fix: rewrite title generator to only use user messages; bump background model to qwen2.5:3b

src/fabledassistant/config.py

@@ -27,7 +27,7 @@ class Config:
     # Lightweight model for background tasks (title generation, tag suggestions,
     # project summaries, RSS classification). Using a separate model keeps the
     # main model's KV cache intact between user messages, enabling prefix cache hits.
-    OLLAMA_BACKGROUND_MODEL: str = os.environ.get("OLLAMA_BACKGROUND_MODEL", "qwen2.5:0.5b")
+    OLLAMA_BACKGROUND_MODEL: str = os.environ.get("OLLAMA_BACKGROUND_MODEL", "qwen2.5:3b")
     # KV cache context window for generation. Keep this as small as practical —
     # a larger context forces more KV cache into CPU RAM, drastically slowing prefill.
     # 16384 covers ~30+ message conversations with our system prompt comfortably.

src/fabledassistant/services/generation_task.py

@@ -118,34 +118,45 @@ _TOOL_LABELS: dict[str, str] = {
 
 
 async def _generate_title(messages: list[dict], user_id: int) -> str:
-    """Ask the LLM for a concise conversation title."""
-    # Build conversation text like summarize_conversation_as_note
-    conv_lines = []
+    """Ask the LLM for a concise conversation title.
+
+    Only uses user messages to avoid feeding tool-call JSON, system prompt
+    fragments, or other noise into the title generator. Caps input length
+    to keep the task fast and focused.
+    """
+    user_texts = []
     for m in messages:
-        if m["role"] in ("system", "tool"):
-            continue
-        label = "User" if m["role"] == "user" else "Assistant"
-        content = m.get("content", "")
-        if not content:
-            continue
-        conv_lines.append(f"{label}: {content}")
-    # Keep only last 6 pairs worth of text
-    conv_lines = conv_lines[-12:]
+        if m["role"] == "user":
+            content = (m.get("content") or "").strip()
+            if content:
+                user_texts.append(content[:300])
+    if not user_texts:
+        return ""
+    # First + last user messages capture intent best
+    if len(user_texts) > 2:
+        user_texts = [user_texts[0], user_texts[-1]]
 
     prompt_messages = [
         {
-            "role": "system",
+            "role": "user",
             "content": (
-                "Generate a concise 3-8 word title for this conversation. "
-                "Reply with ONLY the title, no quotes or punctuation."
+                "Generate a concise 3-8 word title for a conversation that started with:\n\n"
+                + "\n\n".join(user_texts)
+                + "\n\nReply with ONLY the title. No quotes, no punctuation, no explanation."
             ),
         },
-        {"role": "user", "content": "\n\n".join(conv_lines)},
     ]
     bg_model = await get_setting(user_id, "background_model", Config.OLLAMA_BACKGROUND_MODEL)
-    title = await generate_completion(prompt_messages, bg_model, max_tokens=30)
+    title = await generate_completion(prompt_messages, bg_model, max_tokens=30, num_ctx=1024)
+    # Strip common LLM noise: quotes, thinking tags, role labels
     title = title.strip().strip('"\'').strip()
-    return title[:100] if title else ""
+    for prefix in ("Title:", "title:", "Assistant:", "User:"):
+        if title.startswith(prefix):
+            title = title[len(prefix):].strip()
+    # Drop anything after a newline (model sometimes adds explanation)
+    if "\n" in title:
+        title = title.split("\n")[0].strip()
+    return title[:80] if title else ""
 
 
 async def _update_message(

src/fabledassistant/services/projects.py

@@ -124,7 +124,7 @@ async def generate_project_summary(user_id: int, project_id: int) -> None:
         from fabledassistant.services.settings import get_setting
         messages = [{"role": "user", "content": prompt}]
         bg_model = await get_setting(user_id, "background_model", Config.OLLAMA_BACKGROUND_MODEL)
-        summary = await generate_completion(messages, model=bg_model, max_tokens=400)
+        summary = await generate_completion(messages, model=bg_model, max_tokens=400, num_ctx=2048)
         if not summary:
             return