Constructing LLM Apps That Can See, Suppose, and Combine: Utilizing o3 with Multimodal Enter and Structured Output

, the usual “textual content in, textual content out” paradigm will solely take you thus far.

Actual purposes that ship precise worth ought to have the ability to look at visuals, cause by way of advanced issues, and produce outcomes that methods can truly use.

On this put up, we’ll design this stack by bringing collectively three highly effective capabilities: multimodal enter, reasoning, and structured output.

For instance this, we’ll stroll by way of a hands-on instance: constructing a time-series anomaly detection system for e-commerce order knowledge utilizing OpenAI’s o3 mannequin. Particularly, we’ll present find out how to pair o3’s reasoning functionality with picture enter and emit validated JSON, in order that the downstream system can simply eat it.

By the top, our app will:

• See: analyze charts of e-commerce order quantity time collection
• Suppose: establish uncommon patterns
• Combine: output a structured anomaly report

You’ll go away with useful code you possibly can reuse for numerous use circumstances that transcend simply anomaly detection.

Let’s dive in.

Curious about studying the broader panorama of how LLMs are being utilized for anomaly detection? Take a look at my earlier put up: Boosting Your Anomaly Detection With LLMs, the place I summarized 7 rising utility patterns that you simply shouldn’t miss.

1. Case Research

On this put up, we intention to construct an anomaly detection resolution for figuring out irregular patterns in e-commerce order time collection knowledge.

For this case research, we generated three units of artificial each day order knowledge. The datasets symbolize three totally different profiles of the each day order over roughly one month of time. To make seasonality apparent, now we have shaded the weekends. The x-axis exhibits the day of the week.

Every determine incorporates one particular sort of anomaly (can you discover them?). We’ll later use these figures to check our anomaly detection resolution and see if it may precisely recuperate these anomalies.

2. Our Answer

2.1 Overview

Not like the normal machine studying approaches that require tedious function engineering and mannequin coaching, our present strategy is far easier. It really works with the next steps:

1. We put together the determine for visualizing the e-commerce order time collection knowledge.
2. We immediate the reasoning mannequin o3, ask it to take a better take a look at the time collection picture we fed to it, and decide if an uncommon sample exists.
3. The o3 mannequin will then output its findings in a pre-defined JSON format.

And that’s it. Easy.

After all, to ship this resolution, we have to allow o3 mannequin to take picture enter and emit structured output. We’ll see how to try this shortly.

2.2 Establishing the reasoning mannequin

As talked about earlier than, we’ll use o3 mannequin, which is the flagship reasoning mannequin from OpenAI that may sort out advanced multi-step issues with state-of-the-art efficiency. Particularly, we’ll use the Azure OpenAI endpoint to name the mannequin.

Be sure you have put the endpoint, API key, and deployment identify in an .env file, we are able to then proceed to establishing the LLM consumer:

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import matplotlib.dates as mdates

from openai import AzureOpenAI
from dotenv import load_dotenv
import os

load_dotenv()

# Setup LLM consumer
endpoint = os.getenv("api_base")
api_key = os.getenv("o3_API_KEY")
api_version = "2025-04-01-preview"
model_name = "o3"
deployment = os.getenv("deployment_name")

LLM_client = AzureOpenAI(
    api_key=api_key,  
    api_version=api_version,
    azure_endpoint=endpoint
)

We use the next instruction because the system message for the o3 mannequin (tuned by GPT-5):

instruction = f"""

[Role]
You're a meticulous knowledge analyst.

[Task]
You'll be given a line chart picture associated to each day e-commerce orders. 
Your activity is to establish outstanding anomalies within the knowledge.

[Rules]
The anomaly sorts will be spike, drop, level_shift, or seasonal_outlier.
A level_shift is a sustained baseline change (≥ 5 consecutive days), not a single level.
A seasonal_outlier occurs if a weekend/weekday behaves in contrast to friends in its class. 
For instance, weekend orders are often decrease than the weekdays'.
Learn dates/values from axes; should you can’t learn precisely, snap to the closest tick and be aware uncertainty in clarification.
The weekends are shaded within the determine.
"""

Within the above instruction, we clearly outlined the function of the LLM, the duty that the LLM ought to full, and the principles the LLM ought to comply with.

To restrict the complexity of our case research, we deliberately specified solely 4 anomaly sorts that LLM must establish. We additionally offered clear definitions of these anomaly sorts to take away ambiguity.

