Retry truncated Ollama character analysis

ltx_director.py

@@ -438,9 +438,10 @@ _PROVIDER_DEFAULTS = {
 }
 
 _ANALYZE_PROMPT = (
-    "Describe the character's physical appearance in two concise sentences. "
+    "Describe only the visible character in exactly two complete sentences. "
-    "Specify their hair color/style, face details, and their clothing type/color. "
+    "Sentence 1 must cover hair, face, age impression, and any standout physical traits. "
-    "Keep the entire response very brief."
+    "Sentence 2 must cover clothing, accessories, and overall silhouette. "
+    "Do not start with fragments like 'The'. Do not mention image quality, background, framing, or emotions."
 )
 
 
@@ -470,6 +471,17 @@ def _extract_ollama_generated_text(resp_json: dict) -> str:
     return ""
 
 
+def _analysis_text_is_usable(text: str) -> bool:
+    text = (text or "").strip()
+    if not text:
+        return False
+    if len(text) < 24:
+        return False
+    if len(text.split()) < 6:
+        return False
+    return True
+
+
 def _resolve_provider(data):
     provider = (data.get("provider") or "ollama").lower()
     defs = _PROVIDER_DEFAULTS.get(provider, _PROVIDER_DEFAULTS["ollama"])
@@ -515,8 +527,15 @@ async def analyze_character_endpoint(request):
             async with aiohttp.ClientSession() as session:
                 if provider == "ollama":
                     payload = {
-                        "model": model_name, "prompt": _ANALYZE_PROMPT,
+                        "model": model_name,
-                        "images": cleaned_b64_list, "stream": False, "keep_alive": 0,
+                        "prompt": _ANALYZE_PROMPT,
+                        "images": cleaned_b64_list,
+                        "stream": False,
+                        "keep_alive": 0,
+                        "options": {
+                            "temperature": 0.2,
+                            "num_predict": 120,
+                        },
                     }
                     async with session.post(f"{base_url}/api/generate", json=payload, timeout=300) as response:
                         if response.status != 200:
@@ -525,9 +544,9 @@ async def analyze_character_endpoint(request):
                         resp_json = await response.json()
                         generated_text = _extract_ollama_generated_text(resp_json)
 
-                    # Some Ollama model variants return an empty `response` from `/api/generate`
+                    # Some Ollama model variants return an empty or truncated response from
-                    # even though the same request succeeds via the chat endpoint.
+                    # `/api/generate` even though the same request succeeds via the chat endpoint.
-                    if not generated_text:
+                    if not _analysis_text_is_usable(generated_text):
                         chat_payload = {
                             "model": model_name,
                             "messages": [{
@@ -537,6 +556,10 @@ async def analyze_character_endpoint(request):
                             }],
                             "stream": False,
                             "keep_alive": 0,
+                            "options": {
+                                "temperature": 0.2,
+                                "num_predict": 120,
+                            },
                         }
                         async with session.post(f"{base_url}/api/chat", json=chat_payload, timeout=300) as response:
                             if response.status == 200:
@@ -574,10 +597,10 @@ async def analyze_character_endpoint(request):
 
         if "<think>" in generated_text:
             generated_text = generated_text.split("</think>")[-1].strip()
-        if not generated_text:
+        if not _analysis_text_is_usable(generated_text):
             return web.json_response({
                 "status": "error",
-                "message": f"{provider} returned an empty analysis response.",
+                "message": f"{provider} returned an empty or truncated analysis response.",
             })
 
         log.info("[LTXDirector] Analysis complete: %s", generated_text)