diff --git a/src/processor/processor.py b/src/processor/processor.py
index 98ab1e6..c7deef8 100644
--- a/src/processor/processor.py
+++ b/src/processor/processor.py
@@ -70,9 +70,28 @@ class OllamaClient:
 def simple_chunk(text: str, chunk_size: int = 800, overlap: int = 100) -> list[str]:
     """
     Split text into overlapping chunks by approximate word count.
-    Respects paragraph boundaries when possible.
+    Respects paragraph/line boundaries when possible.
+
+    Acepta párrafos separados por uno o más saltos de línea (Wikipedia y
+    trafilatura usan '\n' simple, lo que antes dejaba el documento entero como
+    un único 'párrafo' → un solo chunk gigante). Además subdivide por palabras
+    cualquier párrafo que por sí solo supere chunk_size.
     """
-    paragraphs = [p.strip() for p in text.split("\n\n") if p.strip()]
+    raw_paragraphs = [p.strip() for p in re.split(r"\n+", text) if p.strip()]
+
+    # Subdivide párrafos sobredimensionados en piezas de (chunk_size - overlap)
+    # palabras; así, al reinyectar 'overlap' palabras de solapamiento, ningún
+    # chunk resultante supera chunk_size.
+    piece_size = max(1, chunk_size - max(0, overlap))
+    paragraphs: list[str] = []
+    for para in raw_paragraphs:
+        words = para.split()
+        if len(words) <= chunk_size:
+            paragraphs.append(para)
+        else:
+            for i in range(0, len(words), piece_size):
+                paragraphs.append(" ".join(words[i:i + piece_size]))
+
     chunks = []
     current = []
     current_words = 0
@@ -81,10 +100,12 @@ def simple_chunk(text: str, chunk_size: int = 800, overlap: int = 100) -> list[s
         para_words = len(para.split())
         if current_words + para_words > chunk_size and current:
             chunks.append("\n\n".join(current))
-            # overlap: keep last paragraph
-            if overlap > 0 and current:
-                current = [current[-1]]
-                current_words = len(current[0].split())
+            # overlap: arrastra un tail de 'overlap' palabras (no el párrafo
+            # completo — eso duplicaba el tamaño cuando los párrafos eran grandes)
+            if overlap > 0:
+                tail = "\n\n".join(current).split()[-overlap:]
+                current = [" ".join(tail)]
+                current_words = len(tail)
             else:
                 current = []
                 current_words = 0
@@ -241,7 +262,10 @@ class ContentProcessor:
                              threshold=settings.quality_threshold, words=words)
                 continue
 
-            embedding = await self.ollama.embed(chunk[:1000])
+            # Embeber el chunk completo (ya acotado a ~chunk_size palabras).
+            # Antes truncaba a 1000 chars → el vector solo representaba el
+            # principio de cada chunk, degradando el ranking del RAG.
+            embedding = await self.ollama.embed(chunk)
 
             await self.db.add_chunk(
                 session_id=session_id,