pandas

pandas → .NET

Verdict: use it, accepting more roughness. The equivalent exists but is less mature and less ergonomic than pandas; expect some glue.

pandas need	Recommended .NET
General DataFrame, CSV IO, typed columns	Microsoft.Data.Analysis
Time series, rich indices	Deedle (F# origin, excellent at time series)

dotnet add package Microsoft.Data.Analysis

using Microsoft.Data.Analysis;

DataFrame df = DataFrame.LoadCsv("data.csv");
df["price"] = df["price"].Multiply(1.2);
DataFrame expensive = df.Filter(df["price"].ElementwiseGreaterThan(100));

Pitfalls

• groupby / pivot. Less complete and less fluent than pandas; sometimes it's simpler to group with LINQ over the columns.
• Index. No label index like pandas in Microsoft.Data.Analysis (positional indexing). Deedle is closer.
• Missing values. null/NaN handling differs from pandas; check column by column.
• Lodestar glue. There is no DataFrame ↔ sparse-matrix bridge and none is planned; the join is a LINQ expression, shown below.

Feeding a DataFrame column to a vectorizer

The vectorizers take IEnumerable<string>, so a column reaches them without anything in between. Both of these work — the typed column already enumerates as strings, and the untyped indexer needs a cast:

using Microsoft.Data.Analysis;
using Lodestar.Text.Vectorization;

DataFrame df = DataFrame.LoadCsv("reviews.csv");

// The column is typed: enumerate it directly.
IEnumerable<string> documents =
    ((StringDataFrameColumn)df["review"]).Select(v => v ?? string.Empty);

// Or from the untyped indexer, which enumerates as object.
IEnumerable<string> viaCast =
    df["review"].Cast<string?>().Select(v => v ?? string.Empty);

CsrMatrix counts = new TfidfVectorizer().FitTransform(documents);

Decide what an empty cell means before you write ?? string.Empty. pandas keeps NaN distinct from ""; a vectorizer cannot, and a missing review scored as an empty document is a row of zeros that looks like a legitimate answer. Filter the nulls out if that is what you mean.

Going back is reading, not converting: CsrMatrix exposes Values, ColumnIndices and RowPointers, and ToDense() if the shape is small enough to want a rectangle. Materialising a document-term matrix into a DataFrame is usually the wrong move — a vocabulary of thirty thousand terms is thirty thousand columns, almost all zero, which is exactly the density the sparse format exists to avoid.

Why there is no bridge package

The doctrine in CLAUDE.md is native code only where .NET has a real gap, and this is not one: both sides already exist and the join above is one expression. A bridge would also cost Lodestar.Text its no-third-party-dependency property, or need a fifth package to hold a Select.

Guide to be expanded as real needs arise.