RAG
RAG: answers grounded in your own documentation
In short
RAG (Retrieval-Augmented Generation) is the technique that retrieves the most relevant passages from a knowledge base before a language model generates its answer. In RootNia, RAG works over documents already classified by sector, so every answer is based on your own sources and links to the documents that ground it.
Last updated ·
Why an LLM alone is not enough
A language model answers with what it learned during training. It does not know your policies, contracts or internal procedures, and when it does not know something it may invent a plausible answer. Using it over your own documentation requires giving it the right context with each question.
How RAG works
- Retrieval: for each question, the system searches the knowledge base for the most relevant passages.
- Augmentation: those passages are passed to the model along with the question.
- Generation: the model drafts the answer from that context, and the answer links to its sources.
What RootNia adds to RAG
- Connected, synchronized sources: SharePoint, OneDrive and databases, internal and external.
- Sector classification first: retrieval works over already-structured documents, not an undifferentiated pile of text.
- Governed retrieval: private, public and technical access tiers, and a log of every retrieval.
- Traceability: VERA anchors every value to its region in the document.
- Model-agnostic: ChatGPT, Claude, Kimi or others, swappable without rebuilding the pipeline.
Example: one question, three documents
Illustrative example of how it works, not a specific customer.
Someone asks: “What is the notice period to cancel the contract with supplier X?” RAG retrieves the term clause from the master agreement, a later amendment that changes it and the email confirming the renewal. The model answers with the current notice period under the amendment and cites all three documents, so the person asking can see why the answer differs from the original contract.
Limits and good practices
- If it is not in the documents, it is not in the answer: RAG cannot supply information your documentation does not contain.
- Outdated documentation: when old and new versions coexist, classification and access tiers help prioritize the current ones.
- Verify what matters: use the source links and VERA before acting on an important value.
RAG and structured data
RAG answers over documents. For questions about databases, RootNia uses NL-to-SQL. If you are considering a general-purpose assistant, see the comparison with ChatGPT and Claude.
Frequently asked questions
What is RAG in a nutshell?
Giving the language model the relevant fragments of your documents together with the question, so it answers from them and not only from its general knowledge.
Does RAG eliminate hallucinations?
It reduces them, because the model answers from specific sources, but no system eliminates them entirely. That is why RootNia links every answer to its sources and VERA lets you verify each value in the original document.
What does RootNia add compared with building your own RAG?
The orchestrator keeps sources in sync, the classifier structures documents by sector before retrieval, access is tier-based and every retrieval is logged. You can also switch LLMs without rebuilding the pipeline.