Fenic

Fenic is an open source Python library that puts language model inference inside the DataFrame model. Extraction, classification, embedding and semantic joins are expressed as column operations, so messy text becomes typed rows through the same API a developer would use for ordinary structured data. It is built on a Polars based engine with Rust extensions and installs from the Python package index for Python 3.10 and later. It suits data and AI engineers who want pipelines that can be inspected and rerun rather than one off scripts. Queries combine meaning and metadata in a single expression, and the engine exposes lineage, query plans and cost, so a run can be explained after the fact. Plans are cached, which means a rerun repeats only what changed instead of recomputing everything. A finished pipeline can be promoted to a table or published as an MCP tool, which lets agents call the same logic that an engineer runs interactively. The library is free and open source, and no pricing, paid tier or hosted service is published alongside it. Running costs are whatever the underlying model providers charge for the inference a pipeline triggers, which is the reason the engine reports cost per run. The project is unusually explicit about its boundaries. It describes itself as a semantic DataFrame engine and not an agent framework, orchestration runtime, memory store, vector database or metrics layer for business intelligence, and it makes no claim to distributed, streaming or real time processing.