Research: the graph behind our agents | Rakam AI

Research

A reliable agent starts with a graph.

Our graphs come from research carried out with Lemon Learning and published at ICMLC 2026.

The paper

Building Specialized Software-Assistant ChatBot with Graph-Based Retrieval-Augmented Generation

Mohammed Hilel, Yannis Karmim, Jean de Bodinat, Reda Sarehane, Antoine Gillon

Conference
ICMLC 2026
Subjects
Software engineering · Machine learning
First version
7 November 2025
Revised
14 May 2026

The context

Guiding a user through complex software.

Digital adoption platforms help employees use complex enterprise software: CRM, ERP, HRMS. Lemon Learning has shown that good guidance cuts training costs and speeds up onboarding. But writing and maintaining those guides is still manual work, screen by screen.

The problem

The model makes things up

Without a structured understanding of the target software, a large language model answers confidently, and sometimes wrongly.

It cannot be retrained

Production models are closed APIs: no access to the weights, so no fine-tuning.

The software changes

Every release moves screens and fields. A hand-written guide ages at the pace of updates.

The method

Turning the application into a state-action graph.

  1. 01

    Extract

    The web application is crawled and its interfaces extracted: screens, fields, buttons.

  2. 02

    Structure

    Each screen becomes a state, each gesture an action. The whole software becomes a knowledge graph.

  3. 03

    Retrieve

    For each question, the agent fetches the path in the graph that answers it, instead of guessing.

  4. 04

    Answer, grounded

    The model generates its answer from that path: contextual, checkable, nothing invented.

In production

It did not stay a paper.

The framework was integrated into Lemon Learning’s guidance workflows. The paper details the engineering pipeline, the design of graph retrieval and the lessons learned from deployment: scalability, robustness, industrial use cases.

Read the Lemon Learning case →

And in our agents

The same idea structures every Rakam agent: a workflow graph, named steps, sensitive cases handed to a human. A predictable agent, so a testable one.

See an agent’s anatomy →

In preparation

Two more papers.

Workflow graphs

How a work graph makes an agent predictable and measurable.

Ethical AI

Human oversight, traceability and user permissions in an agent that acts.

Read the full paper on arXiv ↗

See this graph on your software.

Request a demo

Fifteen minutes to see an AI agent working inside software like yours.