What each system receives, what it returns, and what it guarantees. The method
used to get there is not public.
Scraper
In
An organisation and the address of its website.
Out
A package: the organisation profile, a list of services, and for each
service a verbatim quote with the address of the page carrying it.
Guarantee
Every service returned carries a quote of at least eight words,
present literally on a page belonging to the organisation. Comparison ignores whitespace,
invisible characters, and typographic variants of the same character. A service without a
quote is never returned.
Limits
A site that refuses automated requests produces no package. The failure
is recorded as a failure, never filled in with a guess.
Not public
The crawl strategy, the extraction, the coding of services, and the
detail of the internal checks.
RAG
In
A question in plain language.
Out
Relevant passages, each with the organisation and the address it came
from.
Guarantee
No passage without its source. The system answers from the content of
the packages; it does not fill gaps with general knowledge.
Limits
What is not in a package does not exist for it. An empty answer is a
valid answer.
Not public
Chunking, vector representation, retrieval and ranking.
Chatbot
In
A message, sent to an HTTP endpoint.
Out
A written answer with the sources that support it.
Guarantee
The sources shown are the ones the RAG returned, with nothing
added.
Not public
The orchestration and the instructions given to the model.
Public directory
In
The packages.
Out
Static HTML pages, with no scripts and no network calls.
Guarantee
Nothing is displayed without its quote and its link. The build fails
rather than publish an exception.
The pivot format
A package is one folder per organisation. Downstream consumers read only these two
files:
organization_profile.json identity, address, contacts, mission
services_programs.json list of services; per service:
canonical_name the service name
description what it does
grounding_quote the verbatim quote
source_urls where the quote lives
service_classifications its categories
review_decision the review state
A service whose decision begins NOT a SourceLink service was set
aside deliberately, with its reason written down. It stays in the package: a recorded exclusion
is worth more than a silent deletion.