Understand AI

How AI uses your documents: search, context and retrieval

Understand what actually happens when an assistant answers from your company documents, why that is not the same as training, and what to check before you trust the answer.

For: Owners, office managers and professionals who want an assistant to answer from company policies, contracts and records without guessing.

Format: Explainer + source-readiness checklist
Published: September 9, 2026
Last checked: September 9, 2026
Written by: Dryvn AI editorial
Edited by: Dryvn AI editorial
Reviewed by: Not yet reviewed by a named person
Read as: Vendor-reported (from official documentation)

The short answer

When an assistant answers from your documents, it usually searches for the few passages that look relevant, reads them as context for that one question, and writes an answer. It does not memorise your files.
That process is called retrieval. It is separate from training, which changes the model itself. Major vendors document the two as different things; check the terms of the product you actually use.
The answer is only as good as what the search finds. Old versions, duplicates, missing documents and permission gaps produce confident wrong answers.
Good assistants show which passage an answer came from. Treat a citation as the thing to check, not as proof.
Before connecting documents, run the source-readiness checklist: owner, current version, who may see it, what conflicts, and where the assistant should say it does not know.

What actually happens when you ask about a document

People imagine the assistant has read every file in the business and remembers all of it. That is not how the common setups work. When you ask a question, three things happen in order. The system searches your documents for the passages that look most relevant to the question. It hands those passages to the model along with your question. The model writes an answer using what it was just handed.

The search step is the important one. Google Cloud's documentation describes the pattern as splitting documents into chunks, indexing them, then at question time searching the index for relevant pieces and passing them to the model as context [S31]. OpenAI's retrieval documentation describes the same idea: files are chunked, embedded and indexed, and a search returns the closest matches with the file name and a score [S25]. Different vendors, same shape.

So the model only sees a slice of your documents for each question. If the slice is right, the answer is usually good. If the search picks the wrong slice, or the right slice does not exist, the model will still write something. That is where the trouble starts.

How a document-grounded answer is produced

Text version of this diagram

1. Question (from a person) → next step
2. Find documents (search picks a few passages) → next step
3. Read as context (for this question only) → next step
4. Answer with sources (passages shown) → next step
5. Person checks (the cited passage)

Three words that get mixed up

Retrieval, context and training are different things
Term	What it means in plain words	What it means for your documents
Retrieval	Searching your documents for the passages most relevant to a question.	Your files stay where they are; the search reads them when asked.
Context	The text the model is given to work from for one question, including the passages found.	The model works from what it was handed, not from memory of your whole business.
Training	Changing the model itself by learning from large amounts of text.	A separate decision with its own terms. Retrieval does not require it.

The training question matters because it is the one people worry about most. Uploading a document so an assistant can search it is not the same as training a model on it. Vendors publish their own rules on this. OpenAI's data controls page, for example, states that data sent to its API is not used to train its models unless the customer opts in [S33]. Consumer products can have different defaults from business and API products. Read the terms of the product you actually use, not the one in the news.

Where document answers go wrong

Failure modes and what to check
What goes wrong	Why it happens	What to check
Answers from an old version	The old file is still in the search and looks just as relevant as the new one.	One current version per document. Archive or exclude the rest.
Two documents disagree	A policy and an email about the policy say different things; the search returns both.	Decide which is authoritative and label it. Remove or mark the other.
Confident answer to a missing document	The search found something loosely related and the model wrote an answer from it.	The assistant should say not found when nothing relevant is found. Test that it does.
Answer from a document the person should not see	Everything was indexed together without permissions.	Who may see each document must carry into the assistant. If it cannot, do not index the document.
Citation points to the wrong passage	The passage was relevant but did not support the specific claim.	Read the cited passage. A citation is a pointer, not a verdict.
Stale facts inside a current document	The document is current but a rate, name or date inside it is not.	Give documents an owner and a review date, the same as any record.

Why citations help, and why they are not proof

The better setups show you where an answer came from. Anthropic's documentation describes citations that return the exact passages supporting each claim so the answer can be verified and the source shown to users [S32]. OpenAI's retrieval results carry the file name and score for each match [S25]. That is genuinely useful. It turns a black box into something a person can check in thirty seconds.

It is still a pointer. A citation tells you which passage the model used. It does not tell you the passage was the right one, the current one, or that the model read it correctly. The habit to build is simple: when the answer matters, open the cited passage and read it. When the assistant gives no citation for a claim that matters, treat the claim as unverified.

Illustrative example · fictional

Two versions of a maintenance policy

A fictional property management office connects its shared drive to an assistant so the front desk can answer tenant questions faster. A tenant asks whether the landlord or the tenant pays for a blocked drain. The assistant answers confidently: the tenant pays, citing a document called Maintenance Responsibilities. The front desk repeats that to the tenant.

The problem is that the office changed that policy eight months ago. The new version says the landlord covers the first call-out and the tenant covers repeat blockages caused by misuse. Both files sit in the same folder: the old one is called Maintenance Responsibilities and the new one is called Maintenance Responsibilities v2 FINAL. The search found both, the old one matched the question slightly better, and the model wrote from it.

