += 1 | |
| if len(w) - i <= WORDS + 30: | |
| break | |
| pq.write_table(pa.table(cols), dst + ".tmp", compression="zstd", row_group_size=4096); os.replace(dst + ".tmp", dst) | |
| print(src, "->", dst, "passages", n, flush=True) | |