A small chatbot for the team's recurring questions

The same questions arrive every week: where is the expense form, who approves a software purchase, and which checklist applies to a release? A small internal chatbot can help people find those answers without asking a colleague to repeat them. The useful starting point is not a company-wide assistant. It is a narrow tool that answers a few recurring questions from documents your team already trusts.

§ Choose one corner of the team's knowledge

Start with a single subject, such as onboarding or everyday IT requests. Gather a small set of current documents and give each one a clear title, an owner, and a review date. Remove outdated copies before connecting anything to a model.

This preparation matters more than a clever chat interface. If two documents disagree about an approval process, the chatbot cannot reliably decide which policy your team intended to keep.

Write down ten real questions people ask. Include a question the documents cannot answer. These become your first test set, including a test of whether the bot can admit that it does not know.

§ Understand the few pieces involved

A basic setup has a private chat page, a small server application, a collection of approved documents, and a model API. The server receives a question, finds relevant document passages, and sends those passages with the question to the model. The model drafts an answer, which the server returns to the chat page.

This document-finding step is often called retrieval. It does not mean training a new model. You are supplying reference material for each request. For a tiny collection, simple keyword search may be enough. More advanced search can wait until real questions show that you need it.

Vana API provides an OpenAI-compatible endpoint at https://api.vanaapi.com, so an application using an OpenAI SDK can connect by configuring its base URL, API key, and model. It also supports Anthropic Messages and Gemini native protocols. You do not need to use all three; choose the interface your application already understands.

§ Give the assistant a modest job

Use a short instruction that defines the boundary: answer from the supplied team documents, name the supporting document, and say when the available material is insufficient. Ask for brief answers and a practical next step where appropriate.

For an initial trial, a model such as Gemini 2.5 Flash is one option to evaluate. Test it against your actual questions rather than assuming that a model name guarantees useful results.

• Supported question: Give the answer and identify the document and section that support it.
• Missing information: Explain what is missing and suggest the appropriate document owner.
• Conflicting information: Point out the disagreement instead of silently choosing an answer.

Instructions help, but they are not a guarantee. Check that the cited material actually supports each answer. Treat retrieved documents as reference text, not as instructions that can change the chatbot's rules.

§ Keep access and secrets outside the model

Require team sign-in and enforce document permissions on the server before retrieving passages. A person should not receive information from a document they cannot open themselves. Do not rely on a prompt saying that certain information is confidential.

Keep the API key on the server, never in browser code. Begin with low-sensitivity material, and review the applicable data handling and retention terms before sending internal content to any external service. Decide whether chat logs are necessary, who can read them, and when they should be deleted.

§ Run a small, measurable trial

Invite a few colleagues to use the bot for a week. Record unanswered questions, unsupported answers, response times, and request costs. Keep early versions read-only: finding a purchase policy is a safer first task than approving a purchase.

Vana API bills per real token with no monthly fee. Tokens are pieces of text processed by the model; both document excerpts and generated answers affect usage. Send only relevant passages, limit answer length, and track spending in your application.

Create a Vana API account when you are ready to test the connection. Expand the chatbot only when the first set of questions works reliably. A useful internal assistant is one your team can verify, correct, and maintain.