Skip to main content

Export traces

Recommended Reading

Before diving into this content, it might be helpful to read the following:

The recommended way to export runs (the span data in LangSmith traces) is to use the list_runs method in the SDK or /runs/query endpoint in the API.

LangSmith stores traces in a simple format that is specified in the Run (span) data format.

Use filter arguments​

For simple queries, you don't have to rely on our query syntax. You can use the filter arguments specified in the filter arguments reference.

Prerequisites

Initialize the client before running the below code snippets.

from langsmith import Client

client = Client()

Below are some examples of ways to list runs using keyword arguments:

List all runs in a project​

project_runs = client.list_runs(project_name="<your_project>")

List LLM and Chat runs in the last 24 hours​

todays_llm_runs = client.list_runs(
    project_name="<your_project>",
    start_time=datetime.now() - timedelta(days=1),
    run_type="llm",
)

List root runs in a project​

Root runs are runs that have no parents. These are assigned a value of True for is_root. You can use this to filter for root runs.

root_runs = client.list_runs(
    project_name="<your_project>",
    is_root=True
)

List runs without errors​

correct_runs = client.list_runs(project_name="<your_project>", error=False)

List runs by run ID​

Ignores Other Arguments

If you provide a list of run IDs in the way described above, it will ignore all other filtering arguments like project_name, run_type, etc. and directly return the runs matching the given IDs.

If you have a list of run IDs, you can list them directly:

run_ids = ['a36092d2-4ad5-4fb4-9c0d-0dba9a2ed836','9398e6be-964f-4aa4-8ae9-ad78cd4b7074']
selected_runs = client.list_runs(id=run_ids)

Use filter query language​

For more complex queries, you can use the query language described in the filter query language reference.

List all root runs in a conversational thread​

This is the way to fetch runs in a conversational thread. For more information on setting up threads, refer to our how-to guide on setting up threads. Threads are grouped by setting a shared thread ID. The LangSmith UI lets you use any one of the following three metadata keys: session_id, conversation_id, or thread_id. The following query matches on any of them.

group_key = "<your_thread_id>"
filter_string = f'and(in(metadata_key, ["session_id","conversation_id","thread_id"]), eq(metadata_value, "{group_key}"))'
thread_runs = client.list_runs(
    project_name="<your_project>",
    filter=filter_string,
    is_root=True
)

List all runs called "extractor" whose root of the trace was assigned feedback "user_score" score of 1​

client.list_runs(
    project_name="<your_project>",
    filter='eq(name, "extractor")',
    trace_filter='and(eq(feedback_key, "user_score"), eq(feedback_score, 1))'
)

List runs with "star_rating" key whose score is greater than 4​

client.list_runs(
    project_name="<your_project>",
    filter='and(eq(feedback_key, "star_rating"), gt(feedback_score, 4))'
)

List runs that took longer than 5 seconds to complete​

client.list_runs(project_name="<your_project>", filter='gt(latency, "5s")')

List all runs where total_tokens is greater than 5000​

client.list_runs(project_name="<your_project>", filter='gt(total_tokens, 5000)')

List all runs that have "error" not equal to null​

client.list_runs(project_name="<your_project>", filter='neq(error, null)')

List all runs where start_time is greater than a specific timestamp​

client.list_runs(project_name="<your_project>", filter='gt(start_time, "2023-07-15T12:34:56Z")')

List all runs that contain the string "substring"​

client.list_runs(project_name="<your_project>", filter='search("substring")')

List all runs that are tagged with the git hash "2aa1cf4"​

client.list_runs(project_name="<your_project>", filter='has(tags, "2aa1cf4")')

List all "chain" type runs that took more than 10 seconds and​

had total_tokens greater than 5000

client.list_runs(
  project_name="<your_project>",
  filter='and(eq(run_type, "chain"), gt(latency, 10), gt(total_tokens, 5000))'
)

List all runs that started after a specific timestamp and either​

have "error" not equal to null or a "Correctness" feedback score equal to 0

client.list_runs(
  project_name="<your_project>",
  filter='and(gt(start_time, "2023-07-15T12:34:56Z"), or(neq(error, null), and(eq(feedback_key, "Correctness"), eq(feedback_score, 0.0))))'
)

Complex query: List all runs where tags include "experimental" or "beta" and​

latency is greater than 2 seconds

client.list_runs(
  project_name="<your_project>",
  filter='and(or(has(tags, "experimental"), has(tags, "beta")), gt(latency, 2))'
)