Lastly, we injected a little bit of area data about e-commerce patterns, i.e., decrease weekend orders are anticipated in comparison with weekdays. Incorporating area know-how is usually thought of good observe for guiding the mannequin’s analytical course of.

Now that now we have our mannequin arrange, let’s focus on find out how to put together the picture for o3 mannequin to eat.

2.3 Picture preparation

To allow o3’s multimodal capabilities, we have to present figures in a particular format, i.e., both publicly accessible internet URLs or as base64-encoded knowledge URLs. Since our figures are generated domestically, we’ll use the second strategy.

What’s Base64 Encoding anyway? Base64 is a approach to symbolize binary knowledge (like our picture recordsdata) utilizing solely textual content characters which are secure to transmit over the web. It converts binary picture knowledge right into a string of letters, numbers, and some symbols.

And what about knowledge URL? An information URL is a kind of URL that embeds the file content material straight within the URL string, relatively than pointing to a file location.

We will use the next perform to deal with this conversion mechanically:

import io
import base64

def fig_to_data_url(fig, fmt="png"):
    """
    Converts a Matplotlib determine to a base64 knowledge URL with out saving to disk.

    Args:
    -----
    fig (matplotlib.determine.Determine): The determine to transform.
    fmt (str): The format of the picture ("png", "jpeg", and many others.)

    Returns:
    --------
    str: The info URL representing the determine.
    """

    buf = io.BytesIO()
    fig.savefig(buf, format=fmt, bbox_inches="tight")
    buf.search(0)
    
    base64_encoded_data = base64.b64encode(buf.learn()).decode("utf-8")
    mime_type = f"picture/{fmt.decrease()}"
    
    return f"knowledge:{mime_type};base64,{base64_encoded_data}"

Basically, our perform first saves the matplotlib determine to a reminiscence buffer. It then encodes the binary PNG knowledge as base64 textual content and wraps it within the desired knowledge URL format.

Assuming now we have entry to the artificial each day order knowledge, we are able to use the next perform to generate the plot and convert it into a correct knowledge URL format in a single go:

def create_fig(df):
    """
    Create a Matplotlib determine and convert it to a base64 knowledge URL.
    Weekends (Sat–Solar) are shaded.

    Args:
    -----
    df: dataframe incorporates one profile of each day order time collection. 
        dataframe has "date" and "orders" columns.

    Returns:
    --------
    image_url: The info URL representing the determine.
    """

    df = df.copy()
    df['date'] = pd.to_datetime(df['date'])

    fig, ax = plt.subplots(figsize=(8, 4.5))
    ax.plot(df["date"], df["orders"], linewidth=2)
    ax.set_xlabel('Date', fontsize=14)
    ax.set_ylabel('Every day Orders', fontsize=14)

    # Weekend shading
    begin = df["date"].min().normalize()
    finish   = df["date"].max().normalize()
    cur = begin
    whereas cur <= finish:
        if cur.weekday() == 5:  # Saturday 00:00
            span_start = cur                                      # Sat 00:00
            span_end   = cur + pd.Timedelta(days=1)               # Mon 00:00
            ax.axvspan(span_start, span_end, alpha=0.12, zorder=0)
            cur += pd.Timedelta(days=2)                           # skip Sunday 
        else:
            cur += pd.Timedelta(days=1)

    # Title
    title = f'Every day Orders: {df["date"].min():%b %d, %Y} - {df["date"].max():%b %d, %Y}'
    ax.set_title(title, fontsize=16)

    # Format x-axis dates
    ax.xaxis.set_major_formatter(mdates.DateFormatter('%b %d')) 
    ax.xaxis.set_major_locator(mdates.WeekdayLocator(interval=1))

    plt.tight_layout()

    # Receive url
    image_url = fig_to_data_url(fig)

    return image_url

Figures 1-3 are generated by the above plotting routine.

2.4 Structured output

On this part, let’s focus on how to make sure the o3 mannequin outputs a constant JSON format as a substitute of free-form textual content. That is what’s referred to as “structured output,” and it’s one of many key enablers for integrating LLMs into present computerized workflows.

