fix

backend/app/core/llm.py

@@ -16,6 +16,37 @@ from app.models import LlmCallLog
 log = get_logger("llm")
 
 
+# vendor-specific end-of-sequence / control tokens that some local models
+# emit into the content stream. Strip them so they don't leak to the user.
+_EOS_TOKENS = (
+    "<eos>",
+    "</s>",
+    "<|endoftext|>",
+    "<|im_end|>",
+    "<|end|>",
+    "<|eot_id|>",
+    "<|eom_id|>",
+)
+
+
+def _clean_model_text(text: str) -> str:
+    """Remove vendor-specific end-of-sequence tokens and collapse whitespace.
+
+    Some local models leak control tokens into the visible content stream.
+    We strip them so they never reach the user.
+    """
+    if not text:
+        return text
+    cleaned = text
+    for tok in _EOS_TOKENS:
+        cleaned = cleaned.replace(tok, "")
+    while "\n\n\n" in cleaned:
+        cleaned = cleaned.replace("\n\n\n", "\n\n")
+    return cleaned
+
+
+
+
 class LlmResponse:
     """Non-streaming response wrapper."""
 
@@ -93,7 +124,7 @@ class LlmClient:
                 data = resp.json()
             choice = (data.get("choices") or [{}])[0]
             msg = choice.get("message", {})
-            text = msg.get("content") or ""
+            text = _clean_model_text(msg.get("content") or "")
             tool_calls = msg.get("tool_calls") or []
             usage = data.get("usage") or {}
         except httpx.ConnectError as e:
@@ -206,8 +237,10 @@ class LlmClient:
                             continue
                         delta = choices[0].get("delta", {})
                         if delta.get("content"):
-                            full_text_parts.append(delta["content"])
-                            yield {"type": "delta", "content": delta["content"]}
+                            piece = _clean_model_text(delta["content"])
+                            if piece:
+                                full_text_parts.append(piece)
+                                yield {"type": "delta", "content": piece}
                         if delta.get("tool_calls"):
                             for tc in delta["tool_calls"]:
                                 idx = tc.get("index", 0)