o

���g��@s�
dZd
dl
Z
d
dlZd
dlZd
dlZd
dlmZmZmZm	Z	m
Z
mZ
d
dlm
Z
d
dlmZ
d
dlmZ
d
dlmZ
d
dlmZ
dedefd	d
�Zdejded
eefdd�Zeddddd�de	ejdedeedeededed
ejfdd��
Z deded
eejfdd�Z!ed�
Z"ed�
Z#dee"d ee#d
ee
e"e#ffd!d"�Z$ed#dd$�ded%e%d&ee&deed
df
d'd(��
Z'dS))zfBeta utility functions to assist in common eval workflows.

These functions may change in the future.
�N)�DefaultDict�List�Optional�Sequence�Tuple�TypeVar)
�
evaluation)
�	warn_beta)
�Client�run_dict�id_mapcCsf|d
}|
��D]\}}|�
t|�
t|�
�}q||d
<|�d�
r(|
|d|d<|�d�
s1i|d<|S)a
Convert the IDs in the run dictionary using the provided ID map.

    Parameters:
    - run_dict (dict): The dictionary representing a run.
    - id_map (dict): The dictionary mapping old IDs to new IDs.

    Returns:
    - dict: The updated run dictionary.
    �dotted_order�
parent_run_id�extra)�items�replace�str�get)rr�do�
k�
v�r�H/home/arjun/.local/lib/python3.10/site-packages/langsmith/beta/_evals.py�_convert_idss









r�root�run_to_example_map�returncs�|g
}t�
�}|j|i
�g}|rJ|��}|jhd
�
d�
}��|dt�
���|d<�|d|d<�|d|d<|jrC|�|j�

|�|�

|s�f
dd�|D�
}|
|j	|dd<|S)	a&
Convert the root run and its child runs to a list of dictionaries.

    Parameters:
    - root (ls_schemas.Run): The root run to convert.
    - run_to_example_map (dict): The dictionary mapping run IDs to example IDs.

    Returns:
    - List[dict]: The list of converted run dictionaries.
    >�
session_id�
child_run_ids�parent_run_ids)
�exclude�id�trace_idc
sg|]}
t|
���qSr)
r��.0�
r�
rrr�
<listcomp>@�z%_convert_root_run.<locals>.<listcomp>r
�reference_example_id)
�uuid�uuid4r"�pop�dictr�
child_runs�extend�appendr!)rr�runs_r"�results�src�src_dict�resultrr&r�_convert_root_run)s"














�	

r6F)�test_project_name�client�load_child_runs�include_outputs�runs�dataset_namer7r8r9r:c
st
|s	td
|���
�
�pt
����j|
d�
}|rdd�|D�
n
d}�jdd�|D�
|dd�|D�
|jd�
�
s9|}n
��
fd	d�|D�
}|pPd
t��jdd���}t	�j
|
d�
�
}	dd
�|	D�
�|	djrj|	djn|	dj}
�f
dd�|D�
}�j
||jd|
��d�d�}|D])}
|
d|
d}tjjtjjd�
|
d<|
d||
d<�jdi|
�
d|i
�
�

q���|j�
}|S)a�Convert the following runs to a dataset + test.

    This makes it easy to sample prod runs into a new regression testing
    workflow and compare against a candidate system.

    Internally, this function does the following:
        1. Create a dataset from the provided production run inputs.
        2. Create a new test project.
        3. Clone the production runs and re-upload against the dataset.

    Parameters:
    - runs (Sequence[ls_schemas.Run]): A sequence of runs to be executed as a test.
    - dataset_name (str): The name of the dataset to associate with the test runs.
    - client (Optional[Client]): An optional LangSmith client instance. If not provided,
        a new client will be created.
    - load_child_runs (bool): Whether to load child runs when copying runs.
        Defaults to False.

    Returns:
    - ls_schemas.TracerSession: The project containing the cloned runs.

