Skip to main content

How to query traces and runs

LangSmith makes it easy to query for traces and runs. In addition to the filtering experience presented in the UI, you can also use the SDK or API to query for traces and runs.

Using the list_runs method in the SDK or /runs/query endpoint in the API, you can filter runs to analyze and export. Most simple requests can be satisfied using simple top level arguments:

KeysDescription
project_id / project_nameThe project(s) to fetch runs from - can be a single project or a list of projects.
trace_idFetch runs that are part of a specific trace.
run_typeThe type of run to get, such as llm, chain, tool, retriever, etc.
dataset_name / dataset_idFetch runs that are associated with an example row in the specified dataset. This is useful for comparing prompts or models over a given dataset.
reference_example_idFetch runs that are associated with a specific example row. This is useful for comparing prompts or models on a given input.
parent_run_idFetch runs that are children of a given run. This is useful for fetching runs grouped together using the context manager or for fetching an agent trajectory.
errorFetch runs that errored or did not error.
run_idsFetch runs with a given list of run ids. Note: This will ignore all other filtering arguments.
filterFetch runs that match a given structured filter statement. See the run filtering guide below for more information.
trace_filterFilter to apply to the ROOT run in the trace tree. This is meant to be used in conjunction with the regular filter parameter to let you filter runs by attributes of the root run within a trace.
tree_filterFilter to apply to OTHER runs in the trace tree, including sibling and child runs. This is meant to be used in conjunction with the regular filter parameter to let you filter runs by attributes of any run within a trace.
is_rootOnly return root runs.
selectSelect the fields to return in the response. By default, all fields are returned.
query (experimental)Query the experimental natural language API, which translates your query into a filter statement.

Using keyword arguments​

For simple queries, such as filtering by project, run time, name, or run ID's, you can directly use keyword arguments in the list_runs method. These correspond directly to query params in the REST API. All the examples below assume you have created a LangSmith client and configured it with your API key to connect to the LangSmith server.

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 traces in a project​

Root runs (or run traces), are runs that have no parents. These are assigned an 'execution_order' of 1. 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​

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)
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.

Run Filtering​

Listing runs with query params is useful for simple queries, but doesn't support many common needs, such as filtering by metadata, tags, or other fields.

LangSmith supports a filter query language to permit more complex filtering operations when fetching runs. This guide will provide a high level overview of the grammar as well as a few examples of when it can be useful.

If you'd prefer a more visual guide, you can get a taste of the language by viewing the table of runs on any of your projects' pages. We provide some recommended filters to get you started that you can copy and use the SDK.

Grammar​

The filtering grammar is based on common comparators on fields in the run object. Supported comparators include:

  • gte (greater than or equal to)
  • gt (greater than)
  • lte (less than or equal to)
  • lt (less than)
  • eq (equal to)
  • neq (not equal to)
  • has (check if run contains a tag or metadata json blob)
  • search (search for a substring in a string field)

Additionally, you can combine multiple comparisons through and and or operators.

These can be applied on fields of the run object, such as its id, name, run_type, start_time / end_time, latency, total_tokens, error, execution_order, tags, and any associated feedback through feedback_key and feedback_score.

Examples​

The following examples assume you have configured your environment appropriately and have runs stored in LangSmith.

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 Examples​

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()

Export retriever inputs/outputs for traces with a specific feedback score​

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?