RAG system for cultural event recommendation

Python, RAG, LangChain, FAISS, Mistral AI, FastAPI, Docker, Generative AI, NLP, Ragas

RAG System for Cultural Event Recommendation

Context

As part of a generative AI project, I developed a Proof of Concept (POC) for a Retrieval-Augmented Generation (RAG) system to answer natural language questions about cultural events. The system combines semantic search in a vector database with a Large Language Model (LLM) to generate contextualized and reliable answers.

Objectives

• Build a complete RAG chain.
• Collect and index events from the OpenAgenda API.
• Implement semantic search with FAISS.
• Generate contextualized answers with an LLM.
• Expose the system via a REST API.
• Automatically evaluate the quality of generated answers.

Architecture

API OpenAgenda
        │
        ▼
Prétraitement des données
        │
        ▼
Découpage en chunks
        │
        ▼
Embeddings
(Hugging Face)
        │
        ▼
Base vectorielle FAISS
        │
        ▼
LangChain
        │
        ▼
LLM Mistral
        │
        ▼
FastAPI
        │
        ▼
Client

Technologies Used

• Python
• LangChain
• FAISS
• Hugging Face Embeddings
• Mistral AI
• FastAPI
• Docker
• Pandas
• Ragas
• Pytest

Achievements

Data Collection and Preparation

• Event retrieval via the OpenAgenda API.
• Data cleaning and structuring.
• Filtering by period and location.
• Document chunking.
• Metadata preservation (date, location, category, description).

Vector Search

• Embedding generation with a Hugging Face model.
• Construction of a FAISS vector index.
• Semantic search by similarity.
• Automatic index reconstruction.

Augmented Generation (RAG)

• Orchestration of the RAG chain with LangChain.
• Search for relevant documents.
• Context injection into the prompt.
• Answer generation with Mistral.
• Contextualized answers based on indexed events.

REST API

Development of an API with FastAPI including:

• /ask endpoint
• /rebuild endpoint
• Swagger/OpenAPI documentation
• Request validation
• Error handling

Evaluation

Implementation of a system evaluation strategy:

• Annotated question set
• Unit tests
• Automation of evaluations
• Metrics with Ragas
• Analysis of answer relevance and faithfulness

Containerization

• Dockerfile
• Local deployment
• Environment reproducibility
• Full demonstration via Docker container

Demonstrated Skills

Generative AI

• Retrieval-Augmented Generation (RAG)
• Prompt Engineering
• LLM
• Context Augmentation

NLP

• Embeddings
• Semantic Search
• Document Chunking
• Vector Similarity

Backend Development

• FastAPI
• REST API
• Swagger / OpenAPI
• Request Validation

Data Engineering

• Data Collection via API
• Preprocessing
• Metadata Management
• Vector Indexing

MLOps

• Docker
• Automated Tests
• RAG System Evaluation
• Reproducibility
• Industrialization of an AI POC

Outcome

This project allowed me to design a complete RAG system, from data collection to the generation of LLM-augmented answers. It illustrates the integration of modern technologies such as LangChain, FAISS, FastAPI, and Docker, while implementing a performance evaluation approach using Ragas to ensure the quality of the generated answers.

Skills Acquired

• Python
• LangChain
• RAG
• FAISS
• Mistral AI
• Hugging Face Embeddings
• FastAPI
• Docker
• REST API
• Prompt Engineering
• Vector Database
• Semantic Search
• NLP
• Ragas
• Pytest
• OpenAPI
• Data Engineering
• Generative AI