RAG readiness check
Eight questions on your documents and your users: is RAG the right tool, and what would a pilot need?
Pick one answer per question. The verdict, what a pilot would need, and the risks appear once all eight are in. Nothing is sent anywhere until you choose to contact us.
Verdict
0 of 8 answered. The verdict appears when all eight are in.
Why check first
Most failed RAG projects were the wrong project.
Retrieval-augmented generation lets a language model answer from your own documents instead of its training data. It is the right tool when the answer lives in your files, such as manuals, policies, contracts, tickets, and transcripts, and has to be current and traceable. It is the wrong tool more often than vendors admit. If the answer is a number in a database, a model reading documents is slower and less reliable than a query. If you want the model to write differently rather than know more, that is a prompt or fine-tuning question.
This check asks the eight things we ask in the first call. Where the answers live, what kind of files they are, how often they change, who is allowed to see what, what people actually ask, how bad a wrong answer is, how much material there is, and whether the data may leave your network. Your answers decide three things: whether RAG fits, what a first pilot would have to include, and which risks to plan for before any money is spent.
How it works
Each answer maps to work we have done in production.
Some answers stop the check. If the answers live in a database, or the questions are totals and trends, the verdict sends you to the database first. If the knowledge is still in people's heads, there is nothing to retrieve yet. If you want a particular writing voice, RAG is not the lever. We would rather tell you that here than sell you a pilot that cannot work.
The other answers add to what a pilot needs. Scanned PDFs and tables need OCR and structure-aware parsing, the way SigmaCPM reads tender packages with retry handling for scanned pages. Video and audio need transcription first; NoteTube turns lectures, recordings, PDFs, and images into one index. Restricted documents need access control at the chunk level, so people only retrieve passages they are allowed to read, and per-user data needs scoping on every query, which NoteTube does per user and per session in Qdrant. Documents that change daily need an ingestion pipeline that keeps the index in sync, not a one-off upload.
High stakes change the answer format. In SigmaCPM, every value the model extracts from a tender must come with a quote, and that quote must be found on the cited page. If it is not there, the value is dropped instead of guessed. That is the standard we apply when a wrong answer costs money or creates legal risk. When data may not leave your network, the vector store, the pipeline, and an open-weight model can run inside your cloud account or on-premise.
The verdict weighs how many of these hard parts stack up. A simple setup is a contained pilot. Several together still fit, but the pilot should start narrow, with one source type and one group of users, and prove accuracy before it widens.
What a pilot is
One document set, real questions, a score.
Every pilot we run has the same core. One document set and one group of users. Ingestion that survives the real files. Hybrid search, embeddings plus keyword search, with metadata filters and a reranker, tuned to your corpus rather than a public benchmark. Answers that cite the passage they came from and say so when the documents do not cover a question. And an evaluation set built from your real questions, scored on every change, so you know whether it is getting better.
A working pilot on one document set usually takes three to six weeks and starts at $5,000 to $10,000, including the evaluation set. Production hardening after that depends on volume, access rules, and integrations. Run the check, then bring the result to a call: it saves the first twenty minutes.
Questions
Questions we get asked.
What is RAG and when do we need it?
Retrieval-augmented generation lets a language model answer from your documents instead of its training data. You need it when the answer lives in your own files and has to be current and traceable.
When is RAG the wrong choice?
When the answer is a number in a database, query the database. When you need the model to change how it writes rather than what it knows, that is a prompt or fine-tuning question.
Can RAG run without our data leaving the network?
Yes. The vector store, the pipeline, and an open-weight model can be deployed inside your cloud account or on-premise, so documents never leave your network.
How do you stop the model from making things up?
Better retrieval first: structure-aware chunking, hybrid search, and reranking. Then answers must be grounded in retrieved passages and cite them, and the system refuses when nothing relevant is found. We measure this on an evaluation set before launch.
How long does a RAG pilot take and what does it cost?
A working pilot on one document set usually takes three to six weeks and starts at $5,000 to $10,000, including an evaluation set built from your real questions.
Can people only see answers from documents they are allowed to read?
Yes. We apply access control at the chunk level, and per-user or per-client scoping on every query, so retrieval never returns a passage the person could not open themselves.
RAG readiness check
Bring your verdict to a call.
Thirty minutes on your documents and your questions. You leave with a pilot plan: the document set, the evaluation, the timeline, and a number.