LangChain for EDA: Build a CSV Sanity-Check Agent in Python

, brokers carry out actions.

That’s precisely what we’re going to check out in as we speak’s article.

On this article, we’ll use LangChain and Python to construct our personal CSV sanity examine agent. With this agent, we’ll automate typical exploratory information evaluation (EDA) duties as displaying columns, detecting lacking values (NaNs) and retrieving descriptive statistics.

Brokers determine step-by-step which device to name and when to reply a query about our information. This can be a massive distinction from an software within the conventional sense, the place the developer defines how the method works (e.g., by way of if-else loops). It additionally goes far past easy prompting as a result of we’re constructing a system that acts (albeit in a easy means) and doesn’t simply speak.

This text is for you for those who:

• …work with Pandas and need to automate EDA.
• …discover LLMs thrilling, however have little expertise with LangChain to date.
• …need to perceive how brokers actually work (from setup to mini-evaluation) utilizing a easy instance.

Desk of Contents
What we build & why
Hands-On-Example: CSV-Sanity-Check Agent with LangChain
Mini-Evaluation
Final Thoughts – Pitfalls, Tips and Next Steps
Where Can You Continue Learning?

What we construct & why

An agent is a system to which we assign duties. The system then decides for itself which instruments to make use of to unravel these duties.

This requires three parts:

Agent = LLM + Instruments + Management logic

Let’s take a more in-depth take a look at the three parts:

• The LLM offers the intelligence: It understands the query, plans steps, and decides what to do.
• The instruments are small Python features that the agent is allowed to name (e.g., get_schema() or get_nulls()): They supply particular info from the information, akin to column names or statistics.
• The management logic (coverage) ensures that the LLM doesn’t reply instantly, however first decides whether or not it ought to use a device. It thinks step-by-step: First, the query is analyzed, then the suitable device is chosen, then the result’s interpreted and, if needed, a subsequent step is chosen, and eventually a response is returned.

As an alternative of manually describing all information as in traditional prompting, we switch the accountability to the agent: The system ought to act by itself, however solely with the instruments offered.

Let’s take a look at a easy instance:

A consumer asks: “What’s the common age within the CSV?”

At this level, the agent calls up the device we now have outlined, df.describe(). The output is a clearly structured worth (e.g., “imply”: 29.7). Right here we are able to additionally see that this will cut back or decrease hallucinations, because the system is aware of what to use and can’t return a solution akin to “In all probability between 20 and 40.”

LangChain as a framework

We use the LangChain framework for the agent. This enables us to attach LLMs with instruments and construct methods with outlined conduct. The system can carry out actions as an alternative of simply offering solutions or producing textual content. An in depth clarification would make this text too lengthy. However in a earlier article, yow will discover a proof of LangChain and a comparability with Langflow: LangChain vs Langflow: Build a Simple LLM App with Code or Drag & Drop.

What the agent does for us

After we obtain a brand new CSV, we often ask ourselves the next questions first (begin of exploratory information evaluation):

• What columns are there?
• The place is information lacking?
• What do the descriptive statistics seem like?

That is precisely what we would like the agent to do mechanically.

Instruments we outline for the agent

For the agent to work, it wants clearly outlined instruments. It’s best to outline them as small, particular, and managed as attainable. This manner, we keep away from errors, hallucinations or unclear outputs as a result of they make the output deterministic. Additionally they make the agent reproducible and testable as a result of the identical enter ought to produce a constant outcome.

In our instance, we outline three instruments:

• schema: Returns column names and information sorts.
• nulls: Reveals columns with lacking values (together with quantity).
• describe: Offers descriptive statistics for numeric columns.

Later, we’ll add a small mini-evaluation to make sure that our agent is working appropriately.

Why is that this an agent and never an app?

We’re not constructing a traditional program with a hard and fast sequence (e.g., utilizing if-else), however reasonably the mannequin plans itself primarily based on the query, selects the suitable device, and combines steps as essential to arrive at a solution:

Palms-On-Instance: CSV-Sanity-Verify Agent with LangChain

1) Setup

Prerequisite: Python 3.10 or increased should be put in. Many packages within the AI tooling world require ≥ 3.10. Yow will discover the code and the hyperlink to the repo under.

Tip for newbies:
You may examine this by coming into “python –model” in cmd.exe.

With the code under, we first create a brand new venture, create an remoted Python atmosphere and activate it. We do that in order that packages and variations are reproducible and don’t consolidate with different tasks.

