Using LLM for Data Analysis

#product #oxlo #ai

Large language models have become standard tools in data analysis pipelines, handling tasks from semantic SQL generation to automated data cleaning. These workloads often require long system prompts, wide table schemas, and multi-turn reasoning, all of which can drive up inference costs on traditional token-based platforms. Oxlo.ai addresses this with a developer-first, request-based pricing model and a broad catalog of OpenAI-compatible models, making it a strong candidate for production data analysis.

Why LLMs for Data Analysis

Data teams routinely face three bottlenecks: translating business questions into query logic, normalizing inconsistent records, and producing reproducible analysis code. LLMs can directly assist with each. They parse natural language into SQL, infer structure from semi-clean CSVs, and generate Python or R snippets for visualization. The challenge is not capability, but cost control once you start passing large schemas or conducting iterative, agentic investigations.

Text-to-SQL with Oxlo.ai

One of the most reliable production patterns is natural language to SQL. By providing the LLM with a database schema and a business question, you can generate auditable queries without writing them manually. Because Oxlo.ai charges a flat rate per request, you can include extensive schema context and few-shot examples without worrying about token count.

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.oxlo.ai/v1",
    api_key=os.environ["OXLO_API_KEY"]
)

schema = """
CREATE TABLE transactions (
    id INTEGER PRIMARY KEY,
    user_id INTEGER,
    amount DECIMAL(12,2),
    currency VARCHAR(3),
    created_at TIMESTAMP,
    status VARCHAR(20)
);
"""

question = "Which users spent more than $1,000 in status 'completed' during March 2024?"

response = client.chat.completions.create(
    model="MODEL_ID",  # e.g., DeepSeek R1 671B MoE on Oxlo.ai
    messages=[
        {"role": "system", "content": "You are an expert SQL analyst. Return only the query."},
        {"role": "user", "content": f"Schema:\n{schema}\nQuestion: {question}"}
    ]
)

print(response.choices[0].message.content)

For general-purpose schema understanding, Llama 3.3 70B is also a solid choice. If your data source is multilingual, Qwen 3 32B handles cross-lingual schema descriptions well.

Structured Output for Data Cleaning

Raw data often arrives with inconsistent formatting, mixed currencies, and ambiguous categories. You can use JSON mode to enforce a strict output schema and turn messy records into structured objects. Oxlo.ai supports JSON mode across its chat models, so you can build deterministic cleaning pipelines.

import os
import json
from openai import OpenAI

client = OpenAI(base_url="https://api.oxlo.ai/v1", api_key=os.environ["OXLO_API_KEY"])

dirty_records = [
    {"company": "Acme Inc.", "revenue": "$5M", "headcount": "200+"},
    {"company": "ACME International", "revenue": "5 million USD", "headcount": "approx 200"},
]

prompt = f"""Standardize the following records into valid JSON.
Target schema: company_name (string), annual_revenue_usd (integer), employee_count (integer).

Records:
{json.dumps(dirty_records, indent=2)}
"""

response = client.chat.completions.create(
    model="MODEL_ID",  # e.g., Llama 3.3 70B or Kimi K2.6 on Oxlo.ai
    messages=[{"role": "user", "content": prompt}],
    response_format={"type": "json_object"}
)

clean = json.loads(response.choices[0].message.content)
print(clean)

Agentic Analysis and Long Context

Advanced analysis is rarely a single prompt. It involves multi-turn tool use, chaining Python execution, or iterating on visualizations. Models like Kimi K2.6, DeepSeek V4 Flash, and GLM 5 on Oxlo.ai support long context windows and function calling, which lets an agent retain state across reasoning steps.

The cost profile is important here. On token-based platforms, each turn bills every token in the context window. With Oxlo.ai flat per-request pricing, reloading a long schema or conversation history does not inflate the cost. This makes Oxlo.ai significantly cheaper for long-context and agentic workloads.

DeepSeek V4 Flash offers a 1M context window for near state-of-the-art open-source reasoning, while Kimi K2.6 provides advanced reasoning and agentic coding with a 131K context window. Both are available without cold starts.

Code Generation for Visualization

Beyond SQL, LLMs can draft matplotlib, seaborn, or Plotly code from a description of the data. You can pair a generalist model with Oxlo.ai's code-specific endpoints, such as Qwen 3 Coder 30B or Oxlo.ai Coder Fast, to produce analysis scripts.

import os
from openai import OpenAI

client = OpenAI(base_url="https://api.oxlo.ai/v1", api_key=os.environ["OXLO_API_KEY"])

data_description = """
A pandas DataFrame named df with columns:
- date (datetime)
- revenue (float)
- cost (float)
"""

request = "Write a Python script that plots revenue and cost over time on the same line chart with a legend."

response = client.chat.completions.create(
    model="MODEL_ID",  # e.g., Qwen 3 Coder 30B or Kimi K2.6 on Oxlo.ai
    messages=[
        {"role": "system", "content": "You write concise, runnable Python scripts."},
        {"role": "user", "content": f"{data_description}\n{request}"}
    ]
)

print(response.choices[0].message.content)

Selecting a Model for Your Pipeline

Oxlo.ai hosts over 45 models across seven categories. For data analysis, the most relevant are:

DeepSeek R1 671B MoE: Deep reasoning and complex coding for statistical or multi-step analytical pipelines.
Kimi K2.6 / K2.5 / K2 Thinking: Advanced chain-of-thought reasoning, agentic coding, and vision for chart interpretation.
Qwen 3 32B: Multilingual reasoning and agent workflows for global datasets.
Llama 3.3 70B: General-purpose flagship for fast, balanced SQL and text generation.
GLM 5: 744B MoE suited for long-horizon agentic tasks that span many data sources.
Qwen 3 Coder 30B / Oxlo.ai Coder Fast: Dedicated code generation for scripts and notebooks.

Cost Efficiency for Data Workloads

Data analysis workloads are inherently long-context. A single request might include thousands of tokens from a table schema, sample rows, and system instructions. On token-based providers, this scales linearly with input size. Oxlo.ai uses request-based pricing, so the cost is one flat amount per API call regardless of prompt length.

For teams running daily batch analysis, agentic investigations, or embedded analytics with wide schemas, this model can be 10-100x cheaper than token-based alternatives. You can view the exact structure on the Oxlo.ai pricing page.

Conclusion

Integrating LLMs into data analysis does not require switching away from familiar tools. Because Oxlo.ai is fully OpenAI SDK compatible, you can point existing Python scripts to https://api.oxlo.ai/v1 and immediately access models optimized for SQL generation, structured extraction, and coding. With flat per-request pricing and no cold starts on popular models, Oxlo.ai is a strong, relevant option for data teams that want predictable costs and broad model choice.

DEV Community