mloda

Give an AI agent declarative, deterministic data access via mloda (https://github.com/mloda-ai/mloda) - a Python plugin framework where the caller describes WHAT features or data it needs and mloda re

SKILL.md

GitHub

mloda: Declarative Data Access for AI Agents

Check the registry index first

Always do this before hand-writing a chain step. mloda-registry's community set covers common {col}__{op} transforms: aggregation, window/scalar/frame aggregate, scalar/point arithmetic, rank, offset, percentile, binning, datetime, string ops, time bucketization, ffill, ema, sessionization, resample. None of it ships with plain pip install mloda; install what you need, e.g. pip install mloda-community (all) or pip install "mloda-community-rank[pandas]" (one). See its plugins table for the full list and feature-name patterns.

The table drifts, so once installed, confirm what's actually loaded rather than trusting a static list:

from mloda.user import PluginLoader
from mloda.steward import get_feature_group_docs

PluginLoader.all()
for fg in get_feature_group_docs():
    print(fg.name, fg.description)

This only reflects installed packages: an empty or unrelated result means the plugin isn't installed yet, not that it doesn't exist.

Nothing covering it? See the mloda-plugins skill for how to write a FeatureGroup, ComputeFramework, or Extender.

Mental model

mloda separates WHAT a caller needs from HOW it gets computed. A caller (human or LLM) requests features by name; mloda resolves dependencies across a plugin graph and executes them on a compute framework (PandasDataFrame, PyArrowTable, PythonDictFramework, ...), returning the result with lineage back to source.

Three plugin types, most work happens in the first:

Type Role
FeatureGroup Defines a data transformation - the unit you'll write most often
ComputeFramework Execution backend the transformation runs on
Extender Hooks for logging, validation, monitoring

Roles map to modules: mloda.provider (define plugins), mloda.user (request data), mloda.steward (govern execution).

When to reach for mloda

  • The agent should declare a data/feature request as a tool call instead of writing ad hoc fetch/transform code (see "LLM Tool Function" below).
  • Multiple context sources (system prompt, user profile, retrieved docs, history) need declarative assembly before being handed to a model.
  • A RAG pipeline needs chained steps (retrieve -> validate -> redact) expressed as a feature name or config instead of hand-wired code.
  • You need to check whether a requested feature is already covered by a plugin before writing new code (see "Check the registry index first" above); to write one instead, see the mloda-plugins skill.

LLM Tool Function pattern

The agent emits a feature request as JSON, no Python required from the agent itself:

from mloda.user import PluginLoader, load_features_from_config, mloda

PluginLoader.all()  # every mloda example calls this once before the first run_all()

llm_output = '''
[
    "customer_id",
    {"name": "income__sum_aggr"},
    {"name": "total_spend", "options": {"aggregation_type": "sum", "in_features": "income"}}
]
'''

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"], "income": [50000, 75000]}},
)

api_data inlines data under a label (e.g. "SampleData"); features are matched to columns by name. For data on disk, pass a DataAccessCollection instead (see mloda's API Request docs).

Feature-name chaining

A chain like documents__retrieved__pii_redacted expresses a pipeline without wiring it by hand - if you ask for pii_redacted, mloda traces that it depends on retrieved, which depends on documents, and resolves the whole chain from the single requested name:

Feature(name="user_query__injection_checked__retrieved__pii_redacted")

This only works when a FeatureGroup plugin exists for every step in the chain - the chaining syntax alone does not fetch data mloda has no plugin for.

Latency-sensitive / repeated requests

mloda.run_all() rebuilds the full execution plan every call, fine for batch jobs but wasteful when the same features are recomputed per request (model serving, streaming, dashboards). Use the two-phase API to plan once and execute cheaply per request:

session = mloda.prepare(features, compute_frameworks=["PyArrowTable"], data_access_collection=data_access_collection)
result = session.run(api_data={"MyKey": {"col": [1, 2]}})

Reference