Tip for newbies:
I work with Home windows. We open a terminal with Home windows + R > cmd and paste the next code.

mkdir csv-agent

cd csv-agent
python -m venv .venv
.venvScriptsactivate

Then we set up the mandatory packages:

pip set up "langchain>=0.2,<0.3" "langchain-openai>=0.1.7" "langchain-community>=0.2" pandas seaborn

With this command, we pin LangChain to the 0.2 line and set up the OpenAI connection and the neighborhood bundle. We additionally set up pandas for the EDA features and seaborn for loading the Titanic pattern dataset.

Tip for newbies:
For those who don’t need to use OpenAI, you possibly can work regionally with Ollama (e.g., with Llama or Mistral). This feature is on the market later within the code.

2) Put together the information set in prepare_data.py

Subsequent, we create a Python file referred to as prepare_data.py. I take advantage of Visible Studio Code for this, however you may also use one other IDE. On this file, we load the Titanic dataset, as it’s publicly accessible.

# prepare_data.py
import seaborn as sns
df = sns.load_dataset("titanic")
df.to_csv("titanic.csv", index=False)
print("Saved titanic.csv")

With seaborn.load_dataset(“titanic”), we load the general public dataset (891 rows + first row with column names) immediately into reminiscence and reserve it as titanic.csv. The dataset incorporates solely numeric, Boolean and categorical columns, making it supreme for an EDA agent.

Suggestions for newbies:

• sns.load_dataset() requires web entry (the information comes from the seaborn repo).
• Save the file within the venture folder (csv-agent) so htat essential.py can discover it.

Within the terminal, we execute the Python file with the next command, in order that the titanic.csv file is positioned within the venture:

python prepare_data.py

We then see within the terminal that the csv has been saved and see the titanic.csv file within the folder:

Facet Be aware – Titanic dataset

The evaluation is predicated on the Titanic dataset (OpenML ID 40945), which is marked as public on OpenML.

After we open the file, we see the next 14 columns and 891 rows of knowledge. The Titanic dataset is a traditional instance of exploratory information evaluation (EDA). It incorporates info on 891 passengers of the Titanic and is usually used to research the connection between traits (e.g., gender, age, ticket class) and survival.

Listed here are the 14 columns with a short clarification:

• survived: Survived (1) or didn’t survive (0).
• pclass: Ticket class (1 = 1st class, 2 = 2nd class, 3 = third class).
• intercourse: Gender of the passenger.
• age: Age of the passenger (in years, could also be lacking).
• sibsp: Variety of siblings/spouses on board.
• parch: Variety of mother and father/kids on board.
• fare: Fare paid by the passenger.
• embarked: Port of embarkation (C = Cherbourg, Q = Queenstown, S = Southampton).
• class: Ticket class as textual content (First, Second, Third). Corresponds to pclass.
• who: Categorization “man,” “girl,” “baby.”
• adult_male: Boolean discipline: Was the passenger an grownup male (True/False)?
• deck: Cabin deck (usually lacking).
• embark_town: Metropolis of port of embarkation (Cherbourg, Queenstown, Southampton).
• alone: Boolean discipline: Did the passenger journey alone (True/False)?

Non-compulsory for superior readers
If you wish to observe and consider your agent runs later, you need to use LangSmith.

2) Outline instruments in essential.py

Subsequent, we outline the assorted instruments. To do that, we create a brand new Python file referred to as essential.py and reserve it within the csv-agent folder as effectively. We add the next code to it:

# essential.py
import os, json
import pandas as pd

# --- 0) Loading CSV ---
DF_PATH = "titanic.csv"
df = pd.read_csv(DF_PATH)

# --- 1) Defining instruments as small, concise instructions ---
# IMPORTANT: Instruments return strings (on this case, JSON strings) in order that the LLM sees clearly structured responses.

from langchain_core.instruments import device

@device
def tool_schema(dummy: str) -> str:
    """Returns column names and information sorts as JSON."""
    schema = {col: str(dtype) for col, dtype in df.dtypes.gadgets()}
    return json.dumps(schema)

@device
def tool_nulls(dummy: str) -> str:
    """Returns columns with the variety of lacking values as JSON (solely columns with >0 lacking values)."""
    nulls = df.isna().sum()
    outcome = {col: int(n) for col, n in nulls.gadgets() if n > 0}
    return json.dumps(outcome)

