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 at present’s article.

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

Brokers resolve step-by-step which software to name and when to reply a query about our knowledge. It is a huge distinction from an utility within the conventional sense, the place the developer defines how the method works (e.g., through 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 discuss.

This text is for you if you happen to:

• …work with Pandas and wish to automate EDA.
• …discover LLMs thrilling, however have little expertise with LangChain to this point.
• …wish 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 elements:

Agent = LLM + Instruments + Management logic

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

• The LLM supplies the intelligence: It understands the query, plans steps, and decides what to do.
• The instruments are small Python capabilities that the agent is allowed to name (e.g., get_schema() or get_nulls()): They supply particular info from the information, similar 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 software. It thinks step-by-step: First, the query is analyzed, then the suitable software is chosen, then the result’s interpreted and, if essential, a subsequent step is chosen, and at last a response is returned.

As an alternative of manually describing all knowledge as in traditional prompting, we switch the duty 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 person asks: “What’s the common age within the CSV?”

At this level, the agent calls up the software 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 scale back or reduce hallucinations, because the system is aware of what to use and can’t return a solution similar 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 techniques 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, you’ll find 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 knowledge evaluation):

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

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

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 doable. This fashion, 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 consequence.

In our instance, we outline three instruments:

• schema: Returns column names and knowledge varieties.
• nulls: Reveals columns with lacking values (together with quantity).
• describe: Gives descriptive statistics for numeric columns.

Later, we are going to 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 aren’t constructing a traditional program with a hard and fast sequence (e.g., utilizing if-else), however slightly the mannequin plans itself primarily based on the query, selects the suitable software, 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 larger 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 possibly can test this by getting 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 initiatives.

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 capabilities and seaborn for loading the Titanic pattern dataset.

Tip for newbies:
In the event you don’t wish to use OpenAI, you’ll be able to work regionally with Ollama (e.g., with Llama or Mistral). This feature is accessible 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 it’s also possible to use one other IDE. On this file, we load the Titanic dataset, as it’s publicly obtainable.

# 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.

Ideas 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 foremost.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:

Aspect Notice – 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 information. The Titanic dataset is a traditional instance of exploratory knowledge evaluation (EDA). It incorporates info on 891 passengers of the Titanic and is commonly used to research the connection between traits (e.g., gender, age, ticket class) and survival.

Listed below are the 14 columns with a quick 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 dad and mom/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,” “youngster.”
• adult_male: Boolean discipline: Was the passenger an grownup male (True/False)?
• deck: Cabin deck (typically 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 should utilize LangSmith.

2) Outline instruments in foremost.py

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

# foremost.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 software

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

@software
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()
    consequence = {col: int(n) for col, n in nulls.objects() if n > 0}
    return json.dumps(consequence)

@software
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 knowledge varieties as JSON. This provides us an outline of what we’re coping with and is often step one in any knowledge evaluation. Even when a software doesn’t want enter (like schema), it should nonetheless settle for one argument, as a result of the agent all the time 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 be aware that this software solely works for numeric columns. Strings or Booleans, then again, are ignored. This is a vital step within the sanity test or EDA. This enables us to rapidly see the imply, min, max, and so on. of the completely different columns. For giant CSVs, describe() can take a very long time. On this case, you may combine df.pattern(n=10000) as sampling logic, for instance.

These instruments are the managed interfaces via 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, but it surely can’t immediately “see” knowledge. 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 capabilities that the agent is allowed to name. As an alternative of creating the whole lot free, we solely enable very particular, reproducible actions.

What precisely does the code do?

With the @software decorator, LangChain routinely infers the software’s identify, description and argument schema from the perform signature and docstring. This implies we solely want to jot down the perform itself. LangChain takes care of the remainder.

• The mannequin passes arguments that match the software’s schema (typically JSON). On this tutorial we preserve issues easy and settle for a single string argument (e.g., input_str: str or a dummy string we ignore).
• Instruments all the time return a string (textual content). JSON is right for structured knowledge, which we outline with return json.dumps(…).

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

4) Registering instruments for LangChain in foremost.py

We add the code under to the identical foremost.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 capabilities into an inventory. Every perform has already been transformed right into a LangChain software by the @software decorator.

