CRECO: an agentic RAG for credit analysts
Analysts query years of credit assessment memorandums and committee minutes in natural language. Every answer cites its sources, down to the page.
"Have we already financed a comparable deal, and on what terms?"
A simple question for a credit analyst. But the answer is buried in hundreds of documents nobody can reread before every committee.
Credit assessment memorandums of 40 to 80 pages, in PDF, filed by deal.
Committee minutes, where the decisions and the conditions actually imposed are recorded.
Institutional memory, which leaves with experienced analysts.
An agentic RAG, not just a search engine
Instead of returning a list of documents, CRECO does the reading: it plans the search, assigns each source to an agent, then checks every citation before answering.
Understand
The question is rephrased and broken down into precise searches.
Search
Specialised agents query each type of document in parallel.
Read
Relevant passages are extracted, with their page and date.
Verify
Every statement is checked against its source; otherwise it is removed.
Answer
A short answer, clickable citations and the documents to open.
Designed for the analyst, not the IT team
The wireframes set one rule: the analyst must be able to check every answer in one click.
- 1
A single search bar
The analyst asks a question as they would ask a colleague, in French or English.
- 2
The answer, highlighted
Sentences taken from the documents are highlighted, each with its source number.
- 3
Sources alongside
Each citation opens the document at the right page. Nothing is stated without proof.
- 4
Suggested next steps
Compare two deals, list the conditions imposed, prepare a note for the committee.
Built for a bank's most sensitive data
Access rights respected
An analyst only sees in the answers the files they are allowed to consult.
Data stays at the bank
Documents, indexes and models remain inside the bank's information system.
Open source models
Locally hosted models, with no data sent to an external service.
Traceability
Every question, every source consulted and every answer is logged.
No answer without a source
If no document supports an answer, CRECO says so instead of making one up.
Continuously evaluated
A set of reference questions measures answer quality at every change.
Frequently asked questions
What is an agentic RAG?
A system that finds the right passages in your documents, then has AI agents read and check them before answering. Unlike a standard chatbot, it cites its sources.
Can CRECO be wrong?
It may not find an answer. But every statement is linked to a document: the analyst checks in one click, and CRECO says when no source answers.
Do the documents leave the bank?
No. Indexing, models and answers all run inside the bank's information system.
Can it be adapted to other functions?
Yes. The same approach applies to compliance, executive committees or loan agreements: only the documents and the agents change.
Do your analysts spend more time searching than analysing? Let's talk about your documents and your recurring questions.
Book a scoping session ↗