@device
def tool_describe(input_str: str) -> str:
    """
    Returns describe() statistics.
    Non-compulsory: input_str can comprise a comma-separated checklist of columns, e.g. "age, fare".
    """
    cols = None
    if input_str and input_str.strip():
        cols = [c.strip() for c in input_str.split(",") if c.strip() in df.columns]
    stats = df[cols].describe() if cols else df.describe()
    # describe() has a MultiIndex. Flatten it for the LLM to maintain it readable:
    return stats.to_csv(index=True)

After importing the mandatory packages, we load titanic.csv into df as soon as and outline three small, narrowly outlined instruments. Let’s take a more in-depth take a look at every of those instruments:

• tool_schema returns the column names and information sorts as JSON. This offers us an outline of what we’re coping with and is often step one in any information evaluation. Even when a device doesn’t want enter (like schema), it should nonetheless settle for one argument, as a result of the agent at all times passes a string. We merely ignore it.
• tool_nulls counts lacking values per column and returns solely columns with lacking values.
• tool_describe calls df.describe(). You will need to word that this device solely works for numeric columns. Strings or Booleans, however, are ignored. This is a vital step within the sanity examine or EDA. This enables us to rapidly see the imply, min, max, and so forth. of the totally different columns. For giant CSVs, describe() can take a very long time. On this case, you would combine df.pattern(n=10000) as sampling logic, for instance.

These instruments are the managed interfaces by means of which the LLM is allowed to entry the information. They’re deterministic and subsequently reproducible. Instruments ought to ideally be clear and restricted: In different phrases, they need to have just one perform or process.

Why do we’d like instruments in any respect?

An LLM can generate textual content, nevertheless it can not immediately “see” information. To ensure that the LLM to work meaningfully with a CSV, we have to present interfaces. That’s precisely what instruments are for:

Instruments are small Python features that the agent is allowed to name. As an alternative of constructing every part free, we solely permit very particular, reproducible actions.

What precisely does the code do?

With the @device decorator, LangChain mechanically infers the device’s identify, description and argument schema from the perform signature and docstring. This implies we solely want to put in writing the perform itself. LangChain takes care of the remainder.

• The mannequin passes arguments that match the device’s schema (usually JSON). On this tutorial we maintain issues easy and settle for a single string argument (e.g., input_str: str or a dummy string we ignore).
• Instruments at all times return a string (textual content). JSON is good for structured information, which we outline with return json.dumps(…).

This can be a multi-step thought course of. The LLM plans iteratively. As an alternative of responding immediately, it thinks step-by-step: it decides which device to name, interprets the outcome, and should proceed till it has sufficient info to reply.

4) Registering instruments for LangChain in essential.py

We add the code under to the identical essential.py file to register the beforehand outlined instruments for the agent:

# --- 2) Registering instruments for LangChain ---

instruments = [tool_schema, tool_nulls, tool_describe]

With this code, we merely acquire the embellished features into a listing. Every perform has already been transformed right into a LangChain device by the @device decorator.

5) Configuring LLM in essential.py

Subsequent, we configure the LLM that the agent makes use of. Right here, you possibly can both use the variant for OpenAI or for an open-source device with Ollama.

I used OpenAI, which is why we first have to set the API key:

At OpenAI, we create a brand new API key:

We then copy it immediately (it won’t be displayed later) and set it as an atmosphere variable within the terminal with the next command.

setx OPENAI_API_KEY "your_key”

You will need to restart cmd and reactivate .venv afterwards. We are able to use echo to examine whether or not an API key has been saved.

Now we add the next code to the top of essential.py:

# --- 3) Configure LLM ---
# Choice A: OpenAI (easy)
#   export OPENAI_API_KEY=...    # Home windows: setx OPENAI_API_KEY "YOUR_KEY"
#   Use a decrease temperature for extra secure device utilization
USE_OPENAI = bool(os.getenv("OPENAI_API_KEY"))

if USE_OPENAI:
    from langchain_openai import ChatOpenAI
    llm = ChatOpenAI(mannequin="gpt-4o-mini", temperature=0.1)
else:
    # Choice B: Native with Ollama (ensure to tug the mannequin first, e.g. 'ollama run llama3')
    from langchain_community.chat_models import ChatOllama
    llm = ChatOllama(mannequin="llama3.1:8b", temperature=0.1)

The code makes use of OpenAI if an OpenAI_API_KEY is on the market, in any other case Ollama regionally.

We set the temperature to 0.1. This ensures that the responses are extra deterministic, which is necessary for the following check.

We additionally use gpt-4o-mini because the LLM. This can be a light-weight mannequin from OpenAI with a deal with device utilization.

