A Cloud-AI architecture for grassroots financial cooperatives

Cloud-AI Architecture for Grassroots Financial Cooperatives: Integrating Electronic Forms and RAG-Based Knowledge Retrieval

AICCC 2026 · Second author · Accepted, with an oral presentation in Tokyo this December

Background

Credit departments of farmers’ associations face a very practical problem. Senior staff are retiring, front-line headcount is shrinking, and a lot of operational knowledge lives only in those senior employees’ heads. When they leave, the knowledge leaves with them.

From our interviews we identified four front-line pain points:

• Not enough people: repetitive administrative work takes up too much staff time.
• Slow service: filling in forms, looking things up and confirming details all take time.
• Inconvenient service: work depends heavily on the counter process.
• Lack of standardisation: procedures and judgement are scattered across people and documents.

Research questions

1. How can front-line pain points and knowledge loss be turned into concrete system modules and functions?
2. How can electronic forms and RAG knowledge retrieval be integrated under one technical architecture, as a single service platform?
3. How can an evaluation framework be established before the system goes live, as a basis for verifying its effect later?

System architecture

• Cloud application service: personal data entered by the user is merged into the PDF in the browser only, and is not sent to the server.
• AI and knowledge retrieval: dense and sparse retrieval run in parallel, results are fused with RRF and reranked with a cross-encoder, with extra weighting by regulation level.
• Security and access control: encrypted transport, authentication, role-based access, department and business scope, audit logs and sensitive data filtering.

An honest limitation

From a knowledge management point of view, the system does two things: it codifies knowledge into documents (codification), and it uses technology to lower the barrier to finding that knowledge (technology-enabled retrieval).

It does not do the third: passing experience from senior staff to newcomers (mentoring). An AI knowledge base can preserve knowledge that can be written down, but it cannot replace a senior employee’s judgement in a complicated situation. We state this clearly in the paper.

My part

I was responsible for the architecture of RAG, electronic forms, backend services, permissions and data flow, and took part in writing and revising the technical content of the paper.