Search trace trees by full text You can use the search() function without​

any specific field to do a full text search across all string fields in a run. This allows you to quickly find traces that match a search term.

client.list_runs(
  project_name="<your_project>",
  filter='search("image classification")'
)

Check for presence of metadata​

If you want to check for the presence of metadata, you can use the eq operator, optionally with an and statement to match by value. This is useful if you want to log more structured information about your runs.


to_search = {
    "user_id": ""
}

# Check for any run with the "user_id" metadata key
client.list_runs(
  project_name="default",
  filter="eq(metadata_key, 'user_id')"
)
# Check for runs with user_id=4070f233-f61e-44eb-bff1-da3c163895a3
client.list_runs(
  project_name="default",
  filter="and(eq(metadata_key, 'user_id'), eq(metadata_value, '4070f233-f61e-44eb-bff1-da3c163895a3'))"
)

Check for environment details in metadata.​

A common pattern is to add environment information to your traces via metadata. If you want to filter for runs containing environment metadata, you can use the same pattern as above:

client.list_runs(
  project_name="default",
  filter="and(eq(metadata_key, 'environment'), eq(metadata_value, 'production'))"
)

Check for conversation ID in metadata​

Another common way to associate traces in the same conversation is by using a shared conversation ID. If you want to filter runs based on a conversation ID in this way, you can search for that ID in the metadata.

client.list_runs(
  project_name="default",
  filter="and(eq(metadata_key, 'conversation_id'), eq(metadata_value, 'a1b2c3d4-e5f6-7890'))"
)

Combine multiple filters​

If you want to combine multiple conditions to refine your search, you can use the and operator along with other filtering functions. Here's how you can search for runs named "ChatOpenAI" that also have a specific conversation_id in their metadata:

client.list_runs(
  project_name="default",
  filter="and(eq(name, 'ChatOpenAI'), eq(metadata_key, 'conversation_id'), eq(metadata_value, '69b12c91-b1e2-46ce-91de-794c077e8151'))"
)

Tree Filter​

List all runs named "RetrieveDocs" whose root run has a "user_score" feedback of 1 and any run in the full trace is named "ExpandQuery".

This type of query is useful if you want to extract a specific run conditional on various states or steps being reached within the trace.

client.list_runs(
    project_name="<your_project>",
    filter='eq(name, "RetrieveDocs")',
    trace_filter='and(eq(feedback_key, "user_score"), eq(feedback_score, 1))',
    tree_filter='eq(name, "ExpandQuery")'
)

Advanced: export flattened trace view with child tool usage​

The following Python example demonstrates how to export a flattened view of traces, including information on the tools (from nested runs) used by the agent within each trace. This can be used to analyze the behavior of your agents across multiple traces.

This example queries all tool runs within a specified number of days and groups them by their parent (root) run ID. It then fetches the relevant information for each root run, such as the run name, inputs, outputs, and combines that information with the child run information.

To optimize the query, the example:

  1. Selects only the necessary fields when querying tool runs to reduce query time.
  2. Fetches root runs in batches while processing tool runs concurrently.
from collections import defaultdict
from concurrent.futures import Future, ThreadPoolExecutor
from datetime import datetime, timedelta

from langsmith import Client
from tqdm.auto import tqdm

client = Client()
project_name = "my-project"
num_days = 30

# List all tool runs
tool_runs = client.list_runs(
    project_name=project_name,
    start_time=datetime.now() - timedelta(days=num_days),
    run_type="tool",
    # We don't need to fetch inputs, outputs, and other values that # may increase the query time
    select=["trace_id", "name", "run_type"],
)

data = []
futures: list[Future] = []
trace_cursor = 0
trace_batch_size = 50

tool_runs_by_parent = defaultdict(lambda: defaultdict(set))
# Do not exceed rate limit
with ThreadPoolExecutor(max_workers=2) as executor:
    # Group tool runs by parent run ID
    for run in tqdm(tool_runs):
        # Collect all tools invoked within a given trace
        tool_runs_by_parent[run.trace_id]["tools_involved"].add(run.name)
        # maybe send a batch of parent run IDs to the server
        # this lets us query for the root runs in batches
        # while still processing the tool runs
        if len(tool_runs_by_parent) % trace_batch_size == 0:
            if this_batch := list(tool_runs_by_parent.keys())[
                trace_cursor : trace_cursor + trace_batch_size
            ]:
                trace_cursor += trace_batch_size
                futures.append(
                    executor.submit(
                        client.list_runs,
                        project_name=project_name,
                        run_ids=this_batch,
                        select=["name", "inputs", "outputs", "run_type"],
                    )
                )
    if this_batch := list(tool_runs_by_parent.keys())[trace_cursor:]:
        futures.append(
            executor.submit(
                client.list_runs,
                project_name=project_name,
                run_ids=this_batch,
                select=["name", "inputs", "outputs", "run_type"],
            )
        )