Tip for Newbies:
The temperature determines how creatively an LLM responds. If we enter 0.0, it responds deterministically. Which means the mannequin virtually at all times returns the identical reply when the enter is similar. That is good for structured duties akin to device utilization, code or information, for instance. If we specify 1.0, the mannequin responds creatively and with all kinds of choices. Which means the mannequin varies extra and may recommend totally different formulations or options, which is nice for brainstorming or textual content concepts, for instance.

6) Defining the agent’s conduct in essential.py utilizing the coverage

On this step, we outline how the agent ought to behave. The system immediate units the coverage.

# --- 4) Slim Coverage/Immediate (Agent Conduct) ---
from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder

SYSTEM_PROMPT = (
    "You're a data-focused assistant. "
    "If a query requires info from the CSV, first use an acceptable device. "
    "Use just one device name per step if attainable. "
    "Reply concisely and in a structured means. "
    "If no device suits, briefly clarify why.nn"
    "Obtainable instruments:n{instruments}n"
    "Use solely these instruments: {tool_names}."
)

immediate = ChatPromptTemplate.from_messages(
    [
        ("system", SYSTEM_PROMPT),
        ("human", "{input}"),
        MessagesPlaceholder(variable_name="agent_scratchpad"),
    ]
)

_tool_desc = "n".be a part of(f"- {t.identify}: {t.description}" for t in instruments)
_tool_names = ", ".be a part of(t.identify for t in instruments)
immediate = immediate.partial(instruments=_tool_desc, tool_names=_tool_names)

First, we import ChatPromptTemplate to construction our agent’s immediate. A very powerful a part of the code is the system immediate: it defines the coverage, i.e., the “guidelines of the sport” for the agent. In it, we outline that the agent could solely use one device per step, that it must be concise, and that it could solely use the instruments we now have outlined.

With the final two traces within the system immediate, we make sure that {instruments} lists all accessible instruments with their descriptions and with {tool_names}, we make sure that the agent can solely use these names and can’t invent fantasy instruments.

As well as, we use MesagesPlaceholder(“agent_scratchpad”). That is the place the agent shops intermediate steps: The agent shops which instruments it has referred to as and which ends up it has obtained. This enables it to proceed its personal chain of reasoning till it arrives at a ultimate reply.

7) Create tool-calling agent in essential.py

Within the final step, we outline the agent:

# --- 5) Create & Run Software-Calling Agent ---
from langchain.brokers import create_tool_calling_agent, AgentExecutor

agent = create_tool_calling_agent(llm=llm, instruments=instruments, immediate=immediate)
agent_executor = AgentExecutor(
    agent=agent,
    instruments=instruments,
    verbose=False,   # elective: True for debug logs
    max_iterations=3,
)

if __name__ == "__main__":
    user_query = "Which columns have lacking values? Checklist 'Column: Rely'."
    outcome = agent_executor.invoke({"enter": user_query})
    print("n=== AGENT ANSWER ===")
    print(outcome["output"])

With create_tool_calling_agent, we join our LLM, the instruments and the immediate to type a tool-calling agent.

To make sure that the method runs easily, we use the AgentExecutor. It takes care of the so-called agent loop: The agent first plans what must be carried out, then calls up a device, receives the outcome and decides whether or not one other device is required or whether or not it could actually present the ultimate reply. This cycle repeats till the result’s prepared.

With verbose=True, we are able to view the intermediate steps within the terminal, which is extraordinarily useful for debugging. For instance, we are able to see which device was referred to as when or what information was returned. If every part is working easily, we are able to additionally set it to =False to maintain the output clearer.

With max_iterations=3, we restrict what number of reasoning–device–response cycles the agent could carry out. This helps stop infinite loops or extreme device calls. In our instance, the agent would possibly fairly name schema → nulls → describe earlier than answering.

With the final a part of the code, the agent is executed with the pattern enter “Which columns have lacking values?”. The result’s printed within the terminal.

Tip for newbies:
if identify == “essential”: is a regular Python sample: If we execute the file immediately within the terminal with python essential.py, the code on this block shall be began. Nevertheless, if we solely import the file (e.g., later within the mini_eval.py file), this block is skipped. This enables us to make use of the file as a standalone script or reuse it as a module in different tasks.

8) Run the script: Run the file essential.py within the terminal.

Now we enter python essential.py within the terminal to begin the agent. We then see the ultimate reply within the terminal:

Mini-Analysis

Lastly, we need to examine our agent, which we do with a small analysis. This ensures that the agent behaves appropriately and doesn’t introduce any “regressions” once we change one thing within the code afterward.

