mloda.ai: Open Data Access for ML & AI

Declarative data access for AI agents. Describe what you need - mloda delivers it.

pip install mloda

30-Second Example

Your AI describes what it needs. mloda figures out how to get it:

from mloda.user import PluginLoader, mloda
PluginLoader.all()

result = mloda.run_all(
    features=["customer_id", "income", "income__sum_aggr", "age__avg_aggr"],
    compute_frameworks=["PandasDataFrame"],
    api_data={"SampleData": {
        "customer_id": ["C001", "C002", "C003", "C004", "C005"],
        "age": [25, 35, 45, 30, 50],
        "income": [50000, 75000, 90000, 60000, 85000]
    }}
)

Copy, paste, run. mloda resolves dependencies, chains plugins, delivers data.

---

What mloda Does

┌─────────────────────────────────────────────────────────────────┐
│                      DATA USERS                                 │
│  AI Agents  •  ML Pipelines  •  Data Science  •  Analytics      │
└───────────────────────────┬─────────────────────────────────────┘
                            │ describe what they need
                            ▼
                    ┌───────────────┐
                    │     mloda     │  ← resolves HOW from WHAT
                    │   [Plugins]   │
                    └───────────────┘
                            │ delivers trusted data
                            ▼
┌─────────────────────────────────────────────────────────────────┐
│                     DATA SOURCES                                │
│  Databases  •  APIs  •  Files  •  Any source via plugins        │
└─────────────────────────────────────────────────────────────────┘

---

Why mloda?

You want to...	mloda gives you...
Give AI agents data access	Declarative API - agents describe WHAT, not HOW
Trace every result	Built-in lineage back to source
Reuse across projects	Plugins work anywhere - notebook to production
Mix data sources	One interface for DBs, APIs, files, anything

---

AI Use Case: LLM Tool Function

Let LLMs request data without writing code:

# LLM generates this JSON
llm_request = '["customer_id", {"name": "income__sum_aggr"}]'

# mloda executes it
from mloda.user import load_features_from_config
features = load_features_from_config(llm_request, format="json")
result = mloda.run_all(
    features=features,
    compute_frameworks=["PandasDataFrame"],
    api_data={"SampleData": {"customer_id": ["C001", "C002"], "income": [50000, 75000]}}
)

More patterns: Context Window Assembly • RAG Pipelines

---

How mloda is Different

mloda separates WHAT you need from HOW to get it - through plugins. Existing tools solve parts of this, but none bridge the full gap:

Category	Products	What it does	Why it's not enough
Feature Stores	Feast, Tecton, Featureform	Store + serve features	Infrastructure-tied, storage-only
Semantic Layers	dbt Semantic Layer, Cube	Declarative metrics	SQL-only, centralized
DAG Frameworks	Hamilton, Kedro	Dataflows as code	Function-first, no plugin abstraction
Data Catalogs	DataHub, Atlan	Metadata & discovery	No execution, no contracts
ORMs	SQLAlchemy, Django ORM	Database abstraction	Single database, no ML lifecycle

mloda is the connection layer - separating WHAT you compute from HOW you compute it. Plugins define transformations. Users describe requirements. mloda resolves the pipeline.

---

Plugins: The Building Blocks

mloda's architecture follows three roles: providers (define plugins), users (access data), and stewards (govern execution). The module structure reflects this: mloda.provider, mloda.user, mloda.steward.

mloda uses three types of plugins:

Type	What it does
FeatureGroup	Defines data transformations
ComputeFramework	Execution backend (Pandas, Spark, etc.)
Extender	Hooks for logging, validation, monitoring

Most of the time, you'll work with FeatureGroups - Python classes that define how to access and transform data (see Quick Example above).

Why plugins?

• Steps, not pipelines - Build transformations. mloda wires them together.
• Small and testable - Each plugin is a focused unit. Easy to test, easy to debug.
• AI-friendly - Small, template-like structures. Let AI generate plugins for you.
• Share what isn't secret - Your pipeline runs steps a,b,c,d. Steps b,c,d have no proprietary logic? Share them across projects, teams, even organizations.
• Experiment to production - Same plugins in your notebook and your cluster. No rewrite.
• Stand on shoulders - Combine community plugins with your own. Build on what exists.

