structured-extraction

Public

Turn unstructured text into typed, validated JSON — sentiment, classification, entities, PII, invoices, summaries, addresses, and events. The flagship collection: every prompt ships a strict output schema and follows house conventions (reasoning-first, no hallucinated fields, uncalibrated confidence).

9 prompts

README

Structured Extraction

Turn unstructured text into typed, validated JSON — sentiment, classifications, entities, PII, invoices, summaries, addresses, and events. Every prompt ships a strict output schema, so the generated SDK gives you a real type and a parseOutput() you can trust, not a string you have to babysit.

This is the flagship Sufleur collection and a reference for the conventions we hold every official prompt to.

Prompts

Prompt	What it does
`sentiment-analysis`	Overall sentiment (positive / negative / neutral / mixed) with rationale + confidence.
`text-classification`	Assign text to a caller-supplied label set, with an optional fallback.
`entity-extraction`	NER for people, orgs, locations, dates, money, and more — each normalized.
`pii-detection`	Detect and redact PII; returns typed spans plus a redacted copy.
`key-information-extraction`	Generic "messy text → your fields" extractor.
`invoice-extraction`	Strongly-typed invoice/receipt parsing (a specialization of the above).
`structured-summarization`	`{title, key_points, action_items, entities_mentioned}` — not prose.
`address-parsing`	Unstructured address string → structured components + confidence.
`date-event-extraction`	Events with resolved ISO 8601 start/end times.

House conventions

Every prompt in this collection follows the same rules — they're what make the output dependable:

Reasoning before the verdict. Classification and scoring prompts put a reasoning/rationale field first and pin the key order in the system prompt. Reasoning that precedes the answer informs it; reasoning that follows just rationalizes it.
No hallucinated fields. Anything not present in the input comes back null or an empty array — never invented. This is stated explicitly in each prompt.
Confidence is uncalibrated. Where a confidence score is offered, the schema says plainly that LLM self-reported confidence is not a probability — use it as a rough signal for routing and review, nothing more.
Spans use matched-text + occurrence index. Span-based prompts (entity-extraction, pii-detection) locate each hit by its exact text and a 1-based occurrence count, never character offsets — which models produce unreliably.
Strict schemas. additionalProperties: false, complete required arrays, closed enums, and bounded numbers throughout.
Safe inputs. User content is injected with triple-mustache (so &, <, " are never HTML-escaped) and wrapped in XML-style delimiters for clarity and injection-resistance.
Parameterized, not hardcoded. Every prompt exposes at least one meaningful variable beyond the raw input (a label set, a type filter, a reference date, a locale hint).

A note on runtime-defined outputs

A couple of prompts take their value set at runtime (text-classification's labels, key-information-extraction's fields). Because the output schema is fixed per version, parseOutput() can't verify that a returned label/field is one you asked for — add a one-line membership check in your code. Each such prompt's page calls this out.

Prompts9 prompts

sentiment-analysis

Classify the overall sentiment of text (positive / negative / neutral / mixed) with a rationale and confidence score. Strict JSON output.

text-classification

Classify text into a caller-supplied label set (you define the labels at call time). Returns the single best-fit label with a rationale and confidence. Optional fallback label.

entity-extraction

Named-entity recognition: extract people, organizations, locations, dates, money, and more from text — each with a canonical normalized_value and a matched-text + occurrence-index span. Optionally restrict to a caller-supplied set of entity types.

Detect and redact personally identifiable information (PII) — emails, phone numbers, names, addresses, card and account numbers, and more. Returns each detected span plus a fully redacted copy of the text. Optionally restrict to a caller-supplied set of PII types.

key-information-extraction

Generic structured extraction: pull a caller-defined set of named fields out of any messy text — receipts, emails, forms, notes. You describe the fields; it returns each value (canonical string) or null when absent, with no fabrication. See invoice-extraction for a ready-made specialization.

invoice-extraction

Turn a raw invoice or receipt into structured JSON — vendor, invoice/due dates, currency, line items, subtotal, tax, and total. Dates normalized to ISO 8601, amounts to plain decimals. Optional locale hint disambiguates date and number formats. A worked, strongly-typed specialization of key-information-extraction.

structured-summarization

Condense text into a structured JSON summary — title, key points, action items, and entities mentioned — instead of prose. Built for meeting notes, email threads, articles, and agent handoffs. Optional cap on the number of key points.

address-parsing

Parse an unstructured address string into structured components — recipient, street, unit, city, state/region, postal code, and country. Optional country hint guides format expectations. Returns a confidence score.

date-event-extraction

Extract scheduled events from natural-language text with resolved ISO 8601 start/end times. Relative dates ("next Tuesday", "tomorrow at 3") are resolved against a reference date and IANA timezone you supply. Returns title, start, end, all-day flag, and location per event.

@sufleurstructured-extraction