tracecast 0.2.7__tar.gz → 0.2.9__tar.gz

{tracecast-0.2.7/tracecast.egg-info → tracecast-0.2.9}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: tracecast
-Version: 0.2.7
+Version: 0.2.9
 Summary: LLM observability SDK — framework-agnostic tracing for AI applications
 Author: Pedro Castanheira Costa
 License: MIT

{tracecast-0.2.7 → tracecast-0.2.9}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "tracecast"
-version = "0.2.7"
+version = "0.2.9"
 description = "LLM observability SDK — framework-agnostic tracing for AI applications"
 readme = "README.md"
 license = { text = "MIT" }

{tracecast-0.2.7 → tracecast-0.2.9}/tracecast/__init__.py

@@ -8,7 +8,7 @@ from .decorators import trace_cast, trace_span, set_default_tracer
 from .core.scoring import score
 from .prompts import create_prompt, get_prompt, set_label, PromptVersion
 from .integrations.llm import trace_llm_call, wrap_openai, wrap_anthropic
-from .instrument import auto_instrument
+from .instrument import auto_instrument, instrument_openai
 from .exporters.dict_exporter import DictExporter
 
 __all__ = [
@@ -18,5 +18,5 @@ __all__ = [
     "trace_cast", "trace_span", "set_default_tracer", "score",
     "create_prompt", "get_prompt", "set_label", "PromptVersion",
     "trace_llm_call", "wrap_openai", "wrap_anthropic",
-    "auto_instrument", "DictExporter",
+    "auto_instrument", "instrument_openai", "DictExporter",
 ]

{tracecast-0.2.7 → tracecast-0.2.9}/tracecast/instrument.py

@@ -34,6 +34,24 @@ def auto_instrument(tracer: Optional[Tracer] = None) -> None:
         _instrumented = True
 
 
+def instrument_openai(*, chat: bool = True, embeddings: bool = True, audio: bool = True) -> None:
+    """Patch only the requested OpenAI API surfaces, independent of auto_instrument().
+
+    Use `chat=False` when chat completions are already traced through another
+    mechanism (e.g. a manually-wired LangChain callback) and you only want to
+    add embeddings/audio coverage — re-patching chat.completions in that case
+    would produce duplicate spans for every chat call.
+    """
+    from .instrumentors.openai_inst import OpenAIInstrumentor
+    with _lock:
+        inst = _registry.get("openai")
+        if inst is None:
+            inst = OpenAIInstrumentor()
+            _registry["openai"] = inst
+        if not inst.is_patched():
+            inst.patch(chat=chat, embeddings=embeddings, audio=audio)
+
+
 def _reset() -> None:
     global _instrumented
     with _lock:

{tracecast-0.2.7 → tracecast-0.2.9}/tracecast/instrumentors/openai_inst.py

@@ -5,6 +5,22 @@ from typing import Any, Optional
 from .base import BaseInstrumentor, active_parent_id
 
 
+# Only LangChain's own LLM-wrapper packages directly invoke the underlying
+# provider client on their own behalf, triggering LangChain's callback system
+# (which TraceCastCallback hooks into via LangChainInstrumentor). Orchestration
+# packages (langchain_core.tools, langchain.agents, langgraph.*) run *above*
+# a raw client call made by application code (e.g. a @tool function) — that
+# call is invisible to LangChain's callbacks, so it must NOT be deferred here,
+# or it silently disappears from tracing entirely.
+_LANGCHAIN_LLM_WRAPPER_PREFIXES = (
+    "langchain_openai.",
+    "langchain_community.chat_models",
+    "langchain_community.llms",
+    "langchain.chat_models",
+    "langchain.llms",
+)
+
+
 def _handled_by_langchain() -> bool:
     from ..instrument import _registry
 
@@ -15,7 +31,7 @@ def _handled_by_langchain() -> bool:
     depth = 0
     while frame is not None and depth < 60:
         module = frame.f_globals.get("__name__", "")
-        if module.startswith("langchain"):
+        if module.startswith(_LANGCHAIN_LLM_WRAPPER_PREFIXES):
             return True
         frame = frame.f_back
         depth += 1
@@ -32,33 +48,44 @@ class OpenAIInstrumentor(BaseInstrumentor):
         self._original_speech_create: Optional[Any] = None
         self._patched: bool = False
 
-    def patch(self) -> None:
+    def patch(self, *, chat: bool = True, embeddings: bool = True, audio: bool = True) -> None:
+        """Patch the requested OpenAI API surfaces.
+
+        `chat=False` lets callers add embeddings/audio coverage without
+        re-patching chat.completions — needed when chat completions are
+        already traced through another mechanism (e.g. a manually-wired
+        LangChain callback), where re-patching would double-count spans.
+        """
         if self._patched:
             return
-        import openai.resources.chat.completions as mod
 
-        # Sync
-        self._original_create = mod.Completions.create
-        self_ref = self
-        _orig_create = self._original_create
+        if chat:
+            import openai.resources.chat.completions as mod
+
+            # Sync
+            self._original_create = mod.Completions.create
+            self_ref = self
+            _orig_create = self._original_create
 
-        def patched_create(client_self, *args, **kwargs):
-            return self_ref._intercept(client_self, args, kwargs, _orig_create)
+            def patched_create(client_self, *args, **kwargs):
+                return self_ref._intercept(client_self, args, kwargs, _orig_create)
 
-        mod.Completions.create = patched_create
+            mod.Completions.create = patched_create
 
-        # Async
-        if hasattr(mod, "AsyncCompletions"):
-            self._original_acreate = mod.AsyncCompletions.create
-            _orig_acreate = self._original_acreate
+            # Async
+            if hasattr(mod, "AsyncCompletions"):
+                self._original_acreate = mod.AsyncCompletions.create
+                _orig_acreate = self._original_acreate
 
-            async def patched_acreate(client_self, *args, **kwargs):
-                return await self_ref._async_intercept(client_self, args, kwargs, _orig_acreate)
+                async def patched_acreate(client_self, *args, **kwargs):
+                    return await self_ref._async_intercept(client_self, args, kwargs, _orig_acreate)
 
-            mod.AsyncCompletions.create = patched_acreate
+                mod.AsyncCompletions.create = patched_acreate
 
-        self._patch_embeddings()
-        self._patch_audio()
+        if embeddings:
+            self._patch_embeddings()
+        if audio:
+            self._patch_audio()
 
         self._patched = True
 
@@ -108,12 +135,13 @@ class OpenAIInstrumentor(BaseInstrumentor):
     def unpatch(self) -> None:
         if not self._patched:
             return
-        import openai.resources.chat.completions as mod
-        mod.Completions.create = self._original_create
-        if self._original_acreate is not None:
-            mod.AsyncCompletions.create = self._original_acreate
-        self._original_create = None
-        self._original_acreate = None
+        if self._original_create is not None:
+            import openai.resources.chat.completions as mod
+            mod.Completions.create = self._original_create
+            if self._original_acreate is not None:
+                mod.AsyncCompletions.create = self._original_acreate
+            self._original_create = None
+            self._original_acreate = None
 
         if self._original_embeddings_create is not None:
             import openai.resources.embeddings as emb_mod

{tracecast-0.2.7 → tracecast-0.2.9/tracecast.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: tracecast
-Version: 0.2.7
+Version: 0.2.9
 Summary: LLM observability SDK — framework-agnostic tracing for AI applications
 Author: Pedro Castanheira Costa
 License: MIT