diff --git a/.defang/preview b/.defang/preview
new file mode 100644
index 0000000..699cdf2
--- /dev/null
+++ b/.defang/preview
@@ -0,0 +1,20 @@
+# Throwaway Azure stack for previewing changes. Deploy with:
+#   defang compose up -s preview -f compose.preview.yaml
+#
+# No AZURE_CLIENT_ID: that is the CI user-assigned identity. Run locally and the
+# Azure CLI login is used instead.
+#
+# DEFANG_MODE=affordable, not balanced: balanced is gated by subscription tier
+# ("this recipe is not available for your tier"), and a preview wants the cheap
+# shape anyway.
+#
+# DEFANG_TTL so a forgotten preview self-destructs instead of running forever.
+# No effect yet: the feature needs a pulumi-defang release past PR 536, and
+# releases are blocked on DefangLabs/pulumi-defang#414 (RELEASE_PAT). Harmless
+# to set now — it starts working the day that release ships.
+AZURE_LOCATION="westus"
+AZURE_SUBSCRIPTION_ID="f311c4db-e998-4c94-906c-7e2637303a05"
+AZURE_TENANT_ID="12c0f515-fa47-402f-9982-de7646d3cb28"
+DEFANG_PROVIDER=azure
+DEFANG_MODE=affordable
+DEFANG_TTL=1d
diff --git a/app/rag_system.py b/app/rag_system.py
index f46f767..318d935 100644
--- a/app/rag_system.py
+++ b/app/rag_system.py
@@ -282,7 +282,10 @@ def answer_query_stream(self, query):
                 model=os.getenv("MODEL"),
                 messages=messages,
                 temperature=0.25,
-                max_tokens=2048,
+                # GPT-5.x rejects max_tokens ("Unsupported parameter: 'max_tokens' is
+                # not supported with this model. Use 'max_completion_tokens' instead.").
+                # LiteLLM runs with --drop_params, so this stays safe on Bedrock too.
+                max_completion_tokens=2048,
                 # Claude 4.5+ on Bedrock rejects temperature and top_p together
                 # ("`temperature` and `top_p` cannot both be specified for this
                 # model"), so send only temperature. top_p=1 was a no-op anyway.
@@ -293,6 +296,12 @@ def answer_query_stream(self, query):
             for chunk in stream:
                 try:
                     logging.debug(f"Received chunk: {chunk}")
+                    # Azure OpenAI opens the stream with a chunk carrying only
+                    # prompt_filter_results and an empty "choices" list, and emits
+                    # further choice-less chunks for content filtering. Indexing
+                    # [0] on those raises IndexError and aborts the answer.
+                    if not chunk["choices"]:
+                        continue
                     content = chunk["choices"][0]["delta"].get("content", "")
                     collected_messages.append(content)
                     yield content
diff --git a/compose.preview.yaml b/compose.preview.yaml
new file mode 100644
index 0000000..50c64e1
--- /dev/null
+++ b/compose.preview.yaml
@@ -0,0 +1,34 @@
+# Preview overlay: a throwaway Azure stack for testing changes before they reach
+# production. Deploy with:
+#
+#   defang compose up -s <name> -f compose.preview.yaml
+#
+# It deliberately does NOT include compose.yaml wholesale. `extends` pulls in only
+# the services a preview should run, which is how discord-bot is left out: it holds
+# the live Discord token and a second copy would answer real users alongside
+# production. Compose overlays can add and change keys but cannot delete a service,
+# so selecting services here is the only way to drop one.
+services:
+  app:
+    extends:
+      file: compose.yaml
+      service: app
+    # Never claim the production hostname. Azure's DNS writer only touches Azure
+    # zones, so it would silently skip this anyway, but a preview must not race
+    # production for the certificate either. Empty string clears the inherited value.
+    domainname: ""
+
+  redis:
+    extends:
+      file: compose.yaml
+      service: redis
+
+models:
+  # On Azure the AI Foundry deployment is named after the SERVICE, while Defang
+  # wires the model ALIAS into MODEL — so they must match or every request fails
+  # with DeploymentNotFound. Once pulumi-defang#536 ships (the provider then reads
+  # the alias from the service's own --alias) this can become "chat-default", which
+  # is portable across clouds. The alias only picks chat vs embedding, so "llm" is
+  # a valid chat model either way.
+  llm:
+    model: llm
diff --git a/compose.yaml b/compose.yaml
index 9fbf1c4..8c0e791 100644
--- a/compose.yaml
+++ b/compose.yaml
@@ -36,8 +36,12 @@ services:
       test: ["CMD", "curl", "-f", "http://localhost:5050/"]
       interval: 30s
       timeout: 10s
-      retries: 5
-      start_period: 240s
+      # Azure Container Apps caps a probe's InitialDelaySeconds at 60 and Defang maps
+      # start_period onto it, so 240s fails the deploy outright. Keep the startup
+      # budget inside that limit (60 + 10*30 = 360s); ECS treats start_period as a
+      # grace window, so this stays equivalent there.
+      retries: 10
+      start_period: 60s
     depends_on:
       - redis
     models: