Update src/rag_pipeline.py

src/rag_pipeline.py

@@ -156,18 +156,26 @@ GROQ_API_KEY = os.environ.get("GROQ_API_KEY")
 
 # OpenRouter free-model fallback chain. If the primary 429s (upstream provider
 # throttled) or fails, LangChain's .with_fallbacks() transparently retries the
-# next one. Ordered by observed reliability + quality for structured/SOAP output
-# as of July 2026. Multi-provider models (Llama 3.3 70B, GPT-OSS) go first
-# because they're less prone to single-provider throttling than Gemma/Qwen.
+# next one. Ordered for July 2026 by structured-output quality × reliability;
+# openrouter/free (the auto-router across all active free models) sits at the
+# end as a catch-all when every specific pin is throttled.
+# NOTE: the real ceiling here is OpenRouter's PER-ACCOUNT daily cap (50 req/day
+# free, 1,000/day after a one-time $10 top-up). No model choice fixes that.
 OPENROUTER_FALLBACK_MODELS = [
-    "meta-llama/llama-3.3-70b-instruct:free",
     "openai/gpt-oss-120b:free",
+    "nvidia/nemotron-3-super-120b-a12b:free",
+    "meta-llama/llama-3.3-70b-instruct:free",
     "qwen/qwen3-next-80b-a3b-instruct:free",
     "google/gemma-4-31b-it:free",
+    "openrouter/free",
 ]
 
 
 def _build_openrouter_client(model: str, api_key: str):
+    # max_retries=0 is critical: the OpenAI SDK's default is 2 retries with
+    # 30s backoff EACH, so a single 429 would waste ~60s per model before
+    # LangChain's .with_fallbacks() could try the next one. With retries off,
+    # the fallback chain moves through models instantly on 429/5xx.
     return ChatOpenAI(
         model=model,
         openai_api_key=api_key,
@@ -177,6 +185,8 @@ def _build_openrouter_client(model: str, api_key: str):
             "X-Title": "Clinical AI Scribe",
         },
         temperature=0.0,
+        max_retries=0,
+        timeout=25,
     )