language-model-common 2.1.2__tar.gz → 2.1.3__tar.gz

{language_model_common-2.1.2/language_model_common.egg-info → language_model_common-2.1.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: language-model-common
-Version: 2.1.2
+Version: 2.1.3
 Summary: Provides the underlying framework to enhance langchain and add loading configurations
 Author-email: Imran Qureshi <imran.qureshi@bwell.com>
 License: Apache-2.0

language_model_common-2.1.3/VERSION

@@ -0,0 +1 @@
+2.1.3

{language_model_common-2.1.2 → language_model_common-2.1.3/language_model_common.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: language-model-common
-Version: 2.1.2
+Version: 2.1.3
 Summary: Provides the underlying framework to enhance langchain and add loading configurations
 Author-email: Imran Qureshi <imran.qureshi@bwell.com>
 License: Apache-2.0

{language_model_common-2.1.2 → language_model_common-2.1.3}/languagemodelcommon/converters/stream_buffer.py

@@ -18,6 +18,29 @@ class StreamBufferManager:
         self._flush_interval_seconds = flush_interval_seconds
         self._enabled = enabled
         self._buffer = StreamBuffer(chunks=[], last_flush_ts=time.monotonic())
+        self._last_emitted_char: str | None = None
+        self._pending_invocation_boundary = False
+
+    def mark_new_invocation_boundary(self) -> None:
+        """Flag that the next content flows from a fresh chat-model invocation.
+
+        LangGraph re-invokes the chat model independently after each tool
+        call, so the resumed completion has no guaranteed leading whitespace
+        against whatever this request already streamed. The flag is consumed
+        by the next non-empty ``content_text`` passed to ``buffer_content``.
+        """
+        self._pending_invocation_boundary = True
+
+    def _apply_invocation_boundary(self, *, content_text: str) -> str:
+        if not self._pending_invocation_boundary:
+            return content_text
+        self._pending_invocation_boundary = False
+        needs_separator = (
+            self._last_emitted_char is not None
+            and not self._last_emitted_char.isspace()
+            and not content_text[0].isspace()
+        )
+        return f" {content_text}" if needs_separator else content_text
 
     async def buffer_content(
         self,
@@ -25,6 +48,10 @@ class StreamBufferManager:
         content_text: str,
         force_flush: bool = False,
     ) -> str | None:
+        if content_text:
+            content_text = self._apply_invocation_boundary(content_text=content_text)
+            self._last_emitted_char = content_text[-1]
+
         if not self._enabled:
             existing_text = "".join(self._buffer.chunks) if self._buffer.chunks else ""
             self._buffer.chunks.clear()

{language_model_common-2.1.2 → language_model_common-2.1.3}/languagemodelcommon/converters/streaming_manager.py

@@ -293,6 +293,11 @@ class LangGraphStreamingManager(StreamContextMixin):
         chat_request_wrapper: ChatRequestWrapper,
         request_information: RequestInformation,
     ) -> AsyncGenerator[str | None, None]:
+        # Each on_chat_model_start after the first marks a fresh model
+        # invocation (e.g. resuming after a tool call) whose first token has
+        # no guaranteed whitespace against what was already streamed -- see
+        # StreamBufferManager.mark_new_invocation_boundary.
+        self._stream_buffer_manager.mark_new_invocation_boundary()
         yield None
 
     async def _handle_on_chat_model_end(

language_model_common-2.1.2/VERSION

@@ -1 +0,0 @@
-2.1.2