Enterprise RAG over your own documents: The answer is already written down. Now anyone can find it.
Your policies, contracts, manuals and procedures already answer most of what your people ask each other every day. The answers sit in SharePoint, Confluence, shared drives and PDFs nobody reads to the end, so people email the colleague who knows. We build retrieval over those documents that answers in the assistant your people already use, quotes the page it came from, shows each person only what they could already open, and says so when the answer is not written down.
▸ "Can Fabrikam move our support work offshore?"
✓ identity · token for this person
✓ permissions · 412 of 1,180 documents you can open
✓ search · keyword + meaning · 5 passages
✓ Fabrikam MSA · §12.3 · p. 14
✓ Amendment 2 · §1 · notice raised to 30 days
◆ MSA 2022 · superseded, never quoted
✓ citations · document, version, page
→ answered in the assistant you already use
Sound familiar
The question someone asked a colleague this morning.
It is written down somewhere. Finding it costs an email, a meeting or a guess. Here is the same question, answered from the document, with the page attached.
Customer support
“Does the X200 warranty cover water damage on an outdoor install?”
TodayThree tabs, one escalation, and a customer waiting two days.
From your documents
Only if the IP67 enclosure was fitted. Without it, an outdoor install voids water-damage cover. 1
1 · X200 Warranty Terms · v5 · §5.2 · p. 3
Legal & procurement
“What is our liability cap with Litware, and are data breaches inside it?”
TodaySomeone opens the contract, then both amendments, then asks outside counsel.
From your documents
Twelve months of fees. Data breaches are carved out of the cap, so that exposure is uncapped. 1
1 · Litware MSA · §14.1–14.3 · p. 22
Field operations
“What torque do the flange bolts on the P-300 pump take?”
TodayThe technician calls the one engineer who remembers, and waits.
From your documents
85 N·m in a star pattern, rechecked after the first 24 hours of running. 1
1 · P-300 Maintenance Manual · rev. F · p. 47
People & HR
“How many vacation days can I carry into next year?”
TodayAn email to HR, answered on Thursday, and the same email from forty other people.
From your documents
Five, used by 31 March. Your manager can approve up to ten if you were on leave this year. 1
1 · Leave Policy 2026 · §3.4 · p. 5
Why document chatbots get switched off
Finding a paragraph is easy. Trusting it is the work.
The quick version loads every document into a vector database under an admin account and puts a chat box on top. It demos well. Then it quotes the 2022 policy as readily as the 2025 one, shows a junior analyst a salary band because the index could see everything, reads a scanned table as a run of loose numbers, and answers a question nobody ever wrote down with the same confidence as one somebody did.
A confident wrong answer has a cost. In 2024 a Canadian tribunal held an airline to a refund its website chatbot had promised against the airline's own policy, and rejected the argument that the chatbot was responsible for its own words. Legal and security teams know this, which is why so many document assistants stall in review or are quietly switched off after the pilot.
Versions, permissions, layout and honesty about gaps are the specification. They are the parts we build first.
What we build
Four things, each one you can check.
We start with one team's documents, read-only, and grow from there. That start is the First Graph: four weeks, $35,000 fixed, credited in full to the build if you continue, and a document library counts as one of its three systems. Everything runs in your cloud account, from infrastructure as code, behind your identity provider.
01
Documents read the way people read them
PDFs, scans, tables, slides and wiki pages are parsed with their layout intact: a heading stays a heading, a table stays a table, a scanned page is read with OCR. Every passage keeps its document, version, section and page, and when a document changes, only that document is read again.
02
Search that knows who is asking
Access lists are copied from SharePoint, Confluence or your drives along with the text and applied on every query, so a person can only retrieve passages from documents they could already open. Remove someone's access in the source and the index follows on the next sync.
03
Answers that quote the page
Keyword and meaning search run together and a reranker keeps the best passages. The answer is written from those alone and every sentence cites one. The version in force wins over a superseded one, and a question nothing answers gets "not in your documents" and the name of the owner to ask.
