Merge branch 'main' into bagatur/rfc_set_test_vals

js/src/run_trees.ts

@@ -9,6 +9,7 @@ import {
 import {
   RuntimeEnvironment,
   getEnvironmentVariable,
+  getLangSmithEnvironmentVariable,
   getRuntimeEnvironment,
 } from "./utils/env.js";
 import { Client } from "./client.js";
@@ -231,7 +232,7 @@ export class RunTree implements BaseRun {
       id: uuid.v4(),
       run_type: "chain",
       project_name:
-        getEnvironmentVariable("LANGCHAIN_PROJECT") ??
+        getLangSmithEnvironmentVariable("PROJECT") ??
         getEnvironmentVariable("LANGCHAIN_SESSION") ?? // TODO: Deprecate
         "default",
       child_runs: [],

python/langsmith/_internal/_serde.py

@@ -146,7 +146,7 @@ def dumps_json(obj: Any) -> bytes:
         logger.debug(f"Orjson serialization failed: {repr(e)}. Falling back to json.")
         result = json.dumps(
             obj,
-            default=_simple_default,
+            default=_serialize_json,
             ensure_ascii=True,
         ).encode("utf-8")
         try:

python/langsmith/schemas.py

@@ -523,6 +523,10 @@ class FeedbackSourceBase(BaseModel):
     """The type of the feedback source."""
     metadata: Optional[Dict[str, Any]] = Field(default_factory=dict)
     """Additional metadata for the feedback source."""
+    user_id: Optional[Union[UUID, str]] = None
+    """The user ID associated with the feedback source."""
+    user_name: Optional[str] = None
+    """The user name associated with the feedback source."""
 
 
 class APIFeedbackSource(FeedbackSourceBase):

python/langsmith/wrappers/_openai.py

@@ -3,14 +3,14 @@
 import functools
 import logging
 from collections import defaultdict
+from collections.abc import Mapping
 from typing import (
     TYPE_CHECKING,
     Any,
     Callable,
     DefaultDict,
     Dict,
     List,
-    Mapping,
     Optional,
     Type,
     TypeVar,
@@ -82,23 +82,28 @@ def _reduce_choices(choices: List[Choice]) -> dict:
         "content": "",
     }
     for c in reversed_choices:
-        if c.delta.role:
+        if hasattr(c, "delta") and getattr(c.delta, "role", None):
             message["role"] = c.delta.role
             break
     tool_calls: DefaultDict[int, List[ChoiceDeltaToolCall]] = defaultdict(list)
     for c in choices:
-        if c.delta.content:
-            message["content"] += c.delta.content
-        if c.delta.function_call:
-            if not message.get("function_call"):
-                message["function_call"] = {"name": "", "arguments": ""}
-            if c.delta.function_call.name:
-                message["function_call"]["name"] += c.delta.function_call.name
-            if c.delta.function_call.arguments:
-                message["function_call"]["arguments"] += c.delta.function_call.arguments
-        if c.delta.tool_calls:
-            for tool_call in c.delta.tool_calls:
-                tool_calls[c.index].append(tool_call)
+        if hasattr(c, "delta") and getattr(c.delta, "content", None):
+            if getattr(c.delta, "content", None):
+                message["content"] += c.delta.content
+            if getattr(c.delta, "function_call", None):
+                if not message.get("function_call"):
+                    message["function_call"] = {"name": "", "arguments": ""}
+                name_ = getattr(c.delta.function_call, "name", None)
+                if name_:
+                    message["function_call"]["name"] += name_
+                arguments_ = getattr(c.delta.function_call, "arguments", None)
+                if arguments_:
+                    message["function_call"]["arguments"] += arguments_
+            if getattr(c.delta, "tool_calls", None):
+                tool_calls_list = c.delta.tool_calls
+                if tool_calls_list is not None:
+                    for tool_call in tool_calls_list:
+                        tool_calls[c.index].append(tool_call)
     if tool_calls:
         message["tool_calls"] = [None for _ in tool_calls.keys()]
         for index, tool_call_chunks in tool_calls.items():
@@ -108,22 +113,28 @@ def _reduce_choices(choices: List[Choice]) -> dict:
                 "type": next((c.type for c in tool_call_chunks if c.type), None),
             }
             for chunk in tool_call_chunks:
-                if chunk.function:
+                if getattr(chunk, "function", None):
                     if not message["tool_calls"][index].get("function"):
                         message["tool_calls"][index]["function"] = {
                             "name": "",
                             "arguments": "",
                         }
-                    if chunk.function.name:
+                    name_ = getattr(chunk.function, "name", None)
+                    if name_:
                         fn_ = message["tool_calls"][index]["function"]
-                        fn_["name"] += chunk.function.name
-                    if chunk.function.arguments:
+                        fn_["name"] += name_
+                    arguments_ = getattr(chunk.function, "arguments", None)
+                    if arguments_:
                         fn_ = message["tool_calls"][index]["function"]
-                        fn_["arguments"] += chunk.function.arguments
+                        fn_["arguments"] += arguments_
     return {
-        "index": choices[0].index,
+        "index": getattr(choices[0], "index", 0) if choices else 0,
         "finish_reason": next(
-            (c.finish_reason for c in reversed_choices if c.finish_reason),
+            (
+                c.finish_reason
+                for c in reversed_choices
+                if getattr(c, "finish_reason", None)
+            ),
             None,
         ),
         "message": message,

