Build a company knowledge base that answers by citing the document

Who it is for

Mid-sized companies with many procedures, manuals and contracts spread across folders, and a team that asks the same things over and over to the two people who know where everything is.

The problem

Knowledge lives in folders and in the heads of a few people. Finding a procedure takes longer than doing it wrong. And generic chatbots confidently answer things that are not in any company document.

Ingredients

• A server with Docker (your own or in the cloud)
• The documents the team actually consults, organized by folder and with clear names
• An API key for the AI model
• One technical person for the installation and one from the business side to choose and test the documents

Step by step

1. 1

   Choose twenty documents and twenty questions

   The twenty most consulted documents and twenty real questions from the team with their correct answers. This is your exam: the knowledge base passes or fails on these questions.

2. 2

   Bring up the platform with Docker Compose

   Follow the repository's getting-started guide. Leave the model key configured and create a test knowledge base.

3. 3

   Upload the twenty documents and run the exam

   Ask the twenty questions. Measure how many it answers correctly and, above all, whether it cites the right document. If it fails on several, review how the documents are written before touching the configuration.

4. 4

   Connect it where the team works

   According to the README, there are integrations with Telegram, Slack, WeChat and others, and connectors to sync Notion or GitLab. Start with one channel and one pilot group.

5. 5

   Define who maintains the documents

   A knowledge base with old documents confidently answers old things. Someone owns each folder and updates it when the procedure changes.

Replace what is in brackets with your company details, or use the guided mode above: it asks you and fills it in for you.

# Role
You are the internal knowledge assistant for [COMPANY NAME]. You answer only with information from the indexed documents and you always cite the document and section. When the answer is not in them, you say so.

# Context
The documents cover [WHAT THE DOCUMENTS COVER, FOR EXAMPLE "warehouse procedures, return policies and supplier contracts"]. The people asking are [WHO ASKS, FOR EXAMPLE "the operations and sales team"]. The owner of the documents, for questions you cannot resolve, is [WHO MAINTAINS THE DOCUMENTS].

# Task
Answer the question in two to five sentences, cite the document and the section it comes from, and if two documents contradict each other, say so and show both.

# Output format
Brief answer → "Source: document name, section or page". If it is not on record: "Not in the documents; ask [WHO MAINTAINS THE DOCUMENTS]".

# Examples
Question: "How many days does a customer have to return a faulty product?". Expected answer: "Thirty days from delivery, presenting the receipt. Source: Returns Policy v3, section 2".

What usually goes wrong

• ✕Uploading the whole shared drive. Answer quality drops with duplicate documents, drafts and old versions; curate before uploading.
• ✕Measuring with made-up questions. Use real questions the team asked in the last month; those are the ones that matter.
• ✕No document owner. In three months the knowledge base is answering with procedures that have already changed.