Add md2d index + mapping + corpus

.gitattributes

@@ -58,3 +58,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+index/e5_Flat.index filter=lfs diff=lfs merge=lfs -text
+index/emb_e5.memmap filter=lfs diff=lfs merge=lfs -text

index/e5_Flat.index

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:edc7c08bb66f9fffb084014e1efafb7ce78182b330bacd94871eb492566fb644
+size 31374381

index/emb_e5.memmap

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:839b9cd05bcc1253476276b8c5ea9712109bc4d95ee354ec9a3e24bc46538129
+size 31374336

process_corpus.py

@@ -0,0 +1,78 @@
+#!/usr/bin/env python3
+import argparse
+import json
+from pathlib import Path
+
+def iter_spans(obj):
+    """
+    Yield dicts that look like spans with fields:
+    title, id_sp, text_sec
+    Works whether the JSON is a dict-of-dicts (with numeric keys) or nested.
+    """
+    if isinstance(obj, dict):
+        # Detect a span record
+        if all(k in obj for k in ("title", "id_sp", "text_sec")):
+            yield obj
+        # Otherwise, recurse into values
+        for v in obj.values():
+            yield from iter_spans(v)
+    elif isinstance(obj, list):
+        for v in obj:
+            yield from iter_spans(v)
+
+def main():
+    ap = argparse.ArgumentParser(description="Build ID mapping and unique content from multidoc2dial spans.")
+    ap.add_argument("--input", required=True, help="Path to multidoc2dial_doc.json")
+    ap.add_argument("--mapping_out", default="mapping.json", help="Output path for mapping dict (JSON)")
+    ap.add_argument("--contents_out", default="contents.jsonl", help="Output path for contents (JSONL)")
+    # If you ever need a different base suffix than '#1_0', change here:
+    ap.add_argument("--base_suffix", default="#1_0", help="Suffix to append after title before _<id_sp>")
+    args = ap.parse_args()
+
+    input_path = Path(args.input)
+    with input_path.open("r", encoding="utf-8") as f:
+        data = json.load(f)
+
+    # Dedup by exact text_sec (after strip)
+    text_to_int = {}           # normalized text_sec -> int_id
+    next_id = 1
+
+    # Mapping from full key -> int_id
+    key_to_int = {}            # "<title>#1_0_<id_sp>" -> int_id
+
+    # For writing contents.jsonl only once per unique int_id
+    seen_int_ids = set()
+
+    # First pass: assign int IDs by text_sec
+    for span in iter_spans(data):
+        title = span["title"].strip()
+        id_sp = str(span["id_sp"]).strip()
+        text_sec = span["text_sec"].strip()
+
+        full_key = f"{title}{args.base_suffix}_{id_sp}"
+
+        if text_sec not in text_to_int:
+            text_to_int[text_sec] = next_id
+            next_id += 1
+
+        int_id = text_to_int[text_sec]
+        key_to_int[full_key] = int_id
+
+    # Write mapping.json
+    with open(args.mapping_out, "w", encoding="utf-8") as f:
+        json.dump(key_to_int, f, ensure_ascii=False, indent=2)
+
+    # Write contents.jsonl (one line per unique text_sec)
+    # Ensure stable order by int_id
+    # Invert text_to_int to {int_id: text_sec}
+    int_to_text = {v: k for k, v in text_to_int.items()}
+    with open(args.contents_out, "w", encoding="utf-8") as f:
+        for int_id in sorted(int_to_text.keys()):
+            obj = {"id": int_id, "contents": int_to_text[int_id]}
+            f.write(json.dumps(obj, ensure_ascii=False) + "\n")
+
+    print(f"Wrote {len(key_to_int):,} mappings to {args.mapping_out}")
+    print(f"Wrote {len(int_to_text):,} unique contents to {args.contents_out}")
+
+if __name__ == "__main__":
+    main()