python/tests/integration_tests/test_client.py

@@ -19,6 +19,7 @@
 from pydantic import BaseModel
 from requests_toolbelt import MultipartEncoder, MultipartEncoderMonitor
 
+from langsmith._internal._serde import dumps_json
 from langsmith.client import ID_TYPE, Client
 from langsmith.evaluation import aevaluate, evaluate
 from langsmith.schemas import (
@@ -1155,6 +1156,37 @@ def test_surrogates():
     )
 
 
+def test_fallback_json_serialization():
+    class Document(BaseModel):
+        content: str
+
+    raw_surrogates = [
+        ("Hello\ud83d\ude00", "Hello😀"),
+        ("Python\ud83d\udc0d", "Python🐍"),
+        ("Surrogate\ud834\udd1e", "Surrogate𝄞"),
+        ("Example\ud83c\udf89", "Example🎉"),
+        ("String\ud83c\udfa7", "String🎧"),
+        ("With\ud83c\udf08", "With🌈"),
+        ("Surrogates\ud83d\ude0e", "Surrogates😎"),
+        ("Embedded\ud83d\udcbb", "Embedded💻"),
+        ("In\ud83c\udf0e", "In🌎"),
+        ("The\ud83d\udcd6", "The📖"),
+        ("Text\ud83d\udcac", "Text💬"),
+        ("收花🙄·到", "收花🙄·到"),
+    ]
+    pydantic_surrogates = [
+        (Document(content=item), expected) for item, expected in raw_surrogates
+    ]
+
+    for item, expected in raw_surrogates:
+        output = dumps_json(item).decode("utf8")
+        assert f'"{expected}"' == output
+
+    for item, expected in pydantic_surrogates:
+        output = dumps_json(item).decode("utf8")
+        assert f'{{"content":"{expected}"}}' == output
+
+
 def test_runs_stats():
     langchain_client = Client()
     # We always have stuff in the "default" project...

python/tests/integration_tests/wrappers/test_openai.py

@@ -381,13 +381,13 @@ def test_parse_sync_api():
     original_client = openai.Client()
     patched_client = wrap_openai(openai.Client(), tracing_extra={"client": ls_client})
 
-    messages = [{"role": "user", "content": "Say 'Foo' then stop."}]
+    messages = [{"role": "user", "content": "Say 'foo' then stop."}]
 
     original = original_client.beta.chat.completions.parse(
-        messages=messages, model="gpt-3.5-turbo"
+        messages=messages, model="gpt-3.5-turbo", temperature=0, seed=42, max_tokens=3
     )
     patched = patched_client.beta.chat.completions.parse(
-        messages=messages, model="gpt-3.5-turbo"
+        messages=messages, model="gpt-3.5-turbo", temperature=0, seed=42, max_tokens=3
     )
 
     assert type(original) is type(patched)
@@ -413,13 +413,13 @@ async def test_parse_async_api():
         openai.AsyncClient(), tracing_extra={"client": ls_client}
     )
 
-    messages = [{"role": "user", "content": "Say 'Foo' then stop."}]
+    messages = [{"role": "user", "content": "Say 'foo' then stop."}]
 
     original = await original_client.beta.chat.completions.parse(
-        messages=messages, model="gpt-3.5-turbo"
+        messages=messages, model="gpt-3.5-turbo", temperature=0, seed=42, max_tokens=3
     )
     patched = await patched_client.beta.chat.completions.parse(
-        messages=messages, model="gpt-3.5-turbo"
+        messages=messages, model="gpt-3.5-turbo", temperature=0, seed=42, max_tokens=3
     )
 
     assert type(original) is type(patched)