AI knowledge base: build from your files

Updated October 3, 2026 · Ask Your Docs

An AI knowledge base is an organized collection of sources an AI system can consult to answer questions. For a marketing or sales team, those sources may already exist as policies, price lists, product presentations and campaign notes. The work is to make the right information available, keep it current and verify the evidence behind each answer.

Start with one recurring job and a small source set. Use the same four steps throughout: upload a file, organize it, connect your AI and ask with sources. This guide explains what to check at each step, with downloadable fictional examples. It does not assume that every tool supports every format or guarantees correct answers.

Build the first source check

1. Choose one recurring question, such as whether Harbor paid ads can launch.
2. List three to five authoritative files and an owner in the source register; begin with the campaign brief and approval status.
3. Write the expected answer and one exception in the evaluation questions. Harbor still needs finance approval, and a customer quotation adds legal review.
4. Choose a route that can handle those files and the intended access level.
5. Run the approval question and a deliberately unanswerable cancellation-fee question, checking both against the original sources.

These Harbor answers are authored expectations from the public fictional files, not measured assistant results. Record what your chosen workflow actually returns.

Choose a question before choosing software

A useful starting question is one your team can answer today by opening a few documents. “Can this campaign launch?” is concrete: it requires an approval rule and a current status. “Tell us everything about the business” has no clear completion test and makes it difficult to detect missing evidence.

Write down the expected answer and its source before trying an AI workflow. Include one exception and one question that the sources cannot answer. This gives you a practical evaluation set rather than a subjective impression of a polished response. It also shows which files belong in the first library.

For sales, start with product eligibility or an approved claim. For marketing, use a launch checklist or brand rule. For operations, choose a documented handoff. Keep unrelated material out until the first questions work, then expand based on observed gaps.

Prepare PDF, Word, Excel, PowerPoint and Markdown

File upload support is only the first check. The important question is what information survives extraction. A PDF can contain text, scans and charts; a spreadsheet can contain formulas and saved values; a presentation can hide its essential context in images or notes. Inspect the source representation used by the system before trusting an answer about it.

The PDF guide, Word guide and Excel guide give format-specific exercises. Direct upload may be sufficient; converting every document into another format is not automatically an improvement.

Ask Your Docs currently extracts text and stored spreadsheet values for retrieval. That does not imply OCR, formula execution, complete visual analysis or preservation of slide notes. Keep the original files available for checks that depend on those features.

Organize sources by authority and purpose

Give each source a descriptive name and a clear role. A document index should identify the owner, version, review date and questions it answers. Avoid vague filenames such as “final-new-latest” when colleagues need to tell which policy applies. If a source is historical, label its effective period.

Resolve contradictory source material before using it for routine answers. If that is not possible, instruct the assistant to show the conflict and ask for clarification. It should not quietly select whichever passage is easiest to retrieve. A clean source collection reduces ambiguity for people as well as AI.

The downloadable review checklist separates a campaign policy from a sales exercise. Their numbers coexist in one collection, but the sources do not connect them. This is a useful test: an answer should not invent a relationship merely because two files are available.

Choose the route that matches where people work

There are several reasonable ways to make sources available. Native assistant projects keep a collection near its conversations. A connected library can make a maintained collection accessible through supported assistants. An existing knowledge-management product may be appropriate when your team already uses it for writing and approvals. Compare actual workflows instead of declaring one category universally superior.

OpenAI documents reference files and instructions in ChatGPT Projects. Anthropic documents a project knowledge base in Claude Projects. Their capabilities and account controls differ. The ChatGPT setup guide and Claude setup guide use the same examples so you can compare them deliberately.

For Ask Your Docs, use Connect your AI after uploading and organizing the source material. Complete authorization for the intended account and check that the connector is enabled where you ask. If a workspace administrator controls access, resolve that requirement before treating setup as complete.

Verify facts, exceptions and missing information

Use the fictional campaign approval document for a small acceptance test. Ask for the launch date, whether paid advertisements can launch, what changes when an email gains a customer quotation, and the cancellation fee. The answers require different kinds of evidence.

The launch date is October 12, 2026. Paid advertisements still need finance approval. A customer quotation triggers legal approval. The document does not specify a cancellation fee. These are expected source answers, not a claim that every assistant has passed the exercise. Open the document and inspect the cited passage after each response.

Answer using the library sources. Identify the filename and heading or table supporting your conclusion. Check exceptions. Separate facts from calculations and assumptions. If the source does not answer the question, say what is missing.

Score both the conclusion and its evidence. A correct-looking response with an unrelated citation is not reliable proof. For numeric questions, inspect the selected rows and calculation method. For permissions, test from the intended access level rather than assuming an owner-account result applies to everyone.

Make updates part of the workflow

An AI knowledge base needs a way to replace, withdraw and review sources. Assign that work to a named owner. After a material change, rerun the questions affected by it. If the source was removed, check that a fresh question does not rely on an obsolete copy or an earlier conversation answer.

The sample checklist includes an update exercise: change the fictional launch date in a version 1.1 document to October 19, 2026, replace the old reference and ask again in a fresh conversation. Require the answer to identify the new version. This tests your update process, not only the initial upload.

Keep a record of known failures and their causes. A bad answer may come from an incorrect source, missing extraction, ambiguous instructions or retrieval of an outdated passage. Each cause needs a different fix. Adding more documents without diagnosis can make the problem harder to understand.

Measure useful answers rather than library size

For an initial evaluation, record how many of your chosen questions have correct answers with correct supporting sources. Track missing-information handling separately, along with update checks and access checks. Label the sample size and date; a small internal test is not a universal accuracy claim.

Also record the time people spend maintaining the library and checking responses. A workflow that saves a search but creates extensive cleanup may not improve the job. Compare against the existing process using the same task and source set. Do not borrow a vendor’s productivity percentage as evidence about your team.

Expand after the first useful workflow is repeatable. Add the next source or question because it solves a real gap. Keep the source owner, update rule and verification method alongside it. That gives the library a maintainable foundation as more people begin to rely on it.

Frequently asked questions