To attain that, we begin by defining the schema that governs the anticipated output construction. We’ll be utilizing a Pydantic mannequin:

from pydantic import BaseModel, Discipline
from typing import Literal
from datetime import date

AnomalyKind = Literal["spike", "drop", "level_shift", "seasonal_outlier"]

class DateWindow(BaseModel):
    begin: date = Discipline(description="Earliest believable date the anomaly begins (ISO YYYY-MM-DD)")
    finish: date = Discipline(description="Newest believable date the anomaly ends, inclusive (ISO YYYY-MM-DD)")

class AnomalyReport(BaseModel):
    when: DateWindow = Discipline(
        description=(
            "Minimal window that incorporates the anomaly. "
            "For single-point anomalies, use the interval that covers studying uncertainty, if the tick labels are unclear"
        )
    )
    y: int = Discipline(description="Approx worth on the anomaly’s most consultant day (peak/lowest), rounded")
    variety: AnomalyKind = Discipline(description="The kind of the anomaly")
    why: str = Discipline(description="One-sentence cause for why this window is uncommon")
    date_confidence: Literal["low","medium","high"] = Discipline(
        default="medium", description="Confidence that the window localization is appropriate"
    )

Our Pydantic schema tries to seize each the quantitative and qualitative elements of the detected anomalies. For every area, we specify its knowledge sort (e.g., int for numerical values, Literal for a set set of selections, and many others.).

Additionally, we use Discipline perform to offer detailed descriptions of every key. These descriptions are particularly essential as they successfully function inline directions for o3, in order that it understands the semantic which means of every part.

Now, now we have lined the multimodal enter and structured output, time to place them collectively in a single LLM name.

2.5 o3 mannequin invocation

To work together with o3 utilizing multimodal enter and structured output, we use LLM_client.beta.chat.completions.parse() API. A number of the key arguments embody:

• mannequin: the deployment identify;
• messages: the message object despatched to o3 mannequin;
• max_completion_token: the utmost variety of tokens the mannequin can generate in its closing response. Word that for reasoning fashions like o3, they may generate reasoning_tokens internally to “assume by way of” the issue. The present max_completion_token solely limits the seen output tokens that customers obtain;
• response_format: the Pydantic mannequin that defines the anticipated JSON schema construction;
• reasoning_effort: a management knob that dictates how a lot computational effort o3 ought to use for reasoning. The accessible choices embody low, medium, and excessive.

We will outline a helper perform to work together with the o3 mannequin:

def anomaly_detection(instruction, fig_path, 
                      response_format, immediate=None, 
                      deployment="o3", reasoning_effort="excessive"):

    # Compose messages
    messages=[
            { "role": "system", "content": instruction},
            { "role": "user", "content": [  
                { 
                    "type": "image_url",
                    "image_url": {
                        "url": fig_path,
                        "detail": "high"
                    }
                },
            ]} 
    ]

    # Add immediate whether it is given
    if immediate just isn't None:
        messages[1]["content"].append({"sort": "textual content", "textual content": immediate})

    # Invoke LLM API
    response = LLM_client.beta.chat.completions.parse(
        mannequin=deployment,
        messages=messages,
        max_completion_tokens=4000,
        reasoning_effort=reasoning_effort,
        response_format=response_format
    )

    return response.selections[0].message.parsed.model_dump()

Word that the messages object accepts each textual content and picture content material. Since we’ll solely use figures to immediate the mannequin, the textual content immediate is optionally available.

We set the "element": "excessive" to allow high-resolution picture processing. For our present case research, that is more than likely essential as we’d like o3 to higher learn high quality particulars like axis tick labels, knowledge level values, and delicate visible patterns. Nonetheless, keep in mind that high-detail processing would incur extra tokens and better API prices.

Lastly, through the use of .parsed.model_dump(), we flip the JSON output right into a typical Python dictionary.

That’s it for the implementation. Let’s see some outcomes subsequent.

3. Outcomes

On this part, we’ll enter the beforehand generated figures into the o3 mannequin and ask it to establish potential anomalies.

3.1 Spike anomaly

# df_spike_anomaly is the dataframe of the primary set of artificial knowledge (Determine 1)
spike_anomaly_url = create_fig(df_spike_anomaly)

# Anomaly detection
consequence = anomaly_detection(instruction,
                          spike_anomaly_url,
                          response_format=AnomalyReport,
                          reasoning_effort="medium")
print(consequence)

Within the name above, the spike_anomaly_url is the info URL for Determine 1. The output of the result’s proven under:

{
  'when': {'begin': datetime.date(2025, 8, 19), 'finish': datetime.date(2025, 8, 21)}, 
  'y': 166, 
  'variety': 'spike', 
  'why': 'Single day orders leap to ~166, far above adjoining days that sit close to 120–130.', 
  'date_confidence': 'medium'
}

We see that o3 mannequin faithfully returned the output precisely within the format we designed. Now, we are able to seize this consequence and generate a visualization programmatically:

