AI Knowledge Base Search for Internal Teams

Why your team still cannot find the answer

Most companies do not lack documentation. They have too much of it, spread across a SharePoint site from one era, a Confluence space from another, shared drives full of PDFs, and years of resolved support tickets that hold the real answers. A new hire asks a senior colleague instead of searching, because keyword search returns forty files with procedure in the title and no way to tell which one is current.

AI knowledge base search changes the interaction. Instead of a list of files, the employee asks how to process a warranty return for a commercial account, or which forms a new vendor needs, and gets a short answer with links to the two or three passages it came from. Tepia builds these systems so every answer is traceable: if the source is not shown, the employee has no reason to trust it.

The questions that matter differ by team. Support wants answers from past tickets and product manuals. Operations wants standard procedures and safety documents. HR wants policy answers that vary by state or employee class. Tepia starts by collecting fifty to one hundred real questions from each team, since those questions decide which content gets indexed first.

Content sources, connectors and permissions

Tepia connects each source on its own terms, because each one stores files and access rules differently.

Permission handling is where many internal AI projects go wrong. If the index ignores access controls, a warehouse lead can ask about compensation bands and get an answer pulled from an HR folder. Tepia stores the access list alongside every indexed passage and filters results for the signed in user before the model ever sees them, so the assistant cannot reveal anything the employee could not open directly. Sign in through Microsoft Entra ID, Okta or Google Workspace ties it together; see single sign on and identity integration.

Making answers accurate, current and cited

Retrieval quality decides answer quality. Tepia uses hybrid search, combining keyword matching for part numbers, policy codes and acronyms with vector search for meaning, then reranks the top passages before the model writes anything. Documents are split along their real structure, by heading, step or table row, so a procedure is never cut in half.

• Freshness: connectors sync on change events or a schedule, and deleted or archived files drop out of the index the same day.
• Conflicts: when two documents disagree, the answer shows both with their dates and owners rather than silently picking one.
• Refusals: if nothing relevant is retrieved, the assistant says it does not know and names who owns the topic.
• Evaluation: the question set from Discovery becomes a test suite that runs on every change to prompts, chunking or models.

Tepia stores embeddings in PostgreSQL with pgvector or a dedicated vector database, and runs models through Azure OpenAI, AWS Bedrock or the OpenAI and Anthropic APIs, depending on where your data already lives. For the method in depth, read how to build a RAG system.

Where employees ask their questions

A search tool nobody opens does not help. Tepia puts the assistant where your team already works:

• A Microsoft Teams or Slack app that answers in a thread and shows source links inline.
• A search panel inside your intranet, help desk or internal tools.
• A mobile view for field and warehouse staff who need a procedure while standing at the machine.
• A browser side panel for support agents working inside a ticketing system.

Every answer carries a thumbs up or down and an optional comment. Tepia routes negative feedback to the content owner for that document, which turns the assistant into a running list of missing or outdated documentation. Within a few months, most teams find the feedback log is as useful as the answers themselves, because it shows exactly where the written knowledge has gaps.

How Tepia approaches AI knowledge base search

1. Discovery: an inventory of content sources, owners and access models, plus interviews with each team about the questions they ask colleagues every week. Deliverables are an Investigation Summary, an Interview Summary and User Stories such as a support agent finding the fix for a specific error code.
2. Design: the answer layout with citations, the conflict and refusal states, feedback controls and an admin view for content owners, shown as detailed wireframes.
3. Development and Testing: connectors, indexing and the permission filter come first, then the answer layer. The Alpha serves one team; the Beta adds sources and roles. Test plans include the question set, permission leak tests and load tests on sync.
4. Training: short sessions on asking good questions, reading citations and flagging bad answers, plus a separate session for content owners.
5. Launch: a full index build, rollout team by team, and a review of the first weeks of questions with each department.
6. Support: new sources, model upgrades tested against the question set, and monthly reports on unanswered questions.

Typically Discovery takes 1 to 3 months and development and QA 2 to 6 months, with about a month to roll out. One source for one team is often live in the shorter window. More on how Tepia works.

Built in AI search or a custom knowledge assistant

If your content lives almost entirely in Microsoft 365 and your questions are about those documents, the AI search built into that suite may be enough, and Tepia will say so. The same applies to help desk platforms that now suggest answers from their own articles.

A custom assistant is worth building when:

• Answers span several systems that no single vendor indexes together.
• You need answers grounded in structured records from your own apps, not just documents.
• Permission rules are more detailed than folder access, such as by region, license or customer account.
• You want the assistant inside a product or tool your team already uses all day.

Once the internal version proves its accuracy, Tepia can extend it into a chatbot trained on your company data for customers. Browse our work or contact Tepia to map your sources.

Frequently asked questions