Upload 3 files

app.py

@@ -178,7 +178,7 @@ def ensure_db_ready():
             with open("ena_full_data.json", "r", encoding="utf-8") as f:
                 pages = json.load(f)
             
-            ts = RecursiveCharacterTextSplitter(chunk_size=400, chunk_overlap=80, separators=["\n\n", "\n", ".", " "])
+            ts = RecursiveCharacterTextSplitter(chunk_size=600, chunk_overlap=120, separators=["\n\n", "\n", ".", " "])
             all_chunks = []
             
             CATS = {"/concours/":"concours_fr","/formation/":"formation_fr","/gouvernance/":"gouvernance","/ar/concours":"concours_ar","/ar/formation":"formation_ar","/ar/service":"service_ar","/actualites/":"news"}
@@ -381,6 +381,13 @@ def build_citations_context(chunks: list):
     context_parts, ref_map, seen_urls, ref_num = [], {}, set(), 1
     for chunk in chunks:
         url = chunk.get("meta", {}).get("url", "")
+
+        # Filter out low-relevance chunks from unrelated pages
+        if "formation-continue" in url and chunk.get("rrf_score", 1) < 0.3:
+            continue
+        if "formation-continue-ar" in url and chunk.get("rrf_score", 1) < 0.3:
+            continue
+
         if url and url not in seen_urls:
             seen_urls.add(url)
             ref_map[ref_num] = {"url":url, "name":next((k for k,v in PAGE_URLS.items() if v==url), url)}

build_chroma.py

@@ -18,8 +18,8 @@ print(f"📄 عدد الصفحات: {len(pages)}")
 
 # chunking
 text_splitter = RecursiveCharacterTextSplitter(
-    chunk_size=400,
-    chunk_overlap=80,
+    chunk_size=600,
+    chunk_overlap=120,
     separators=["\n\n", "\n", ".", " "]
 )
 

chroma_ena_db.zip

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5c12ea10be12174bce9b33bb00df0708cffd58b25eb89f213e33741256b49649
-size 13924202
+oid sha256:e7f160c51d7c6a1de27f8a7cf3d149a015ebce731a6b79a69d6da93edd1aeed3
+size 17946089