BUSINESS AI GUIDE · CREATED BY INRI INC

How to Build a Small Company Knowledge Library Your Team Can Trust

Start with current SOPs, clear owners, narrow questions, and source review instead of importing every company document.

Prepared for InriGPT by Inri Inc · Updated 27 October 2026 · Fictional practice material

Worked example: test a library with an unanswered question

Our fictional returns SOP names the inventory manager as the person who approves a stock adjustment. It gives a deadline for notifying the shift supervisor. It does not define a refund authority, amount limit, or weekend exception. A useful pilot tests all of those boundaries, including questions the library should not answer.

Run the four questions in the document-question walkthrough. Mark a response supported only when its actual claim matches the source. An answer that quotes the correct role but invents a refund threshold is not supported. An honest “not specified” is an acceptable result for the missing-policy questions.

What to recordExample entry
Source and scopeReturns SOP v1, sections 2–3; fictional warehouse returns only
Supported resultInventory manager approves stock adjustments
Unsupported extensionAny manager can authorize a refund up to a guessed amount
Owner actionAsk the policy owner to add approved refund and weekend rules if the team needs them

After correcting a source, repeat the same test questions. That distinguishes a document improvement from a lucky new answer. Keep the question wording, expected source, review result, and change date together. The test set is small enough for an owner to inspect, and its unanswered questions make gaps visible before a team relies on the library.

Choose one operational question set

A useful company knowledge pilot begins with a small group of questions people already ask. For example, a logistics team may need handover instructions, a return process, and the next step when a delivery exception occurs. Collect those questions from the people doing the work. Avoid starting with an unfiltered archive of everything the company has written. A narrow set makes it easier to inspect the results, discover missing procedures, and decide whether the workflow is actually improving access to information.

Select current sources and name their owners

Use approved procedures rather than rough notes when a question concerns official policy. Give each document a clear title and identify who can update it. Remove obsolete copies and explain any differences between locations, roles, or product versions. Split unrelated material so the relevant text can be found more easily. A library can return conflicting instructions when the underlying sources disagree; the model cannot resolve a policy decision that the business has not made. Keeping sources current is an ongoing responsibility.

Understand what retrieval does and does not do

The INRI GPT organization library stores supplied text and selects bounded excerpts using keyword relevance when a teammate asks a question. The excerpts accompany the request and can be inspected beside the answer. This does not train a new foundation model, automatically synchronize external accounts, or verify every statement in a reply. Ask a specific question using terminology that appears in the sources. If there is no useful match, locate or add the correct procedure instead of accepting a plausible unsupported answer.

Test questions that expose missing information

Create examples with a direct answer, an exception case, conflicting sources, and a topic outside the library. Ask the document owner to compare replies with the underlying policy. Record where the question needs more context and where the source needs revision. Include a reviewer who did not prepare the documents; they may interpret an instruction differently from its author. A good pilot should reveal gaps, not conceal them behind fluent language. Decide how a teammate escalates uncertainty to a person.

Set the sharing boundary before expanding

Only import material the team is allowed to use, and review sensitive-data requirements before working with personal or regulated records. The organization package supports shared text documents, conversations, and templates within stated limits. Google Workspace linking verifies identity; it is not automatic document synchronization. Begin with a few sources and inspect the excerpts for real questions. Expand when owners can maintain the content and reviewers understand the limits. The practical goal is easier access to approved information, with a clear route to a human when the sources are incomplete.

Try a source-based workflow

Explore the related product and its limits, or use a sample in the free demo before choosing a package.

When the source lives in Google Drive

Use the Google Workspace and Drive guide to distinguish identity linking from document synchronization. It covers selecting a supported source, checking its library copy, testing an unanswered question and verifying an update. A connected status with no selected files does not mean a source has been imported.