for future in tqdm(futures):
    root_runs = future.result()
    for root_run in root_runs:
        root_data = tool_runs_by_parent[root_run.id]
        data.append(
            {
                "run_id": root_run.id,
                "run_name": root_run.name,
                "run_type": root_run.run_type,
                "inputs": root_run.inputs,
                "outputs": root_run.outputs,
                "tools_involved": list(root_data["tools_involved"]),
            }
        )

# (Optional): Convert to a pandas DataFrame

import pandas as pd

df = pd.DataFrame(data)
df.head()

Advanced: export retriever IO for traces with feedback​

This query is useful if you want to fine-tune embeddings or diagnose end-to-end system performance issues based on retriever behavior. The following Python example demonstrates how to export retriever inputs and outputs within traces that have a specific feedback score.

from collections import defaultdict
from concurrent.futures import Future, ThreadPoolExecutor
from datetime import datetime, timedelta

import pandas as pd
from langsmith import Client
from tqdm.auto import tqdm

client = Client()
project_name = "your-project-name"
num_days = 1

# List all tool runs
retriever_runs = client.list_runs(
    project_name=project_name,
    start_time=datetime.now() - timedelta(days=num_days),
    run_type="retriever",
    # This time we do want to fetch the inputs and outputs, since they
    # may be adjusted by query expansion steps.
    select=["trace_id", "name", "run_type", "inputs", "outputs"],
    trace_filter='eq(feedback_key, "user_score")',
)

data = []
futures: list[Future] = []
trace_cursor = 0
trace_batch_size = 50

retriever_runs_by_parent = defaultdict(lambda: defaultdict(list))
# Do not exceed rate limit
with ThreadPoolExecutor(max_workers=2) as executor:
    # Group retriever runs by parent run ID
    for run in tqdm(retriever_runs):
        # Collect all retriever calls invoked within a given trace
        for k, v in run.inputs.items():
            retriever_runs_by_parent[run.trace_id][f"retriever.inputs.{k}"].append(v)
        for k, v in (run.outputs or {}).items():
            # Extend the docs
            retriever_runs_by_parent[run.trace_id][f"retriever.outputs.{k}"].extend(v)
        # maybe send a batch of parent run IDs to the server
        # this lets us query for the root runs in batches
        # while still processing the retriever runs
        if len(retriever_runs_by_parent) % trace_batch_size == 0:
            if this_batch := list(retriever_runs_by_parent.keys())[
                trace_cursor : trace_cursor + trace_batch_size
            ]:
                trace_cursor += trace_batch_size
                futures.append(
                    executor.submit(
                        client.list_runs,
                        project_name=project_name,
                        run_ids=this_batch,
                        select=[
                            "name",
                            "inputs",
                            "outputs",
                            "run_type",
                            "feedback_stats",
                        ],
                    )
                )
    if this_batch := list(retriever_runs_by_parent.keys())[trace_cursor:]:
        futures.append(
            executor.submit(
                client.list_runs,
                project_name=project_name,
                run_ids=this_batch,
                select=["name", "inputs", "outputs", "run_type"],
            )
        )

for future in tqdm(futures):
    root_runs = future.result()
    for root_run in root_runs:
        root_data = retriever_runs_by_parent[root_run.id]
        feedback = {
            f"feedback.{k}": v.get("avg")
            for k, v in (root_run.feedback_stats or {}).items()
        }
        inputs = {f"inputs.{k}": v for k, v in root_run.inputs.items()}
        outputs = {f"outputs.{k}": v for k, v in (root_run.outputs or {}).items()}
        data.append(
            {
                "run_id": root_run.id,
                "run_name": root_run.name,
                **inputs,
                **outputs,
                **feedback,
                **root_data,
            }
        )

# (Optional): Convert to a pandas DataFrame
import pandas as pd
df = pd.DataFrame(data)
df.head()

Was this page helpful?


You can leave detailed feedback on GitHub.