# Create picture
fig, ax = plt.subplots(figsize=(8, 4.5))
df_spike_anomaly['date'] = pd.to_datetime(df_spike_anomaly['date'])
ax.plot(df_spike_anomaly["date"], df_spike_anomaly["orders"], linewidth=2)
ax.set_xlabel('Date', fontsize=14)
ax.set_ylabel('Every day Orders', fontsize=14)

# Format x-axis dates
ax.xaxis.set_major_formatter(mdates.DateFormatter('%b %d'))  
ax.xaxis.set_major_locator(mdates.WeekdayLocator(interval=1)) 

# Add anomaly overlay
start_date = pd.to_datetime(consequence['when']['start'])
end_date = pd.to_datetime(consequence['when']['end'])

# Add shaded area
ax.axvspan(start_date, end_date, alpha=0.3, coloration='crimson', label=f"Anomaly ({consequence['kind']})")

# Add textual content annotation
mid_date = start_date + (end_date - start_date) / 2  # Center of anomaly window
ax.annotate(
    consequence['why'], 
    xy=(mid_date, consequence['y']), 
    xytext=(10, 20),  # Offset from the purpose
    textcoords='offset factors',
    bbox=dict(boxstyle='spherical,pad=0.5', fc='yellow', alpha=0.7),
    arrowprops=dict(arrowstyle='->', connectionstyle='arc3,rad=0.1'),
    fontsize=10,
    wrap=True
)

# Add legend
ax.legend()

plt.xticks(rotation=0)
plt.tight_layout()

The generated visualization seems like this:

We will see that the o3 mannequin accurately recognized the spike anomaly offered on this first set of artificial knowledge.

Not unhealthy, particularly contemplating the truth that we didn’t do any typical mannequin coaching, simply by prompting an LLM.

3.2 Degree shift anomaly

# df_level_shift_anomaly is the dataframe of the 2nd set of artificial knowledge (Determine 2)
level_shift_anomaly_url = create_fig(df_level_shift_anomaly)

# Anomaly detection
consequence = anomaly_detection(instruction,
                          level_shift_anomaly_url,
                          response_format=AnomalyReport,
                          reasoning_effort="medium")
print(consequence)

The output of the result’s proven under:

{
  'when': {'begin': datetime.date(2025, 8, 26), 'finish': datetime.date(2025, 9, 2)}, 
  'y': 150, 
  'variety': 'level_shift', 
  'why': 'Orders all of the sudden leap from the 120-135 vary to ~150 on Aug 26 and stay elevated for all subsequent days, indicating a sustained baseline change.', 
  'date_confidence': 'excessive'
}

Once more, we see that the mannequin precisely recognized {that a} “level_shift” anomaly is current within the plot:

3.3 Seasonality anomaly

# df_seasonality_anomaly is the dataframe of the third set of artificial knowledge (Determine 3)
seasonality_anomaly_url = create_fig(df_seasonality_anomaly)

# Anomaly detection
consequence = anomaly_detection(instruction,
                          seasonality_anomaly_url,
                          response_format=AnomalyReport,
                          reasoning_effort="medium")
print(consequence)

The output of the result’s proven under:

{
  'when': {'begin': datetime.date(2025, 8, 23), 'finish': datetime.date(2025, 8, 24)}, 
  'y': 132, 
  'variety': 'seasonal_outlier', 
  'why': 'Weekend of Aug 23-24 exhibits order volumes (~130+) on par with surrounding weekdays, whereas different weekends persistently drop to ~115, making it an out-of-season spike.', 
  'date_confidence': 'excessive'
}

This can be a difficult case. However, our o3 mannequin managed to sort out it correctly, with correct localization and a transparent reasoning hint. Fairly spectacular:

4. Abstract

Congratulations! We’ve efficiently constructed an anomaly detection resolution for time-series knowledge that labored solely by way of visualization and prompting.

By feeding each day order plots into the o3 reasoning mannequin and constraining its output to a JSON schema, the LLM managed to establish three totally different anomaly sorts with correct localization. All of this was achieved with out coaching any ML mannequin. Spectacular!

If we take a step again, we are able to see that the answer we constructed illustrates the broader sample of mixing three capabilities:

• See: multimodal enter to let the mannequin eat figures straight.
• Suppose: step-by-step reasoning functionality to sort out advanced issues.
• Combine: structured output that downstream methods can simply eat (e.g., producing visualizations).

The mix of multimodal enter + reasoning + structured output actually creates a flexible basis for helpful LLM purposes.

You now have the constructing blocks prepared. What do you wish to construct subsequent?