---

AI Use Case Patterns

1. LLM Tool Function

Give LLMs deterministic data access - they declare what, mloda handles how:

from mloda.user import PluginLoader, load_features_from_config, mloda
PluginLoader.all()

# LLM generates this JSON (no Python code needed)
llm_output = '''
[
    "customer_id",
    {"name": "income__sum_aggr"},
    {"name": "age__avg_aggr"},
    {"name": "total_spend", "options": {"aggregation_type": "sum", "in_features": "income"}}
]
'''

# mloda parses JSON into Feature objects
features = load_features_from_config(llm_output, format="json")

result = mloda.run_all(
    features=features,
    compute_frameworks=["PandasDataFrame"],
    api_data={"SampleData": {
        "customer_id": ["C001", "C002", "C003"],
        "income": [50000, 75000, 90000],
        "age": [25, 35, 45]
    }}
)

LLM-friendly: The agent only declares what it needs - mloda handles the rest.

2. Context Window Assembly

Gather context from multiple sources declaratively - mloda validates and delivers. Why not let an AI agent do it?

Example: This shows the API pattern. Requires custom FeatureGroup implementations for your data sources.

from mloda.user import Feature, mloda

# Build complete context from multiple sources
features = [
    Feature(name="system_instructions", options={"template": "support_agent"}),
    Feature(name="user_profile", options={"user_id": user_id, "include_preferences": True}),
    Feature(name="knowledge_base", options={"query": user_query, "top_k": 5}),
    Feature(name="conversation_history", options={"limit": 20, "summarize_old": True}),
    Feature(name="available_tools", options={"category": "customer_service"}),
    Feature(name="output_format", options={"format": "markdown", "max_length": 500}),
]

result = mloda.run_all(
    features=features,
    compute_frameworks=["PythonDictFramework"],
    api_data={"UserQuery": {"query": [user_query]}}
)

# Each feature resolved via its plugin, validated

3. RAG with Feature Chaining

Build RAG pipelines declaratively - mloda chains the steps for you.

Example: This shows the chaining syntax. Requires custom FeatureGroup implementations for retrieval and processing.

# String-based chaining: query -> validate -> retrieve -> redact
Feature(name="user_query__injection_checked__retrieved__pii_redacted")

# Configuration-based chaining: explicit pipeline
Feature(
    name="safe_context",
    options=Options(context={
        "in_features": "documents__retrieved__pii_redacted",
        "redact_types": ["email", "phone", "ssn"]
    })
)

mloda resolves the full chain - you declare the end result, not the steps.

Automatic dependency resolution: You only declare what you need. If pii_redacted depends on retrieved which depends on documents, just ask for pii_redacted - mloda traces back and resolves the full chain.

---

Compute Frameworks

Mix multiple backends in a single pipeline - mloda routes each feature to the right framework:

result = mloda.run_all(
    features=[...],
    compute_frameworks=["PandasDataFrame", "PolarsDataFrame", "SparkFramework"]
)

# Results may come from different frameworks based on plugin compatibility

Add your own frameworks - mloda is extensible.

---

Extenders

Wrap plugin execution for logging, validation, or lineage tracking:

import time
from mloda.steward import Extender, ExtenderHook

class LogExecutionTime(Extender):
    def wraps(self):
        return {ExtenderHook.FEATURE_GROUP_CALCULATE_FEATURE}

    def __call__(self, func, *args, **kwargs):
        start = time.time()
        result = func(*args, **kwargs)
        print(f"Took {time.time() - start:.2f}s")
        return result

# Use it
result = mloda.run_all(features, function_extender={LogExecutionTime()})

Built-in and custom extenders give you full lineage - trace any result back to its source.

---

When to Use mloda

Use mloda when:

• Your agents need data from multiple sources
• You want consistent, validated data access
• You need traceability (audit, debugging)
• Multiple agents share the same data patterns

Don't use mloda for:

• Single database, simple queries → use an ORM
• One-off scripts → just write the code
• Real-time streaming (<5ms) → use Kafka/Flink

---

Documentation

• Getting Started - Installation and first steps
• Plugin Development - Build your own plugins
• API Reference - Complete API docs

---

Ecosystem

Most plugins currently live in mloda_plugins/ within this repository. The goal is to gradually migrate them to standalone packages in the registry.

Repository	Description
mloda-registry	Official plugin packages and 40+ development guides
mloda-plugin-template	Cookiecutter template for creating standalone plugins

---

Contributing

We welcome contributions! Build plugins, improve docs, or add features.

• GitHub Issues - Report bugs or request features
• Development Guide - How to contribute