Allow disabling chat completions persistence via null store

src/ogx/core/routers/__init__.py

@@ -61,15 +61,15 @@ async def get_auto_router_impl(
     # TODO: move pass configs to routers instead
     if api == Api.inference:
         inference_ref = run_config.storage.stores.inference
-        if not inference_ref:
+        if inference_ref is None:
             raise ValueError("storage.stores.inference must be configured in run config")
-
-        inference_store = InferenceStore(
-            reference=inference_ref,
-            policy=policy,
-        )
-        await inference_store.initialize()
-        api_to_dep_impl["store"] = inference_store
+        if inference_ref.enabled:
+            inference_store = InferenceStore(
+                reference=inference_ref,
+                policy=policy,
+            )
+            await inference_store.initialize()
+            api_to_dep_impl["store"] = inference_store
     elif api == Api.vector_io:
         api_to_dep_impl["vector_stores_config"] = run_config.vector_stores
         api_to_dep_impl["inference_api"] = deps.get(Api.inference)

src/ogx/core/storage/README.md

@@ -54,3 +54,18 @@ The tenancy mode is set process-wide during startup via `set_default_tenancy_mod
 Storage is configured in `StackConfig.storage` via `StorageConfig`. The `stores` field contains typed references (`KVStoreReference`, `SqlStoreReference`, `InferenceStoreReference`) that point to specific backend configurations.
 
 See `datatypes.py` for all config types and `StorageBackendType` for the enum of supported backends.
+
+### Inference Store (enabled flag)
+
+Setting `inference.enabled: false` on the inference store reference explicitly
+disables Chat Completions persistence. Omitting the flag (or the whole
+reference) keeps the store enabled for backward compatibility, and setting the
+reference itself to `null` remains a configuration error.
+
+When disabled, no `InferenceStore` is constructed, no `inference_store` table is
+created, and no background write workers run. Streaming and non-streaming Chat
+Completions continue to work. The history endpoints (`list`, `retrieve`, and
+`messages`) report that persistence is not configured (HTTP 501).
+
+Other stores (`responses`, `datasets`, `eval`, `files`, `prompts`, `vector_io`)
+are unaffected by disabling the inference store.

src/ogx/core/storage/datatypes.py

@@ -278,8 +278,8 @@ class KVStoreReference(BaseModel):
 ]
 
 
-class InferenceStoreReference(SqlStoreReference):
-    """Inference store configuration with queue tuning."""
+class _QueuedSqlStoreReference(SqlStoreReference):
+    """Base for SQL store references with background write-queue tuning."""
 
     max_write_queue_size: int = Field(
         default=10000,
@@ -291,7 +291,18 @@ class InferenceStoreReference(SqlStoreReference):
     )
 
 
-class ResponsesStoreReference(InferenceStoreReference):
+class InferenceStoreReference(_QueuedSqlStoreReference):
+    """Inference store configuration with queue tuning."""
+
+    enabled: bool = Field(
+        default=True,
+        description=(
+            "Whether the store is enabled; when false, the store is not constructed and payloads are not persisted"
+        ),
+    )
+
+
+class ResponsesStoreReference(_QueuedSqlStoreReference):
     """Responses store configuration with queue tuning."""
 
     table_name: str = Field(

src/ogx/distributions/README.md

@@ -52,3 +52,27 @@ ogx run starter
 # or
 ogx stack run --config path/to/config.yaml
 ```
+
+## Disabling Chat Completions Persistence
+
+By default the `inference` store reference is enabled, so chat completion
+request/response payloads are persisted and the history endpoints (`list`,
+`retrieve`, `messages`) are served from it. An operator can disable that
+persistence for the inference API by setting `inference.enabled: false` in a run
+config:
+
+```yaml
+storage:
+  stores:
+    inference:
+      enabled: false
+```
+
+When the store is disabled, OGX does not construct an `InferenceStore` at all:
+no `inference_store` table is created and no background write workers run. Chat
+completions still work for both streaming and non-streaming requests; the
+history endpoints report that persistence is not configured (HTTP 501) rather
+than returning an empty list or a 404. Other stores (`responses`, `datasets`,
+`eval`, `files`, `prompts`, `vector_io`) stay enabled, so disabling inference
+persistence is independent of the rest of the storage layer. See the storage
+module README for details.

src/ogx/distributions/ci-tests/config.yaml

@@ -274,6 +274,7 @@ storage:
       backend: sql_default
       max_write_queue_size: 10000
       num_writers: 4
+      enabled: true
     conversations:
       table_name: openai_conversations
       backend: sql_default

src/ogx/distributions/ci-tests/run-with-postgres-store.yaml

@@ -287,6 +287,7 @@ storage:
       backend: sql_default
       max_write_queue_size: 10000
       num_writers: 4
+      enabled: true
     conversations:
       table_name: openai_conversations
       backend: sql_default

src/ogx/distributions/nvidia/config.yaml

@@ -58,6 +58,7 @@ storage:
       backend: sql_default
       max_write_queue_size: 10000
       num_writers: 4
+      enabled: true
     conversations:
       table_name: openai_conversations
       backend: sql_default

src/ogx/distributions/oci/config.yaml

@@ -73,6 +73,7 @@ storage:
       backend: sql_default
       max_write_queue_size: 10000
       num_writers: 4
+      enabled: true
     conversations:
       table_name: openai_conversations
       backend: sql_default

src/ogx/distributions/open-benchmark/config.yaml

@@ -106,6 +106,7 @@ storage:
       backend: sql_default
       max_write_queue_size: 10000
       num_writers: 4
+      enabled: true
     conversations:
       table_name: openai_conversations
       backend: sql_default

src/ogx/distributions/starter/config.yaml

@@ -268,6 +268,7 @@ storage:
       backend: sql_default
       max_write_queue_size: 10000
       num_writers: 4
+      enabled: true
     conversations:
       table_name: openai_conversations
       backend: sql_default

src/ogx/distributions/starter/run-with-postgres-store.yaml

@@ -281,6 +281,7 @@ storage:
       backend: sql_default
       max_write_queue_size: 10000
       num_writers: 4
+      enabled: true
     conversations:
       table_name: openai_conversations
       backend: sql_default

src/ogx/distributions/watsonx/config.yaml

@@ -71,6 +71,7 @@ storage:
       backend: sql_default
       max_write_queue_size: 10000
       num_writers: 4
+      enabled: true
     conversations:
       table_name: openai_conversations
       backend: sql_default

tests/integration/inference/store_disabled_support.py

@@ -0,0 +1,27 @@
+# Copyright (c) The OGX Contributors.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+"""Shared support for inference-store-disabled integration tests."""
+
+from ogx.core.datatypes import StackConfig
+from ogx.core.stack import get_stack_run_config_from_distro
+
+TEXT_MODEL = "openai/gpt-4o"
+
+NON_STREAMING_PROMPT = "Say hello."
+STREAMING_PROMPT = "Say hello in one sentence."
+
+
+def build_inference_store_disabled_run_config() -> StackConfig:
+    """Build the minimal ci-tests configuration used by this test scenario."""
+    run_config = get_stack_run_config_from_distro("ci-tests")
+    # The ci-tests distribution always configures the inference store; disable
+    # persistence through the explicit flag while keeping the reference valid.
+    run_config.storage.stores.inference.enabled = False
+    # Vector-store model validation is unrelated to chat completion persistence
+    # and loads the sentence-transformers stack during an in-process boot.
+    run_config.vector_stores = None
+    return run_config