---
schema_version: "newruntime-agent-readable-v0.2"
type: "disappearing_action"
stable_id: "disappearing_action:agents-can-define-document-extraction-schemas-via-natural-language"
slug: "agents-can-define-document-extraction-schemas-via-natural-language"
title: "Conversational document-schema creation is already a general capability delta"
description: "Retired delta: one vendor's paired announcements do not yet demonstrate a reusable cross-product capability change."
retrieval_nugget: "Retired delta: one vendor's paired announcements do not yet demonstrate a reusable cross-product capability change. Human ritual: Promote a same-vendor launch and duplicate signal pair into a durable, general capability delta. Old cost: The record overstates portability and user impact while hiding schema complexity, validation, error rates, and vendor coupling."
status: "published"
human_ritual: "Promote a same-vendor launch and duplicate signal pair into a durable, general capability delta."
old_cost: "The record overstates portability and user impact while hiding schema complexity, validation, error rates, and vendor coupling."
replacement_pattern: "Keep the launch as a source-inspected signal; publish a delta only when independent implementations demonstrate the same bounded capability with inspectable reliability limits."
adoption_status: "challenged"
where_old_interface_wins: "Hand-authored schemas remain clearer when fields are stable, compliance matters, or a deterministic validation contract is required."
first_seen: "2026-09-28"
last_verified: "2026-09-30"
record_date: "2026-09-30"
date_kind: "last_verified"
related_shift: "generated-interface-from-answer"
proof_projects: []
related_posts: []
topics: ["document-processing","capability-delta"]
source_urls: ["https://www.llamaindex.ai/blog/ai-document-extraction-schema","https://x.com/jerryjliu0/status/2092663785542783161"]
featured: false
routes: {"html":"https://newruntime.com/graveyard/agents-can-define-document-extraction-schemas-via-natural-language/","markdown":"https://newruntime.com/graveyard/agents-can-define-document-extraction-schemas-via-natural-language.md","json":"https://newruntime.com/graveyard/agents-can-define-document-extraction-schemas-via-natural-language.json"}
source_format: "markdown"
---

# Conversational document-schema creation is already a general capability delta

## Retrieval answer

Retired delta: one vendor's paired announcements do not yet demonstrate a reusable cross-product capability change. Human ritual: Promote a same-vendor launch and duplicate signal pair into a durable, general capability delta. Old cost: The record overstates portability and user impact while hiding schema complexity, validation, error rates, and vendor coupling.

## Why this delta is retired

Natural-language schema creation is an appealing capability claim: describe the
fields needed from a document and let an agent construct an extraction setup.
The evidence here does not yet establish that it is a general, reusable
capability. It is one vendor announcement and closely related communication
from the same source family, not independent confirmation across products or
workflows.

The difference matters because a schema is not merely a list of field names.
It contains validation rules, optionality, normalization, provenance,
confidence thresholds, exception routing, and consequences for downstream
systems. A demonstration can make the first draft easier while leaving the
hardest reliability work unchanged.

## Better unit of observation

Keep the underlying release as a source-inspected signal. Ask a bounded
question instead: for a stated document class and field set, can the system
produce a schema that validates, exposes evidence for each extracted value,
reports errors, and survives an independent replay? If the answer is yes in
multiple unrelated implementations, that can support a future Delta.

## Where the old interface wins

Hand-authored schemas remain clearer for stable documents, compliance-sensitive
fields, and deterministic integration contracts. They let a team specify
validation before extraction and diagnose a failed record without interpreting
an agent's explanation. Conversational assistance may speed schema design, but
it does not remove the need to own those rules.

## Reopen condition

Reopen only after independent tools demonstrate the same bounded workflow with
explicit validation and error evidence, a reproducible task boundary, and
counterexamples that reveal where the capability fails. Until then, calling it
a general Delta would confuse a promising launch with verified portability.

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- related record: [Answer -> generated interface](https://newruntime.com/shifts/generated-interface-from-answer.md): Shift replacing this disappearing action.
- related record: [Hand-crafted OCR pipelines are already being replaced](https://newruntime.com/graveyard/document-processing-shifts-from-hand-crafted-ocr-pipelines-to-llm.md): Shares document processing.
- related record: [Reading every changelog bullet to understand capability](https://newruntime.com/graveyard/changelog-bullet-reading.md): Continue with a related New Runtime material.
- related record: [Manual context handoff between AI clients](https://newruntime.com/graveyard/manual-context-handoff.md): Continue with a related New Runtime material.
- related record: [ABBEL Treats Memory as an Explicit Belief State](https://newruntime.com/posts/abbel-belief-state-memory.md): Continue with a recent New Runtime material.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)