Jevframe

Row count drives it. The engine sends one input row per request with up to 16 in flight, so 100 rows with three questions produce 100 requests and 100 result rows, not one. The levers are an opt-in in-memory cache, selecting rows before evaluating, and evaluate() sharing a single request across all questions about a row.

It reshapes the output, never the input. Row order and pandas indexes are preserved including duplicates and MultiIndexes, the input frame is never mutated, choice and score return full distributions as p__ columns, evaluate prefixes fields by question name, and nothing is thresholded or silently renormalized.

No. A failed or skipped row leaves a missing result, None in pandas and null in Polars, and context errors always raise even when you pass errors='coerce'. EvaluationError.position and EvaluationWarning.failures point at the failing row without confusing duplicate index labels.

Not in this version. The README places LazyFrame expressions outside v0 and covers eager pandas and Polars only, where results are native Series and DataFrames and empty outputs keep their dtype, including the full Polars struct schema. Importing an integration registers its .jev accessor on the frame.

Install with uv, for example uv add 'jevframe[pandas]==0.1.0', and export TYPESAFE_API_KEY. Python 3.10 or newer is required, the official typesafe-sdk>=0.7,<0.8 handles inference, and dataframe dependencies are optional. The library itself needs your own credentials or an explicitly supplied client; the demo's sponsored fallback is configured only in that notebook.

Pass a column name, a list of column names, or a callable that receives a detached mapping of all columns and returns text, an object, or an array. Indexes are never sent implicitly, missing scalars become JSON null, dates and times become ISO text, and Polars temporal values stay ISO strings to keep nanosecond precision. The callable runs once per row before the cache lookup.

Every method takes max_concurrency, default 16 simultaneous row evaluations including retries, an optional reusable in-memory cache, a progress flag, errors set to raise or coerce, and a nulls policy of include, skip, or raise that applies after the context is built. Progress snapshots expose completed, total, failed, skipped, and cache_hits.