EntwineAgents.Intake 0.3.3

EntwineAgents.Intake

Turn messy sources into clean, model-safe input.

• RecordTableReader — one Table abstraction over XLSX, CSV, and OCR'd document text (parses the [TABLE] markdown emitted by EntwineAgents.Ocr), with tolerant header detection (banner rows above the real header are fine) and case/space-insensitive column lookup.
• DateNormalizer — messy real-world date columns to DateOnly + active/flag signals.
• PseudonymMap — session-local pseudonymisation: typed, stable tokens (ACCOUNT_01, PARTNER_02, ...) replace real identities before text reaches an LLM, and hydrate back for the human-facing output. The map lives only in memory and dies with the session.

var map = new PseudonymMap();
var safe = map.Scrub("Acme Corp renewal at Globex");   // identities out
var text = map.Hydrate(modelOutput);                    // identities back, for humans only

• TableShapeClassifier — read a table for what it is. Give it your target schemas (name, columns with synonyms, optional value-shape predicates) and it binds each schema's columns to the table's headers — exact name, synonym, whole-word containment, then the values' shape — and scores every schema; the best complete reading wins. That path is deterministic and free. Only the residue (nothing fits, or a tie) goes to a model through a ShapeResidueCall you supply, with the headers and a few sample rows you scrub; the answer is validated against your schemas and the table's headers, so it cannot invent a column or a header. ColumnMap.Project(table) re-heads the table with your canonical names; Describe() is the "Partner = Reseller · Customer = Account name" line to show a person before anything runs on it.

var engagements = new TargetSchema("Engagements", new[]
{
    SchemaColumn.Of("Partner",  required: true, "Partner name", "Reseller"),
    SchemaColumn.Of("Customer", required: true, "Account", "Client"),
    new SchemaColumn("Start", new[] { "Start date", "Kick-off" }, ValueShape: LooksLikeDates),
}, "one row per piece of partner work on a customer");

var shape = new TableShapeClassifier(new[] { engagements, sourcing, outcomes });
var result = await shape.ClassifyAsync(table,
    residue: (prompt, ct) => chat.CompleteAsync(prompt, ct),   // any chat client; null = deterministic only
    scrub: map.Scrub);                                          // sample rows leave pseudonymised
if (!result.IsUnknown)
    Console.WriteLine($"Read as {result.Best.Schema.Name}: {result.Best.Describe()}");
var canonical = result.Best?.Project(table);                    // headers are now your names

Part of EntwineAgents — a lean, composable agent runtime for .NET.