feat: implement async retries in model chain for summary and retouch functions

Author: arinadi

model_manager.py

@@ -197,49 +197,33 @@ def try_model_chain(
     task_name: str = "task"
 ) -> Optional[object]:
     """
-    Try models in chain order (primary → fallbacks).
+    Try models in chain order (primary → fallbacks), one attempt per model.
     Returns first successful response, or None if all fail.
+    Retries are handled by the async caller.
     """
     models_to_try = model_chain.get("all", [])
     if not models_to_try:
         log("ERROR", f"No models available for {task_name}")
         return None
 
-    max_retries = getattr(config.Config, 'MODEL_RETRY_COUNT', 2)
-    retry_delay = getattr(config.Config, 'MODEL_RETRY_DELAY_SECONDS', 2.0)
-
     for model_name in models_to_try:
-        attempt = 1
-        while attempt <= max_retries:
-            try:
-                log("MODEL", f"Trying {model_name} for {task_name} (attempt {attempt}/{max_retries})...")
-                kwargs = {"model": model_name, "contents": contents}
-                if generation_config:
-                    kwargs["config"] = generation_config
-
-                response = gemini_client.models.generate_content(**kwargs)
-
-                if response and getattr(response, 'text', None):
-                    log("MODEL", f"Success with {model_name} for {task_name}")
-                    return response
-
-                log("MODEL", f"{model_name} returned empty response for {task_name}")
-                break
-
-            except Exception as e:
-                transient = is_transient_gemini_error(e)
-                log("MODEL", f"{model_name} failed for {task_name} (attempt {attempt}/{max_retries}): {e}")
-                if attempt >= max_retries or not transient:
-                    break
-
-                sleep_time = retry_delay * attempt
-                log("MODEL", f"Transient error detected, retrying {model_name} after {sleep_time}s...")
-                import time
-                time.sleep(sleep_time)
-                attempt += 1
-                continue
-
-        log("MODEL", f"Moving to next model after {model_name} for {task_name}")
+        try:
+            log("MODEL", f"Trying {model_name} for {task_name}...")
+            kwargs = {"model": model_name, "contents": contents}
+            if generation_config:
+                kwargs["config"] = generation_config
+
+            response = gemini_client.models.generate_content(**kwargs)
+
+            if response and getattr(response, 'text', None):
+                log("MODEL", f"Success with {model_name} for {task_name}")
+                return response
+
+            log("MODEL", f"{model_name} returned empty response for {task_name}")
+
+        except Exception as e:
+            log("MODEL", f"{model_name} failed for {task_name}: {e}")
+            continue
 
     log("ERROR", f"All models failed for {task_name}")
     return None

utils.py

@@ -117,7 +117,7 @@ def build_retouch_prompt() -> str:
 
 async def summarize_text(transcript: str, gemini_client) -> str:
     """Generates a journalist-friendly summary of the transcript.
-    Uses model chain: primary (gemma) → fallbacks.
+    Uses model chain: primary (gemma) → fallbacks with async retries.
     """
     if not gemini_client:
         return "Summarization disabled: Gemini API key not configured or client failed to load."
@@ -129,25 +129,35 @@ async def summarize_text(transcript: str, gemini_client) -> str:
     from model_manager import try_model_chain
 
     chain = get_model_chain("summary")
-    config = types.GenerateContentConfig(temperature=0.3)
-
-    response = await asyncio.to_thread(
-        try_model_chain,
-        gemini_client,
-        chain,
-        [prompt, transcript],
-        config,
-        "summary"
-    )
+    gen_config = types.GenerateContentConfig(temperature=0.3)
+    max_retries = config.Config.MODEL_RETRY_COUNT
+    retry_delay = config.Config.MODEL_RETRY_DELAY_SECONDS
+
+    for attempt in range(max_retries):
+        response = await asyncio.to_thread(
+            try_model_chain,
+            gemini_client,
+            chain,
+            [prompt, transcript],
+            gen_config,
+            "summary"
+        )
+
+        if response and response.text:
+            return response.text
+
+        # If all models failed and this wasn't the last attempt, retry
+        if attempt < max_retries - 1:
+            sleep_time = retry_delay * (attempt + 1)
+            log("GEMINI", f"All models failed for summary, retrying after {sleep_time}s...")
+            await asyncio.sleep(sleep_time)
 
-    if response and response.text:
-        return response.text
     return "❌ Error generating summary: all models failed"
 
 
 async def retouch_transcript(transcript: str, gemini_client) -> str:
     """Retouch/clean up transcript: fix typos, punctuation, add paragraph breaks.
-    Uses model chain: primary (gemma) → fallbacks.
+    Uses model chain: primary (gemma) → fallbacks with async retries.
     """
     if not gemini_client:
         return transcript  # Return original if no client
@@ -159,19 +169,29 @@ async def retouch_transcript(transcript: str, gemini_client) -> str:
     from model_manager import try_model_chain
 
     chain = get_model_chain("retouch")
-    config = types.GenerateContentConfig(temperature=0.3)
-
-    response = await asyncio.to_thread(
-        try_model_chain,
-        gemini_client,
-        chain,
-        contents,
-        config,
-        "retouch"
-    )
+    gen_config = types.GenerateContentConfig(temperature=0.3)
+    max_retries = config.Config.MODEL_RETRY_COUNT
+    retry_delay = config.Config.MODEL_RETRY_DELAY_SECONDS
+
+    for attempt in range(max_retries):
+        response = await asyncio.to_thread(
+            try_model_chain,
+            gemini_client,
+            chain,
+            contents,
+            gen_config,
+            "retouch"
+        )
+
+        if response and response.text:
+            return response.text
+
+        # If all models failed and this wasn't the last attempt, retry
+        if attempt < max_retries - 1:
+            sleep_time = retry_delay * (attempt + 1)
+            log("GEMINI", f"All models failed for retouch, retrying after {sleep_time}s...")
+            await asyncio.sleep(sleep_time)
 
-    if response and response.text:
-        return response.text
     return transcript  # Return original on error