The documentation that today lives scattered across different tools, turned into a single source that can answer when someone asks it a question
CODEPRESS builds an enterprise RAG system: documentation scattered across different tools becomes an indexed knowledge base that answers one precise question instead of handing back a list of documents to open. The pilot starts from €6,000 and runs four to six weeks, on a real, bounded corpus: it leaves you with a working assistant on your own documents and answer quality that has been measured.
from €6,000pilot · 4–6 weeks
What an enterprise RAG is for
Every company has a gap between what's written down somewhere and what a person needs to know to answer right away: procedures sitting in a shared drive, decisions made in a meeting nobody wrote up, the knowledge of whoever has been there longest and has never put it on paper. Whoever needs an answer opens three different tools before finding it, if they find it at all, and often ends up interrupting a colleague instead of checking the right document.
An enterprise RAG system takes those documents, scattered as they are, and turns them into a single base a language model can query to answer one precise question, citing where the answer came from instead of inventing it. The same setup serves different audiences without duplicating anything: the internal team looking for a procedure, a new hire finding their footing in the first weeks, an outside user asking a question to an assistant the company has published.
What the RAG pilot covers
An indexed knowledge base
Documents, manuals, procedures and wiki pages you already have get turned into an index the system can query, not just archive: it stays current by following the original sources, so it doesn't drift from their content.
An internal wiki and onboarding by role
The same base feeds a queryable wiki for people already on the team and an onboarding path tailored by role: a new salesperson and a new developer get different answers from the same source, with no two separate sets of documentation to keep in sync.
Assistants you can publish to your users
The same setup can also face outward, as an assistant that answers your customers using your public documentation: the technical upkeep stays inside the service; it doesn't become another system for you to run.
A pilot, not a year-long project
It starts on a real, bounded subset of your documents, not the whole company archive: first we prove the answers hold up on a small corpus; then we decide whether and how to extend the coverage.
How an enterprise RAG works
Choosing the corpus and access
Together we decide which documents go into the pilot and who inside the company should be able to query them through the system: the boundary of permissions gets set before anything is indexed, not after.
Indexing and retrieval trials
The chosen documents are split into queryable chunks and indexed; then we run the system against real questions from whoever will use the assistant, to see whether it retrieves the right fragments before checking the final answer.
Tuning the answers
Where the system retrieves the right fragment but answers it badly, we adjust how the question reaches the model; where it doesn't retrieve it at all, we revisit how the documents were split or indexed.
Delivering the pilot and the extension plan
The pilot closes with the assistant running on the chosen corpus, answer quality measured against real questions, and a written plan for extending coverage to the rest of the company's documentation.
What you get from the RAG pilot
- The assistant running on the corpus chosen for the pilot, ready to query with real questions
- The index built from your documents, with the logic for keeping it current from the original sources
- Answer quality measured against a real set of questions, not a demo case prepared for the occasion
- The internal wiki and the role-based onboarding path, where they fall inside the pilot's scope
- The written plan for extending coverage to the rest of the company's documentation, with priorities marked
- The decision on which model answers the questions, made together with you before any confidential document is indexed
When an enterprise RAG is not for you
- The documents don't exist yet in a form you could query: scattered notes, meetings nobody wrote down, knowledge that lives only in someone's head. RAG indexes what is already written somewhere; it doesn't write it for you.
- The questions it would answer have a deterministic answer that a database query would already give you, faster and cheaper: a language model is the wrong tool for a question with one correct, checkable answer.
- Nobody at the company has taken ownership of keeping the corpus current: a RAG system built on documents nobody touches again goes stale quietly, and inside a year it starts answering with information that is no longer true.
What an enterprise RAG costs, and why
What pushes the figure past €6,000 is how many systems need connecting to build the corpus, how messy the starting documentation is, and whether the pilot also covers an assistant published to outside users, which carries different requirements from an internal-only wiki.
The figure gets set after seeing where your documents live, not before: one tidy shared folder lands near the floor; several tools to connect cost more. From there it holds steady, unchanged across the four to six weeks the pilot runs.
Enterprise RAG: frequently asked questions
- How much does an enterprise RAG system cost?
- The pilot starts at €6,000; the final figure climbs from there based on how many systems need connecting to build the corpus and how messy the starting documentation is. We set the number before the pilot begins, once we've seen where your documents live, and it holds steady for the four to six weeks.
- What is the difference between RAG and a chatbot?
- A generic chatbot answers from whatever the model picked up during training, and on a detail specific to your company it can produce a plausible answer that's simply wrong. A RAG system searches your real documents first, retrieves the fragments relevant to the question, and hands them to the model to build the answer only from those, citing where it came from. If the documents don't say it, the system can say so, instead of making something up.
- Do our documents end up in a public model?
- On every question, the model receives only the document fragments relevant to that question, never the whole corpus at once. If the documents are confidential, which model answers, a public API or an alternative running inside a closed environment, is a decision we make together with you before the pilot starts, not one we've already made for you.
- What kinds of documents does it work on?
- It works on whatever you've already written down somewhere, even in a messy form: procedures, manuals, internal wiki pages, PDF or Word documents, company policies. It doesn't work on knowledge that exists only in someone's head and has never been written down: that has to be written first, then indexed, and writing it isn't part of this service.
- How long does the pilot take?
- Four to six weeks, from choosing the corpus to delivering the assistant with answer quality measured. Where a pilot falls in that window comes down to how messy the starting documentation is: a corpus already organised in a handful of sources needs less time to index and test than one scattered across many different tools.