Emtract Ingest

Coding agent

Command: claude-agent-acp
Model: Used for both the parse and the crawl. Passed to the coding agent as-is, so use its own model names (e.g. sonnet, opus), not a Requesty id.
Gateway: https://router.requesty.ai
Parse timeout: 60 minutes
Crawl timeout: 60 minutes

Set per deployment through ACP_AGENT_COMMAND, ACP_GATEWAY_URL, ACP_PARSE_TIMEOUT and ACP_CRAWL_TIMEOUT.

What to extract

The target audience are companies that need to comply with the regulation, not countries, governments or regulators.
Do not create entries that are only relevant to government entities, countries or regulators. Do not invent anything
that is not grounded in the text. Do not use the term undertaking, use company instead. In titles and descriptions, do not
cite articles, sections, paragraphs or lines; summarize the relevant information instead.

Every source_excerpt is a short verbatim quote (a phrase or single sentence, never the full paragraph), not HTML encoded.
Every *_source_reasoning explains why its field was set the way it was, mentioning the source reference and an excerpt
where possible, and stays empty when the field is not found in the text.

### Regulation metadata

Fill in the regulation's names, descriptions, reference number, jurisdiction, topic, instrument type, status and
dates, and name its versions. Set repealed_on only when the text states the date the regulation is or was repealed.

### Milestones

A milestone is a dated event in the regulation's lifecycle (adoption, entry into force, application start,
transposition deadline, repeal, or another notable date). Only add one when the text states its date, only when it is
relevant for companies rather than for countries, regulators or legislative bodies, and combine events that share
both date and type into one milestone. Dates must lie between 1900-01-01 and 2100-01-01, otherwise leave the date null.

### Rule sets and their conditions

A rule set set of QUALIFYING criteria: the regulation applies when all conditions of any one rule
set hold. Never add exclusions or exemptions as rule sets (a rule set without conditions always matches).
A rule set's label names only its criteria, never obligations, penalties, timing or years.

Resolve references to annexes into specific condition values instead of referring to the annex. Use the activities
and objects attributes as a pair: a "charging station operator" is activity operator and object charging station.

Operators: gt/gte/lt/lte/eq take exactly one value, between exactly two (low, high), included_in/not_included_in/contains_any
one or more, is_true/is_false none. Use the attribute's default unit unless the text states a different one. Keep the
threshold as written and let the operator carry strictness: "more than"/"exceeds" is gt, "at least" is gte, "less
than"/"below" is lt, "at most" is lte — never shift the value (gte 501) to fake a strict comparison.

An activities or objects value names a kind of actor or thing and never carries a quantity. A figure qualifying it (size,
capacity, mass, volume, turnover, headcount, share, date) moves into its own condition on a numeric or date attribute of
the same rule set and is never dropped: "combustion installations with a rated thermal input of 50 MW or more" is
objects: combustion installations, and rated_thermal_input gte 50 MW. Exception: company size expressed through combined
thresholds ("at least two of the three" of balance sheet total, net turnover, headcount) is one condition on
company_size_class.

Set effective_from/effective_until only when the text states the date the rule set starts or stops applying, and point
applies_to_milestone_index/expiring_milestone_index at the milestone with that date. Duties imposed on companies belong in
obligations, not in rule sets, even when the text states thresholds for them.

Prefer active attributes, fall back to proposed ones, and never use one that is deprecated or has a removal reason
(follow its guidance if it names replacements). Prefer combining existing attributes over adding a new one.

### Condition attribute values

For every condition on an attribute with values in the catalogue, list in attribute_values each value the condition
refers to, with the phrases from the condition that denote it in extracted_values (copied exactly as written).
- A value matches when it denotes the same concept; a synonym, translation or the law's own definition is fine.
- An activities value naming a role together with what it acts on ("manufacturer of packaging") matches the role, and
  the thing belongs in an objects condition of the same rule set; add that condition when it is missing.
- Numeric or date thresholds rarely match; only match when the attribute is genuinely categorical.
- Never use a value key that is not in the catalogue for that attribute.

### Obligations

An obligation is a duty imposed on companies (reporting, documentation, risk analysis, registration and similar),
even when the text states thresholds for it; thresholds and other qualifying criteria belong in rule sets. Point
rule_set_index at the rule set it applies to, or null when it applies to every company subject to the regulation.

### Measures

A measure is one concrete, actionable step a company takes towards fulfilling obligations (e.g. "Register with the
national competent authority"). obligation_indices lists every obligation the very same action fulfils, almost always
one. Break an obligation into as many measures as there are genuinely separate actions; a simple obligation may need
only one. Set due_date only when the text states a deadline for that action distinct from its obligation's frequency.

### Penalties

A penalty is a consequence imposed on a company for failing to comply (a fine, licence revocation, suspension,
criminal liability). Set responsible_authority when the text names the enforcing authority. Point rule_set_index at a
rule set only when the penalty applies only to companies matching it, otherwise null. Group similar penalties rather
than adding many.

Appended to the agent's fixed instructions (file layout, editing rules, entry shape). Keep one "### " section per step so each maps to a task. Leave empty to use the template.

Settings

Coding agent

What to extract