RAG stands for retrieval-augmented generation. A language model looks things up before every answer: a search step pulls the matching passages from your own texts, and the model writes its answer from those passages.
In CIRO NOISE, the AI assistant Isa is a RAG system. She searches your wiki, documents, requests, meetings, tasks and websites and lists the sections each answer came from below it.
Why RAG for company knowledge
A language model doesn’t know your prices, contracts or processes. Ask it anyway, and it often writes an answer that sounds plausible and is wrong. With RAG, the model gets the matching sections from your data along with every question. When search finds nothing, the AI says so. When someone edits a wiki page, the new version counts from the next index run, and no model has to be retrained for it.
How RAG works in CIRO NOISE
CIRO NOISE splits every source into short sections at its headings and refreshes the search index every five minutes. For a question, CIRO NOISE searches two ways: by meaning, so “approval” also finds “sign-off”, and by the exact words, so an invoice number or a contract clause is found too. A second step checks whether the top hits fit the question before the AI reads them.
The AI writes its answer from those sections, and every link below it leads to the section it came from. Search only covers what the person asking may see: without read access to the wiki, no wiki section reaches the answer, and confidential pages need their own permission.
Documents and websites as sources
CIRO NOISE reads PDFs, including scanned ones, Word files and text files once and creates a profile: document type, summary, parties, key dates and deadlines. In your file storage, photos of documents count too. Deadlines from documents go into a list, and CIRO NOISE reminds the person responsible 30, 7 and 1 day ahead and on the day itself. Because an extracted deadline hasn’t been checked legally, every reminder leads to the document.
You can add up to five websites as sources, such as your company website or your help center. CIRO NOISE reads up to 100 pages per site, follows its robots.txt and reads it again every week.
An everyday example
Lena from sales asks: “What’s the notice period in the maintenance contract with Winter GmbH?” The contract is a PDF attached to a task on the customer board. The AI finds the “Term” section in it, answers “three months to the end of a quarter, in writing” and links the file. Lena neither has to dig through folders nor read the whole contract.
Where RAG reaches its limits
RAG makes an answer checkable, but it doesn’t make it error-free. If the wiki holds an outdated rule, the AI repeats the outdated rule, with a link to the page. That’s why the knowledge area shows under “Quality” which questions went unanswered and which answers your team marked as not helpful. Whoever maintains the knowledge answers them with an FAQ entry.