How it works

Each portal reads one knowledge box: a private, searchable collection of your documents on the Progress Agentic RAG platform. What people can search, ask and browse is what is in that collection.

CorpusKit reads a copy of your documents. The originals stay where they are.

Text and tables are extracted page by page, so every passage can be traced to where it sits in its document. Each document and its passages are indexed for search by meaning and by exact words, and the names, places and topics the documents mention go into the knowledge graph.

On a Pro licence, CorpusKit also writes a readable title, summary and key takeaways for each document, and labels it by topic.

The portal reads the question and chooses how to search for it: an exact lookup for an identifier, newest first for a question about what is recent, and the full text for a broad question. The choice is shown beside the answer, and you can change it and ask again.

The answer is written only from the passages found. Each sentence that states a finding carries a numbered citation, and opening it shows the passage in the document, at the page used.

Before the answer is shown, figures, populations, named studies and years are tested against the cited text, sentence by sentence. The check is a plain text comparison against the extracted text of your documents, with no language model in the loop.

A sentence that fails is removed, and a note under the answer says so. Any conclusion that rested on a removed sentence goes with it.

Trust, citations and confidence

Each answer is rated high, moderate or low confidence, led by that check. High is given only when every figure in the answer was found in the cited text.

The answer appears as it is drafted, marked as unchecked. The check runs in view, and the checked answer then replaces the draft.

When the collection holds nothing that answers a question, the portal says so and shows the closest passages it found. In Generate, sections the collection cannot support are left out and named.

Each portal has an MCP connector. A portal administrator creates a key, and any MCP-capable AI assistant or agent can then search the collection, ask for cited answers, fetch a document and browse the catalogue, with the same retrieval the portal uses.

The connector is read-only, and each key reaches one portal and can be revoked at any time.

Connecting an AI assistant

CorpusKit runs on Progress Agentic RAG, which provides the knowledge box, retrieval, answer generation and source attribution. CorpusKit adds the numbered citations, the question routing, the answer check and the reading tools: the reader, investigations, Generate, knowledge checks and saved searches.

The code is public at github.com/noicework/corpuskit, and the whole process is described in the CorpusKit documentation.