On the finish of essential.py, we add the code under:

def ask_agent(question: str) -> str:
    return agent_executor.invoke({"enter": question})["output"]

With ask_agent, we encapsulate the agent name in a perform that merely returns a string. This enables us to name the agent later from different recordsdata.

The decrease block ensures {that a} check run is carried out when essential.py known as immediately. If, however, we import essential into one other file, solely the perform is offered.

Now we create the mini_eval.py file and insert the next code:

# mini_eval.py

from essential import ask_agent

exams = [
    ("Which columns have missing values?", ["age", "embarked", "deck", "embark_town"]),
    ("Present me the primary 3 columns with their information sorts.", ["survived", "pclass", "sex"]),
    ("Give me a statistical abstract of the 'age' column.", ["mean", "min", "max"]),
]

def handed(q, out, must_include):
    textual content = out.decrease()
    return all(any(tok in textual content for tok in (m.decrease(), str(m).decrease())) for m in must_include)

if __name__ == "__main__":
    okay = 0
    for q, should in exams:
        out = ask_agent(q)
        outcome = handed(q, out, should)
        print(f"[{'OK' if result else 'FAIL'}] {q}n{out}n")
        okay += int(outcome)
    print(f"Handed {okay}/{len(exams)}")

Within the code, we outline three check instances. Every check consists of a query for the agent and a listing of key phrases that should seem within the reply. The handed() perform checks whether or not these key phrases are included.

Anticipated check outcomes

• Take a look at 1: “Which columns have lacking values?”
  Anticipated: Output mentions age, deck, embarked, embark_town.
• Take a look at 2: “Present me the primary 3 columns with their information sorts.” Anticipated: Output incorporates survived, pclass, intercourse with sorts akin to int64 or object.
• Take a look at 3: “Give me a statistical abstract of the ‘age’ column.” Anticipated output: Output incorporates imply ≈ 29.7, min = 0.42, max = 80.

If every part runs appropriately, the script studies “Handed 3/3” on the finish.

We get this output within the terminal. So the check works:

Yow will discover the code & the csv within the repo on GitHub.

On my Substack Data Science Espresso, I share sensible guides and bite-sized updates from the world of Knowledge Science, Python, AI, Machine Studying, and Tech — made for curious minds like yours.

Take a look and subscribe on Medium or on Substack if you wish to keep within the loop.

Remaining Ideas – Pitfalls, ideas and subsequent steps

LangChain may be very sensible for this instance as a result of it already consists of and properly illustrates the complete agent loop (planning, device calling, management). For small or clearly structured duties, nonetheless, alternate options akin to pure perform calling (e.g., by way of the OpenAI API) or traditional EDA frameworks like Nice Expectations may be enough. That mentioned, LangChain does add some overhead. For those who solely want fastened EDA checks, a plain Python script can be leaner and quicker. LangChain is very worthwhile if you need to lengthen issues flexibly or orchestrate a number of instruments and brokers.

When working with brokers, there are some things you must consider:

One frequent pitfall is unclear device descriptions: If the descriptions are too imprecise, the mannequin can simply select the unsuitable device (misrouting). With exact and concrete descriptions, we are able to drastically cut back this.

One other necessary level is testing: Even a small mini-evaluation with three easy exams is useful in detecting regressions (errors that keep unnoticed attributable to subsequent adjustments) at an early stage.

It’s additionally price beginning small: In our instance, we solely labored with three clearly outlined instruments, however now we all know that they work reliably.

With regard to this agent, it may also be helpful to include sampling (for instance, df.pattern(n=10000)) for very giant CSV recordsdata to keep away from efficiency points. Take into account that LLM brokers may develop into expensive if each query triggers a number of device calls.

On this article, we constructed a single agent that checks CSV recordsdata. In apply, a number of brokers would usually work collectively: For instance, one agent might guarantee information high quality whereas a second agent creates visualizations. Such multi-agent methods are the following step in fixing extra advanced duties.

As a subsequent step, we might additionally incorporate LangGraph to increase the agent loop with states and orchestration. This could permit us to assemble brokers as in a flowchart, together with interruptions, reminiscence, or extra versatile management logic.

Lastly, in our instance, we manually outlined the three instruments schema, nulls, and describe. With the Model Context Protocol (MCP), we might join instruments in a standardized means. For instance, we might join databases, APIs or IDEs.

The place Can You Proceed Studying?