Aghiles Chougar

Intervention with the SOS group to build a RAG to automatically respond to project funding calls for tender

Answering complex questions, requiring the gathering, analysis, and summarization of concepts scattered across multiple pages

1. Naive RAG based on Langchain

2. RAG with grid search:

- Chunks: Variable size from 500 to 6000 characters

- Embedding: 2 embedding models

- Databases: Chroma and Faiss

- Retriever: Semantic retriever with scoring, minimum relevance threshold, and up to 10 chunks returned per query

3. Advanced RAG:

- Chunk: Optimal size of 500 characters

- Database: Faiss

- Hybrid Retriever:

* Semantic retriever, TFIDF, and lexical matching with n-grams (1-2)

- Rerankers:

* ms-marco-MiniLM, BGE Gemma V2, gpt4o

4. Self-RAG:

Based on https://arxiv.org/abs/2310.11511

5. RAPTOR:

Steps:

1. Chunking

2. Clustering

3. Reconstructing chunks into a tree structure (organization by main themes, secondary themes, etc.)

Source: https://arxiv.org/abs/2401.18059

Knowledge graph RAG

- Principle: Structuring information as a knowledge graph

- Advantage: Optimal linking of different concepts to improve response relevance and accuracy, especially for open-ended questions requiring multi-step reasoning (multi-hop reasoning)

- Based on a research paper from 02-2025: https://arxiv.org/abs/2502.14902

Context and problem:

A large and difficult-to-explore supply of housing for sale: 1 million transactions in 2021 More than 500k properties for sale on seloger.com Listings spread across thousands of municipalities Little context on property location provided by portals: Evolution and composition of prices by municipality/neighborhoods/types of housing Socio-economic situation of municipalities/neighborhoods Location and proximity of shops and services The property supply does not consider the profile and needs of buyers: Workplace, family situation, profession, hobbies...

Solution:

Enable property location analysis with: a municipality and neighborhood comparator detailed analysis of property context through several sections and dozens of indicators a summary and a score for each location studied

Data used:

Real estate listings collected weekly 6 million listings, since 02/2019 INSEE data (housing, population, activity...)

Sitadel database (new constructions)

Permanent equipment database (shops and services)

Crime and offense database (Ministry of the Interior)

Other sources to be exploited in the future: Education, internet, transport, air quality, natural and industrial risks, climate

Context: Legal merger of Pernod and Ricard entities, redesign of sales organizations, redesign of the management model, analytical accounting plan, and customer hierarchy, change of tools (ETL and data visualization)

Work:

• Assistance in defining the business and IT target (planning, roles, responsibilities, stakeholders...)

• Formalization of processes and rules for producing reporting/financial analysis information

• Impact analysis of organizational and accounting redesigns on the BI production process

• Drafting of functional requirements for the development of new BI solutions

• Preparation and facilitation of requirements gathering workshops (Agile Scrum principles)