How citations are produced04
A test set made of your own questions
Before launch we collect the questions your teams actually ask, each with the page that answers it. That set runs on every change to the parser, the index or the prompt, and a change that finds the right page less often does not ship.
3
Coordinates on every quote: document, version, page
100%
Answers that link the page they quote
0
Documents searched that the person could not already open
0
Superseded versions quoted
Walkthrough
Watch one question find its page.
“Can Fabrikam move our support work offshore?” Six steps from that question, typed by a procurement lead into the assistant they already use, to an answer that quotes the contract, a gap it admits, and a new amendment it follows.
-
01 / 06
Your documents become passages, with their place kept.
The parser reads your libraries the way a person would: headings, tables, scanned pages. Each passage keeps its document, version, section and page, and the list of people allowed to open the document travels with it.
Fabrikam MSA.pdf38 pages · 214 passagesAmendment 2.pdf3 pages · replaces §12.3 noticeaccessLegal · Procurement
unreadable pages · 0version in force · 2025 -
02 / 06
Someone asks, in the assistant they already use.
Copilot, Claude or ChatGPT turns the question into a call to your retrieval server. The server checks the token on this very request: issued by your identity provider, for this server, for this person.
tool docs.search args { "q": "Can Fabrikam move our support work offshore?" } caller the procurement lead who asked token audience = this server ✓ -
03 / 06
It searches only what this person can open.
Before anything is ranked, the index drops every document this person could not open in the source system. Keyword and meaning search then run together, a reranker keeps the best five passages, and superseded versions are left out.
permissions412 of 1,180 documents visiblecandidates40 passages by keyword and meaningkept5 passages after reranking
superseded versions · excludedMSA 2022 -
04 / 06
The answer quotes the page.
The assistant receives the five passages with their coordinates and writes its answer from those alone. Every sentence cites one, so anyone can open the page and check it in seconds. A sentence that cites something it was never given is removed rather than shown.
MSA §12.3 · p. 14subcontracting allowed with noticeAmend. 2 §1 · p. 1notice raised to 30 daysanswer“Yes, with 30 days' written notice.”
every sentence citedsigned March 2025 -
05 / 06
Not written down? It says so.
She follows up: does that include our customer data? Nothing in the documents she can open answers it. The assistant says exactly that, names the closest passage, and sends the question to the owner of the contract instead of guessing.
answernot in your documentsclosestMSA §9 · data handling, no location namedownercontracts team · question sent
guesses · 0sent to a person -
06 / 06
A new version is signed. The answers follow it.
When Amendment 3 is saved to the library, it is read within minutes and the clause it replaces is marked superseded. From then on the answer quotes the new wording, and every question along the way has left one audit line: who asked, which passages, which answer.
09:14docs.search · procurement lead09:16not found · sent to contracts team16:02Amendment 3 indexed · §1 of Amend. 2 superseded
superseded quoted · 0shared service keys · 0
Anatomy of a question
Found, it quotes the page. Not found, it says so.
Every question passes the same checks in the same order, whichever assistant sent it. The model writes the words of the answer; it never decides who is asking, what they may read, or which version is in force.
- any MCP client
-
1 · ASSISTANT Copilot, Claude, ChatGPT The model picks the tool and phrases the question. That is all it decides.
-
2 · IDENTITY Token checked on every request Issued by your IdP, for this server, for this person.
-
3 · PERMISSIONS Only documents this person can open Access lists from the source system, applied before anything is ranked.
-
4 · RETRIEVAL Keyword and meaning, reranked Versions in force only. Superseded passages never reach the model.
-
every question
5 · AUDIT LOG One line: the person, the passages, the answer An answer someone marks wrong is added to the test set.
Built to be relied on
Answers your legal team will let people act on.
