fix: improve perf

private_gpt/server/chat/interceptors/deduplicate_event_interceptor.py

@@ -17,16 +17,20 @@ class DeduplicateEventInterceptor(ChatResponseLoopInterceptor):
     ``input_json_delta`` with an unchanged ``partial_json_obj``). Those
     duplicates add noise for clients without changing state, so they are
     suppressed. Ping keepalives are always forwarded.
+
+    Equality is checked via pydantic's structural ``__eq__`` (field-by-field
+    comparison) rather than ``model_dump_json``, since serializing every
+    streamed event just to compare it adds significant latency.
     """
 
     def __init__(self) -> None:
-        self._last_fingerprint: str | None = None
+        self._last_event: Event | None = None
 
     async def on_iteration_start(self, context: ChatInterceptorContext) -> None:
-        self._last_fingerprint = None
+        self._last_event = None
 
     async def on_iteration_end(self, context: ChatInterceptorContext) -> None:
-        self._last_fingerprint = None
+        self._last_event = None
 
     async def intercept_event(
         self,
@@ -36,11 +40,10 @@ class DeduplicateEventInterceptor(ChatResponseLoopInterceptor):
         if isinstance(event, PingEvent):
             return event
 
-        fingerprint = event.model_dump_json()
-        if fingerprint == self._last_fingerprint:
+        if event == self._last_event:
             return None
 
-        self._last_fingerprint = fingerprint
+        self._last_event = event
         return event
 
     def model_copy(

tests/sse/test_sse.py

@@ -681,6 +681,42 @@ async def test_unit_duplicate_content_block_deltas_are_dropped() -> None:
     assert deltas[1].delta.partial_json_obj == {"object": "deals", "query": "x"}
 
 
+def test_unit_content_block_delta_equality_ignores_object_identity() -> None:
+    """Two distinct RawContentBlockDeltaEvent objects compare equal when their
+    field values match, and unequal when they differ -- this is what
+    DeduplicateEventInterceptor relies on instead of model_dump_json()."""
+    from private_gpt.events.models import InputJSONDelta
+
+    bid = _block_id()
+
+    equal_a = RawContentBlockDeltaEvent(
+        index=1,
+        block_id=bid,
+        delta=InputJSONDelta(
+            partial_json="", partial_json_obj={"object": "deals", "query": None}
+        ),
+    )
+    equal_b = RawContentBlockDeltaEvent(
+        index=1,
+        block_id=bid,
+        delta=InputJSONDelta(
+            partial_json="", partial_json_obj={"object": "deals", "query": None}
+        ),
+    )
+    assert equal_a is not equal_b
+    assert equal_a == equal_b
+
+    different = RawContentBlockDeltaEvent(
+        index=1,
+        block_id=bid,
+        delta=InputJSONDelta(
+            partial_json='{"object":"deals","query":"x"}',
+            partial_json_obj={"object": "deals", "query": "x"},
+        ),
+    )
+    assert equal_a != different
+
+
 @pytest.mark.asyncio
 async def test_unit_ping_events_are_never_deduplicated() -> None:
     """Ping keepalives always pass through even when consecutive."""