    Examples:
    --------
    .. code-block:: python

        import langsmith
        import random

        client = langsmith.Client()

        # Randomly sample 100 runs from a prod project
        runs = list(client.list_runs(project_name="My Project", execution_order=1))
        sampled_runs = random.sample(runs, min(len(runs), 100))

        runs_as_test(runs, dataset_name="Random Runs")

        # Select runs named "extractor" whose root traces received good feedback
        runs = client.list_runs(
            project_name="<your_project>",
            filter='eq(name, "extractor")',
            trace_filter='and(eq(feedback_key, "user_score"), eq(feedback_score, 1))',
        )
        runs_as_test(runs, dataset_name="Extraction Good")
    z1Expected a non-empty sequence of runs. Received: )
r<c
S�g|]}
|
j�qSr)
�outputsr#rrrr'�z(convert_runs_to_test.<locals>.<listcomp>Nc
Sr=r)
�inputsr#rrrr'�r?c
Sr=r)
r!r#rrrr'�r?)r@r>�source_run_ids�
dataset_idc
sg|]
}
�j|
j
�
d��qS)
)
r9)�read_runr!r#)r8r9rrr'�s
�zprod-baseline-�c
Ssi|]}
|
j|
j
�qSr)�
source_run_idr!)r$�
errr�
<dictcomp>�r(z(convert_runs_to_test.<locals>.<dictcomp>r
c
s g|]}
t|
��D]}|�q	qSr)
r6)r$�root_runr)
rrrr'�s
���z
prod-baseline)�which�dataset_version)�project_name�reference_dataset_id�metadata�end_time�
start_time)
�tzrKr)�
ValueError�rt�get_cached_client�create_dataset�create_examplesr!r*r+�hex�list�
list_examples�modified_at�
created_at�create_project�	isoformat�datetime�now�timezone�utc�
create_run�update_project)r;r<r7r8r9r:�dsr>�runs_to_copy�examplesrJ�	to_create�project�new_run�latency�
_r)r8r9rr�convert_runs_to_testEsP6








�

�
�
�


��	




�rkrKc	Cs�|
j|d
�
}t
�t�
}g}i}|D]}|jdu
r!||j�|�

n|�|�

|||j<q|��D]\}}t|dd�d�||_	q0|S)N)
rKc


Ss|jS�
N)
r
)
r%rrr�<lambda>�sz%_load_nested_traces.<locals>.<lambda>)
�key)
�	list_runs�collections�defaultdictrWrr0r!r�sortedr.)	rKr8r;�treemapr2�all_runs�run�run_idr.rrr�_load_nested_traces�s


�









rw�
T�
U�list1�list2cCstt
�||
��
Srl)rW�	itertools�product)rzr{rrr�_outer_product�s
r~�
)�max_concurrencyr8�
evaluatorsr�c
Cs�d
dlm
}
g}|
D]#}t|tj�r|�|�

q
t|�
r%|�t�|�
�

q
tdt	|�
���
�
|p3t
��}t||�}||d�
�}|j
|jg
tt||���
R�}	Wd�
n1sXw



Y
|	D]}
q_dS)aCompute test metrics for a given test name using a list of evaluators.

    Args:
        project_name (str): The name of the test project to evaluate.
        evaluators (list): A list of evaluators to compute metrics with.
        max_concurrency (Optional[int], optional): The maximum number of concurrent
            evaluations. Defaults to 10.
        client (Optional[Client], optional): The client to use for evaluations.
            Defaults to None.

    Returns:
        None: This function does not return any value.
    r
)
�ContextThreadPoolExecutorz5Evaluation not yet implemented for evaluator of type )
�max_workersN)�	langsmithr��
isinstance�ls_eval�RunEvaluatorr0�callable�
run_evaluator�NotImplementedError�typerRrSrw�map�evaluate_run�zipr~)rKr�r�r8r��evaluators_�func�traces�executorr2rjrrr�compute_test_metrics�s,





�




�

��
�r�)(�__doc__rpr]r|r*�typingrrrrrr�langsmith.run_trees�	run_treesrR�langsmith.schemas�schemas�
ls_schemasr�rr��#langsmith._internal._beta_decoratorr	�langsmith.clientr
r-r�Runr6r�bool�
TracerSessionrkrwrxryr~rW�intr�rrrr�<module>
sf


 







�
�������k
*
�
�����