RAG chatbot knowledge base: prepare company documents

Create a simple register with Source, Topic, Audience, Owner, Approval status, Effective date, Review date and WebChatAgent destination. Include websites, PDFs, workspace pages and text snippets. “Shared drive” is not an owner; name the team or person who can approve a correction.

For each topic, mark one current source of truth. If the help center and a PDF disagree about support hours, stop and resolve the policy with the owner. Do not ask the chatbot to choose between two equally authoritative versions.

Group near-duplicates by topic. Typical duplicates are a current article plus a print view, an old PDF beside a new PDF, parameterized website URLs, exported workspace copies and translations that repeat the same facts. Compare their effective dates and approval state, then keep only the approved version in the chatbot source set.

Archive superseded material outside the connected folder or crawl scope. Search Priority can influence ranking later, but it is not a safe fix for contradictory policies. A stale document remains retrievable even when its weight is lower.

Replace generic headings such as “General” or “More information” with the question the section answers: “Support hours for Northstar customers” or “Annual-plan cancellation eligibility”. Keep one main topic per section and use short paragraphs under a real heading hierarchy.

Write explicit sentences. “It is available for 30 days” becomes “The Northstar onboarding review is available for 30 calendar days after the kickoff date.” Add exceptions and the effective date in the same section. Avoid references such as “see above”, because the retrieved passage may not contain the referenced text.

Write the full term on first use, for example “Service-level agreement (SLA)”. Keep product and plan names consistent across sources. If visitors say “refund” while the policy says “reimbursement”, include both terms naturally in the section.

Use unambiguous dates such as “1 August 2026” or an ISO date in structured fields. Attach currencies and units to every number. State the timezone for hours and deadlines. A bare “50” is not a usable fact when it could mean euros, percent, items or minutes.

Simple tables with one header row and one fact pattern per row are easier to extract. Split merged cells, repeat the row subject where needed and add a short paragraph that states the conclusion a visitor may ask about. For a comparison table, include product, condition, price and effective date in each relevant row.

Add text alternatives for diagrams and screenshots. A screenshot of opening hours is not a knowledge source until those hours also exist as readable text. Keep decorative images, signatures and navigation outside the answer-bearing content when possible.

Try selecting and copying a sentence from every page type. If the page is a scan, run optical character recognition before upload. Review codes, dates, decimal separators, accented names and page headers manually; these are common OCR failure points.

Paste the copied text into a plain-text editor and read it without the page layout. Check multi-column order, table rows, hyphenated line breaks and repeated headers. Re-export a clean accessible PDF when extraction mixes columns or drops essential labels. The demo control fact must remain exactly `NORDSTERN-42`.

Open the isolated assistant, choose Data Sources and select Add Data Source → Documents. Upload the approved Northstar demo PDF. Set the document name to “Northstar Knowledge Base” and the category to “Support” so a reviewer can identify its purpose without opening the file.

Leave AI optimization off for a clean, selectable-text document unless you have a specific extraction problem and will review the transformed result. Wait until the row shows Completed. Record the indexed character or chunk count as a baseline, not as a quality score.

Open Edit for Northstar Knowledge Base. Read the stored extracted text around the first heading, one table or list, the final section and NORDSTERN-42. Confirm that the editor still shows the intended Document Name and Category.

Keep Search Priority at 0 unless repeatable retrieval tests show a justified conflict. A higher value boosts a source relative to others; it does not repair bad OCR, missing context or an obsolete rule. Save only verified corrections, and keep the approved source file as the maintained original.

Choose Content Search and search for `NORDSTERN-42`. The matching result should name Northstar Knowledge Base and show the stored passage containing the fact. Repeat with one ordinary phrase such as “support hours” to make sure retrieval is not limited to an unusual code.

If the unique fact has no result, stop before changing the model or system prompt. Recheck source status, extracted text and the selected assistant. If Content Search finds the passage but the chatbot answer fails, investigate question wording, competing sources, role instructions and model behavior.

Start a fresh Test chatbot conversation and ask: “What is the support verification code in the Northstar Knowledge Base? Answer in one sentence and name the source.” The expected answer contains exactly NORDSTERN-42 and identifies the PDF. Record the actual wording, source version, model and date.

Start another fresh conversation and ask for Northstar’s 2028 warranty policy. The demo source intentionally has no answer. Passing means the assistant says the information is unavailable instead of inventing a duration or condition. Keep both questions in the release set after every source, prompt or model change.

When an owner approves a change, update the maintained source, effective date and register entry together. Re-index or replace the WebChatAgent source, confirm Completed, recheck the changed passage in Content Search and rerun the fixed question set.

Remove the superseded source from the active assistant only after the new version passes. Watch Questions, Feedback and Conversations for real wording that the test set missed. Add a question when it represents a recurring user need, then review critical topics on their documented cadence.