5) Configuring LLM in foremost.py

Subsequent, we configure the LLM that the agent makes use of. Right here, you’ll be able to both use the variant for OpenAI or for an open-source software 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 is not going to 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 test whether or not an API key has been saved.

Now we add the next code to the tip of foremost.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 steady software 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 drag 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 accessible, in any other case Ollama regionally.

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

We additionally use gpt-4o-mini because the LLM. It is a light-weight mannequin from OpenAI with a concentrate on software utilization.

Tip for Newbies:
The temperature determines how creatively an LLM responds. If we enter 0.0, it responds deterministically. Which means that the mannequin virtually all the time returns the identical reply when the enter is identical. That is good for structured duties similar to software utilization, code or info, for instance. If we specify 1.0, the mannequin responds creatively and with all kinds of choices. Which means that the mannequin varies extra and may counsel completely different formulations or options, which is sweet for brainstorming or textual content concepts, for instance.

6) Defining the agent’s conduct in foremost.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 applicable software. "
    "Use just one software name per step if doable. "
    "Reply concisely and in a structured means. "
    "If no software suits, briefly clarify why.nn"
    "Accessible 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 part of(f"- {t.identify}: {t.description}" for t in instruments)
_tool_names = ", ".be 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 might solely use one software per step, that it must be concise, and that it could solely use the instruments we now have outlined.

With the final two strains within the system immediate, we make sure that {instruments} lists all obtainable 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 closing reply.

7) Create tool-calling agent in foremost.py

Within the final step, we outline the agent:

# --- 5) Create & Run Device-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? Record 'Column: Rely'."
    consequence = agent_executor.invoke({"enter": user_query})
    print("n=== AGENT ANSWER ===")
    print(consequence["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 executed, then calls up a software, receives the consequence and decides whether or not one other software is required or whether or not it may 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 software was referred to as when or what knowledge was returned. If the whole lot 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–software–response cycles the agent might carry out. This helps forestall infinite loops or extreme software calls. In our instance, the agent may moderately 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 == “foremost”: is a typical Python sample: If we execute the file immediately within the terminal with python foremost.py, the code on this block can 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 initiatives.

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

Now we enter python foremost.py within the terminal to start out the agent. We then see the ultimate reply within the terminal:

Mini-Analysis

Lastly, we wish to test 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 in a while.

On the finish of foremost.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 information.

The decrease block ensures {that a} check run is carried out when foremost.py is named immediately. If, then again, we import foremost 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 foremost import ask_agent

assessments = [
    ("Which columns have missing values?", ["age", "embarked", "deck", "embark_town"]),
    ("Present me the primary 3 columns with their knowledge varieties.", ["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 assessments:
        out = ask_agent(q)
        consequence = handed(q, out, should)
        print(f"[{'OK' if result else 'FAIL'}] {q}n{out}n")
        okay += int(consequence)
    print(f"Handed {okay}/{len(assessments)}")

Within the code, we outline three check circumstances. Every check consists of a query for the agent and an inventory 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 knowledge varieties.” Anticipated: Output incorporates survived, pclass, intercourse with varieties similar 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 the whole lot runs appropriately, the script reviews “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, suggestions and subsequent steps

LangChain may be very sensible for this instance as a result of it already consists of and properly illustrates all the agent loop (planning, software calling, management). For small or clearly structured duties, nonetheless, alternate options similar to pure perform calling (e.g., through the OpenAI API) or traditional EDA frameworks like Nice Expectations is perhaps enough. That stated, LangChain does add some overhead. In the event you solely want fastened EDA checks, a plain Python script could be leaner and quicker. LangChain is particularly worthwhile if you wish 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 software descriptions: If the descriptions are too imprecise, the mannequin can simply select the improper software (misrouting). With exact and concrete descriptions, we are able to vastly scale back this.

One other essential level is testing: Even a small mini-evaluation with three easy assessments is useful in detecting regressions (errors that keep unnoticed as a result of subsequent modifications) 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 massive CSV information to keep away from efficiency points. Needless to say LLM brokers may turn out to be expensive if each query triggers a number of software calls.

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

As a subsequent step, we might additionally incorporate LangGraph to increase the agent loop with states and orchestration. This may enable 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?