RAG and knowledge bases
RAG (retrieval-augmented generation) finds the relevant passages in your own documents, then asks the AI to answer based on them and cite its sources. Your teams find the answer instead of hunting for the document.
Updated October 11, 2026
The reality
of working time spent searching for information, from 3 hours a week to a quarter of the week, according to vendor surveys.
retrieval failures with contextual retrieval combined with reranking, on Anthropic's datasets (−49% without reranking).
Source: Anthropic, Contextual Retrieval
of organizations don't have, or aren't sure they have, the right data management practices for AI.
Source: Gartner, via Freevacy
Anthropic's results come from their own datasets: they need to be validated on your documents.
How it works
Preparation happens once, then again whenever your documents are updated. Answers are generated for each question, in a few seconds.
Once: preparation
Procedures, contracts, product sheets, emails, internal databases.
Each document is split into short passages, enriched with their context.
Each passage is indexed by meaning (embeddings) and by exact keywords.
For each question: the answer
An employee asks a question in natural language.
The most relevant passages are retrieved, then ranked.
The AI writes the answer using only those passages.
The answer lists the source documents for verification.
Approach described by Lewis et al. (Facebook AI, NeurIPS 2020) and detailed by Anthropic.
Do you really need RAG?
We'll be upfront: if your knowledge base fits in a few hundred pages, RAG isn't always necessary. These three approaches complement each other more than they compete.
Choose it if: Your knowledge changes, it's large, and you want citations and per-user access rights.
Limitation: Quality depends on retrieval: an irrelevant passage can degrade the answer.
Choose it if: You want to enforce a consistent format, tone or vocabulary.
Limitation: Poorly suited to adding facts: they go stale and the model isn't updated.
Choose it if: Your knowledge base is small and stable: about 500 pages at most, according to Anthropic.
Limitation: Weak citations, and quality can degrade with very long contexts.
Not sure which approach is right? One conversation is enough to decide.
Talk to a consultantSources: Anthropic · Denser · AI Builder Club.
What makes a good RAG
The answer can't be better than the passages retrieved: quality is mostly decided at the retrieval stage.
An isolated excerpt loses its meaning (which company? which period?). We add its context before indexing it.
Meaning (embeddings) and exact keywords (references, acronyms, proper names) complement each other.
A second model reranks the retrieved passages to keep only the most useful ones.
Passing around twenty passages gives better results than five or ten, according to Anthropic.
Settings are tested on your own questions, not on generic examples.
French, Arabic, dialect
Sources: ArabicMTEB / Swan · Optimizing RAG Pipelines for Arabic.
Security and privacy
OWASP identifies vector database weaknesses as a risk category in its own right for AI applications. Here is what we put in place.
Your document access rights are enforced on every search, including in the index.
Departments, subsidiaries or clients don't mix: without isolation, an index can return someone else's documents.
An unprotected vector index is equivalent to an unprotected database.
The original text can be approximately reconstructed from vectors, so we treat them as sensitive data.
Sources: OWASP GenAI · Giskard. Regulatory framework: AI governance and the AI Act.
Measuring quality
You don't judge a RAG system by impression: you measure it with four metrics, computed on questions representative of your business. They are scored by a model and vary from one run to the next, so we track them over time.
Is every claim in the answer supported by the retrieved documents?
Does the answer actually address the question asked?
Are the right passages ranked ahead of irrelevant ones?
Did we retrieve everything needed to answer?
Our approach
Document sources and each user's access rights.
Clean, structure and chunk the content.
Semantic and exact-keyword search, with context for each passage.
Measure on your own questions before opening it to your teams.
Access rights, cited sources, monitoring and ongoing updates.
FAQ
RAG (retrieval-augmented generation) finds the relevant passages in your documents, then provides them to the AI so it can answer based on them and cite its sources.
No. It reduces them by grounding answers in your documents, but an irrelevant or contradictory passage can degrade the answer. That's why we measure quality on your questions.
For knowledge that changes and must be cited, RAG is generally the better choice. Fine-tuning is more for enforcing a format or tone. The two can be combined.
Maybe not. According to Anthropic, a knowledge base of about 500 pages or fewer can be placed directly in the prompt. Beyond that, or when access rights must be managed, RAG becomes relevant.
Confidentiality relies on access rights enforced on every search, data isolation and protection of the index. Vectors are not anonymous: we treat them as sensitive data.
Yes, but quality varies by model, especially for dialects. We test several approaches on a sample of your documents before committing to a result.
Go further
Tell us about your documents: we'll tell you whether RAG is the right solution and how to measure it.