Nothing in this example is a fault of the model. The office had two current-looking versions of one policy and no owner deciding which one counts. The fix is on the checklist below: one authoritative version, the old one archived, an owner named, and a test question asked before the assistant went live.

Get your sources ready before you connect them

Most of the work in a document assistant is not the assistant. It is deciding which documents count, who owns them, who may see them and what the assistant should do when it finds nothing. Do that on paper first. The checklist below works for a shared drive, a policy binder or a folder of contracts.

01List the documents that answer real questionsStart with the ten questions your team gets asked most. Find the document that should answer each one. If two documents answer the same question, that is your first conflict.
02Pick one authoritative version of eachRename or archive the others. The assistant cannot tell FINAL from FINAL v2; you can.
03Name an owner and a review dateThe same as any record in the business. Documents without owners drift.
04Decide who may see each oneIf the assistant cannot respect that, leave the document out. Convenience is not a reason to widen access.
05Decide what not found should look likeWrite down three questions the documents do not answer. The assistant should say so, not improvise. Test it.
06Check five answers against the cited passagesBefore anyone relies on it, read the passage behind five real answers. Fix the documents, not the prompt, when an answer is wrong.

Worksheet

Source-readiness checklist

One row per document you intend to connect. Do not connect a document with a blank in the owner or permissions columns.

Document: Name and location
Owner: One person
Last updated: Date, and whether it is the current version
Who may see it: Everyone, team only, named people
Conflicts with: Any other document answering the same question
Questions it answers: The two or three it exists for
Where the assistant should say not found: Questions this document is often mistaken for answering

Downloads · no email required

Source-readiness checklist
The checklist above as a fill-in Markdown file, with a separate fictional example of a completed row.
Download .md

What this page does not cover

This is a business-facing explanation of a pattern that several vendors document. It does not tell you how to build a retrieval service or a vector store, and Dryvn does not describe its own internal setup here. Vendor claims above are taken from official documentation checked on the dates shown; they describe what those products say they do, not tests we ran. Products change. If you are choosing one, the AI tools guide covers what to compare, and the approval boundaries worksheet covers what an assistant should be allowed to do once it can read your files.

Questions people ask

How does AI answer questions using company documents?: In the common setups, the system searches your documents for the passages most relevant to the question, hands those passages to the model as context, and the model writes an answer from them, often with a citation back to the passage. It does not read everything and it does not remember your files between questions.
Does uploading a document train the AI?: Not by itself. Retrieval reads documents at question time; training changes the model. Vendors publish separate rules for training, and business or API products often differ from consumer ones. Check the terms of the exact product you use before assuming either way.
Why did the assistant answer from an old version of our policy?: Because the old version was still in the search and looked relevant. The search does not know which file is current. Keep one authoritative version and archive the rest.
Is a citation proof the answer is right?: No. It tells you which passage was used. Open it and read it when the answer matters.

Knowledge graphs and workflow graphs, explained for business ownersTell knowledge graphs from workflow graphs, see each drawn with a fictional example, and decide which relationships and decisions your business should record.
AI models and tools: how to choose for the jobKnow what you are comparing across OpenAI, Claude, Gemini, Grok, Llama and Perplexity, and pick a tool for one business job with a trial plan, not a ranking.
What should AI be allowed to do without asking?Decide, action by action, what an AI assistant or automation may do on its own, what it may do inside an approved scope, and what always waits for a person.
Turn a repeated task into a usable SOPTake a task that lives in someone's head and write it down so a new person could do it on day one, with the approvals, exceptions and evidence of completion spelled out.
BlogWhat Building Dryvn Taught Me About AI, Security and Serious SoftwareA first-person founder article: what building Dryvn actually taught me about security architecture, AI-assisted development, human engineering review and what it means to be trusted with real business data.

Sources

Dates are when each source was last checked by the editor. Sources support specific claims; they are not endorsements.

S31RAG Engine on Gemini Enterprise Agent Platform overview · Google Cloud documentation · checked September 9, 2026Describes the retrieval pattern: documents are split into chunks, indexed, searched at question time, and the relevant pieces are passed to the model as context. Supports local files, Cloud Storage and Google Drive as ingest options. Vendor documentation, not a test result.
S25Retrieval · OpenAI developer documentation · checked September 9, 2026States that files are automatically chunked, embedded and indexed, that search returns semantically similar results even with few matching keywords, and that results carry the file name and a score. Vendor documentation.
S32Citations · Claude Platform documentation (Anthropic) · checked September 9, 2026States that citations return the exact passages that support each claim so answers can be verified and sources shown. Vendor documentation.
S33Data controls in the OpenAI platform · OpenAI developer documentation · checked September 9, 2026States that data sent to the OpenAI API is not used to train OpenAI models unless the customer opts in. Used only to illustrate that training and retrieval are separate decisions; check the terms of the specific product you use.

R16 · Published September 9, 2026 · Next scheduled review December 9, 2026 · Teaches process management; not legal, warranty, safety or engineering advice. Examples are fictional unless stated. Part of the Dryvn resource library (17 resources).