.svg)
Answers with a source
Knowledge systems that answer from your own material and state where each statement came from. Across language boundaries too.
- RAG (retrieval-augmented generation)
- RAG is short for retrieval-augmented generation and is a method that searches your own documents first and only then lets a language model write the answer from the passages it found. Because the source comes before the answer, every statement can name where it came from.
Your knowledge exists. It just cannot be found.
Manuals on the file server, agreements in the inbox, details in meeting notes. Formally everything is there. In practice people ask a colleague.
Keyword search only helps when you already know the keyword. A language model on its own does not know your material at all and fills the gaps with something plausible.
A RAG system sits between the two: it searches your sources, puts the passages it found in front of the model and ships the source alongside. What is not in the material does not get asserted.
How the pipeline works
Four sources, three parallel search paths, one sourced answer. The self test runs against the same system your people use.
What such a system is made of
Connecting the sources
File stores, mailboxes, databases and line of business applications through their interfaces. Where none exists, through text recognition.
Preparation
Documents are split sensibly, given metadata and kept current. This is where it is decided whether anything useful gets found later.
Several search paths
Vector search for meaning, language aware full text for exact terms, a knowledge graph for relationships. The results are reranked together.
Answer tiers
Simple questions are answered immediately, demanding ones run in a longer tier. The interface never waits for the slowest case.
Sources and limits
Every statement comes with its source. When the system finds nothing suitable it says so instead of guessing.
Self test in production
An automated run puts scenarios to the system like a real user and scores the answers. Degradation shows up before your people report it.
When RAG suits and when something else does
Not every question about knowledge needs a knowledge system.
| Situation | Better choice | Why | What we advise |
|---|---|---|---|
| Few, stable documents | Improve the existing search | The effort does not pay off | Tune search first |
| Many sources, changing content | RAG | The index can be refreshed daily | The classic RAG case |
| Fixed rules, clear decision | Rules or automation | A model only blurs it | Solve it without a model |
| Several languages in the material | RAG with its own full text search | Word search fails without spaces | Index per language |
| Data cannot leave | RAG on your own hardware | Index and model stay in house | Set it up on premise |
Scroll sideways for all columns →
Go live is where it starts
A knowledge system ages with its sources. Without care the hit quality drops noticeably within a year.
That is why the self test belongs in operation rather than in acceptance.
RAG systems: common questions
Related topics
Bring a question nobody can answer quickly today.
Thirty minutes on that question show whether a knowledge system would carry in your company.