INDEXATION VECTOR DB QUERY RETRIEVAL GENERATION
RAG - Retrieval Augmented Generation

AI that answers from your documents.

Ask a question and get the answer from your own documents, with a link to the source. That is how RAG works: the system finds the right passages in your knowledge base, then a language model writes the answer.

Let's connect AI to your documents

≈ 1 MIN

Your need

Which document sources?

You

Company size (optional)

Your details are only used to get back to you.

  • Reply within 1 business day
  • Free 30-minute scoping call
  • You talk to the person who builds the project
Since 2017 of automation, data and AI work for large companies: food, energy, spirits, media, retail.

Companies we have worked for

  • Disneyland Paris
  • Pernod Ricard
  • Nestlé
  • TotalEnergies
  • Radio France
  • Castorama
  • Enpal
  • Solpedinn

How RAG works

Preparation

Text is extracted (OCR for scans) and split along headings and tables, and each passage keeps its metadata: source, date, access rights.

PDF Word Notion Web

Indexing

Each passage gets a vector (embedding) and goes into a keyword index, both stored in PostgreSQL with pgvector.

embed(passage) → pgvector

Search

The question is matched by meaning and by keywords (BM25), filtered by the user's access rights, then the best passages are re-ranked.

hybrid_search → rerank → top_k

Answer and checks

The model writes from the retrieved passages, cites its sources and says when the information is missing. A test question set checks quality after every change.

LLM(passages + question) → réponse + sources

Use Cases

Internal Support

An assistant answers staff questions from internal documentation, with a link to the source document.

Technical Documentation

Manuals, procedures and technical knowledge bases can be queried in plain language, and every answer links back to the original page.

Guest concierge

The WhatsApp concierge built for independent hotels answers guests in several languages from the hotel's website and documents, indexed with Mistral embeddings in pgvector.

Contracts and procedures

Contracts, terms and conditions and quality procedures can be queried in plain language, and the assistant cites the article or clause the answer comes from.

Compatible Sources

PDF
Documents
DOC
Word
XLS
Excel
N
Notion
C
Confluence
API
Web

Companies we have worked for

Case studies

View all

Frequently asked questions

What is RAG (Retrieval Augmented Generation)?

RAG (Retrieval Augmented Generation) combines a search through your documents with text generation. Before answering, the system finds the relevant passages in your knowledge base, then the model writes its answer from those passages and can cite its sources. Hallucinations drop without disappearing, which is why they need to be measured.

RAG vs fine-tuning: which approach should you choose?

RAG suits data that changes often: you only need to update the index. Fine-tuning teaches the model a specific style or format. RAG is cheaper to set up and lets every answer cite its source, so it is the approach we choose in most projects.

Which documents can RAG index?

Any document you can extract text from: PDF, Word, Excel, PowerPoint, web pages, emails, wikis (Notion, Confluence), support tickets, SQL databases. Documents are split into passages, turned into vectors (embeddings) and stored, in our case in PostgreSQL with pgvector.

How do you measure the quality of a RAG system?

We build a set of questions with their expected answers, then measure whether retrieval finds the right passage (recall, precision) and whether the answer stays faithful to the sources. An LLM judge scores the answers, and the set becomes a regression test for every change of prompt, model or retrieval method.

AI Act and GDPR compliance

Indexed documents, access rights and retention periods under the GDPR and the AI Act.

Our guides Agents & RAG

All articles in this theme →