Fix BM25: lecture paginee de la collection (evite 'too many SQL variables' sur 84k chunks -> BM25 enfin actif)

app.py

@@ -163,10 +163,17 @@ def build_bm25_index() -> None:
     if not _BM25_AVAILABLE:
         return
     try:
-        data = collection.get(include=["documents", "metadatas"])
-        _bm25_ids = data.get("ids", []) or []
-        _bm25_docs = data.get("documents", []) or []
-        _bm25_metas = data.get("metadatas", []) or []
+        # Lecture PAGINEE : un collection.get() global depasse la limite SQLite
+        # "too many SQL variables" sur un gros corpus (84k chunks) -> BM25 jamais construit.
+        total = collection.count()
+        _bm25_ids, _bm25_docs, _bm25_metas = [], [], []
+        _off, _step = 0, 2000
+        while _off < total:
+            _batch = collection.get(include=["documents", "metadatas"], limit=_step, offset=_off)
+            _bm25_ids.extend(_batch.get("ids", []) or [])
+            _bm25_docs.extend(_batch.get("documents", []) or [])
+            _bm25_metas.extend(_batch.get("metadatas", []) or [])
+            _off += _step
         if not _bm25_docs:
             _bm25_index = None
             logger.info("BM25 index empty (no documents in store yet).")