These controls are in every retrieval system we build. Each one is something a reviewer can test before launch.
Permissions follow the source
If a person cannot open a document where it lives, the assistant cannot quote it to them. Access changes in SharePoint, Confluence or the drive reach the index on the next sync, and a deleted document leaves the answers with it.
The version in force wins
Every document carries its version and effective date. When a policy or contract is replaced, the old passages are marked superseded and never quoted, so the answer matches what applies today.
Document text is data, never instructions
A line hidden in a PDF that tells the assistant to ignore its instructions is read as text in a document, not obeyed. What the model may call is fixed before the session starts, so nothing it reads can widen that list.
How injected instructions are containedYour documents stay in your cloud
The parser, the index and the embeddings run in your cloud account. To answer a question, only the passages that answer it go to the model your assistant already uses, under the agreement you already have with that provider.
What BYOC meansMeasured on your questions, not assumed
Before launch you see how often the right page was found for your own questions, and how often the answer stayed inside the passages it cited. The same report runs after every change, so quality is a number you can watch.
Retrieval or a graph?
Documents answer what was written. A graph answers how things connect.
Ask what a policy says and retrieval over the document is the right tool. Ask which systems that policy covers and who changed them last quarter, and you need a knowledge graph. Most companies need both, and both answer through the same server, under the same sign-in.
Policies & procedures
HR, security, finance and quality policies, quoted in the version that applies today.
Contracts & agreements
Clauses, amendments and what replaced what, shown only to the people allowed to read them.
Manuals & technical documents
Maintenance manuals, specifications and runbooks, with their tables and part numbers intact.
Proposals & questionnaires
Approved answers to security questionnaires and RFPs, reused instead of rewritten from memory.
Before you hire anyone
When we will tell you not to build one.
Your assistant already covers it
If everyone has Microsoft 365 Copilot and the documents live in SharePoint with tidy permissions, its own search may be enough. Try it on your five hardest questions first; we will help you read the results.
The answer lives in rows, not prose
Revenue by region or open tickets by team are database questions. They need a governed source and a semantic model, not document search.
Agentic business intelligenceNobody owns the documents
If three versions of a policy contradict each other and no one can say which is current, retrieval will quote the confusion faithfully. Ownership comes first, and we will show you where it matters most.
Asked in the first meeting.
Already have a document assistant stalled in security review? The Security Review Memo answers it in five business days, without access to your data.
In the assistant they already use. The retrieval server speaks the Model Context Protocol, so Copilot, Claude, ChatGPT business plans and Gemini Enterprise can all call it. Where a team works in Teams or Slack instead, the same server can answer there.
The ones your people already open. PDFs, scanned ones included, Word, PowerPoint, Excel and wiki pages, from SharePoint, OneDrive, Confluence, Google Drive, Box, network shares and most document systems with an API. Tables, headings and page numbers are kept, and you get a list of any page that could not be read.
No. The libraries, the parser, the index and the embeddings all run in your cloud account. To answer a question, the few passages that answer it go to the model behind your assistant, under the agreement you already have with that provider. You receive the full source, its history and its tests, with no licence and no per-seat fee.
Every sentence links its page, and the system is measured on your questions. Anyone can open the cited page and check a claim in seconds. Before launch we run a test set built from questions your teams really ask, and report how often the right page was found and whether the answer stayed inside it. The same test runs on every change afterwards.
It is read again within minutes, and the version it replaces stops being quoted. Access changes travel the same way, so someone who loses access to a folder stops seeing its passages on the next sync.
At a fixed, published price. Most teams start with a four-week, $35,000 First Graph over up to three systems, where a document library counts as one, credited in full to the build if you continue. A build for one team runs from $75,000. The pricing page shows what moves the number.
Bring the policy nobody reads to the end.
The one people email a colleague about instead of opening. That is where retrieval pays for itself first, and a four-week First Graph is the fastest way to see your own questions answered with the page attached.