Organize documents into searchable units
Long documents are split into clauses and context units, together with the source, language and other details needed for search.
CONVOFIN LEARN · TECHNOLOGY
With RAG and machine learning,
we connect financial documents to their evidence.
Technology that finds what you need in financial documents, classifies what each clause means, and verifies the results. Meet ConvoFin Learn.
This page covers how document search and model training and validation work, and how they could extend to finding financial products that match customer needs.
Explore applications01 / RETRIEVAL-AUGMENTED GENERATION
What is RAG? Finding the evidence behind an explanation
RAG (retrieval-augmented generation) first searches for material related to a question, then generates an explanation that draws on what it found. It suits information that has a source text to check, such as terms and conditions or product descriptions.
Search for evidence related to the question
It finds clauses related to the input sentence. Even when similar words appear, the conditions and surrounding context must be checked.
Explain with reference to the evidence
The material found is linked to the explanation as a reference. Generated explanations should be compared with the source, and anything with weak evidence checked separately.
ConvoFin Learn implements document organization, search and search-result review. The flow above explains how RAG works; how much of it applies to a given analysis depends on each feature's settings and validation status.
02 / MACHINE LEARNING
Learning the features of financial clauses with machine learning
Machine learning learns patterns from example data to classify or predict new input. ConvoFin Learn can train and evaluate a classification model on reviewed financial clauses.
Classifying clauses into fees, cancellation, obligations, losses and so on helps organize what to check. A classification is not a legal judgment on the clause or a statement of suitability for any customer.
From training to validation
- Prepare reviewed examplesClause labels are reviewed and the training material is organized.
- Split training, validation and test dataRelated documents are kept apart so that evaluation results are not inflated.
- Model evaluation and versioningResults and errors are reviewed, and whether to deploy is decided by validation status.
How do RAG and machine learning differ?
RAG asks “what material should this explanation rest on?” while machine learning asks “what kind of clause is this sentence?” Search can find context and learning can help classify, but each result must be verified on its own.
03 / APPLICATIONS
Applications that connect customer needs with financial documents
Directions this could grow in, built on document search and clause classification.
Planned use
Finding financial products that match customer requirements
The customer's preferred product type, purpose, amount, term, interest conditions and repayment method are organized, then candidates that match are found in product descriptions and terms.
Machine learning can classify documents and clause types, and RAG can explain a candidate product by linking its conditions to the supporting sentences.
Example: turning a customer need into search conditions
“I want to repay the same amount every month,
and compare what it costs to pay off early.”
- Conditions
- Repayment method · rate type · early repayment terms
- Candidates
- Search product documents that state those conditions
- Evidence
- Check fees, waivers and where they appear in the source
A hypothetical example. Not a real product list or search result.
Offering real product search would require obtaining and updating current product data and verifying conditions. A match on paper does not guarantee eligibility, approval or personal suitability.
Understanding financial documents
Find clauses on costs, cancellation and obligations in insurance terms, loan product descriptions and card notices, and organize what to check.
Financial document analysis guide →Support for institutional counseling
Search for product explanations and supporting terms during a consultation, and organize conditions the counselor should check further.
Support for financial education
Organize document examples by type to learn easily missed points such as fees, cancellation terms and customer obligations.
Browse financial terms (Korean) →04 / QUALITY & SCOPE
Both the evidence found and the learned results need verification
Source and effective date
Even products with the same name can have different conditions depending on the document date and who it applies to. Check that the retrieved text applies to your situation.
Evaluation data and errors
We need to check whether validation data reflects real use. This page does not state unverified accuracy figures or promise performance.
Permitted use of data
Data used for document search and model training must be checked for usage rights and personal information. See our Privacy Policy for how the service handles personal data.
Do you recommend or broker financial products?
This page describes using customer conditions for document search. We do not recommend or broker specific products, or provide loan screening results.