An assistant that answers from your own documentation and does not make things up

Who it is for

Companies with extensive manuals, policies, catalogs or FAQs, where the team answers the same thing many times by searching through documents.

The problem

The correct answers exist, but they are scattered across PDFs, old emails and the heads of two people. A generic chatbot confidently answers things that are not about your company. And the real fear is that it invents a return policy that does not exist.

Ingredients

• Your documents as text: manuals, policies, catalog, FAQs
• A database with similarity search (vectors), or the provider's equivalent service
• API access to a model
• A process to update the documents when they change, or the assistant goes stale

Step by step

1. 1

   Gather and clean the documents

   Convert everything to plain text or Markdown. Remove old and duplicate versions: if two documents contradict each other, the assistant will contradict itself too. This step is 60% of the work and nobody wants to do it.

2. 2

   Split into fragments with context

   Divide each document into chunks of 300 to 800 words. Prepend to each chunk a line saying which document and section it comes from. Anthropic measured that this added context reduces retrieval failures by 49%; it is the technique the source calls "contextual retrieval".

3. 3

   Store the fragments in a database with semantic search

   Each user question retrieves the 5 to 10 most similar fragments. Those fragments, and only those, are what the model can use to answer.

4. 4

   Write the prompt that forces citations and admitting gaps

   Use the prompt below. The key rule is the "I don't know" rule: without it, the model fills in with general knowledge, and that is where invented policies come from.

5. 5

   Test with the 30 most frequent real questions

   Pull them from your emails or WhatsApp. Note how many it answers correctly, how many it correctly says it does not know, and how many it makes up. The made-up ones are the only serious ones; adjust until they are zero.

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

You are the assistant for [COMPANY NAME]. You answer customer questions using ONLY the information that appears in the reference documents I give you below.

Strict rules:
1. If the answer is in the documents, answer it clearly and finish by indicating which document you took it from, in square brackets: [SOURCE: DOCUMENT NAME].
2. If the answer is NOT in the documents, reply exactly: "I don't have that information in the official documentation. I'll pass you to the team to confirm it." Do not fill in with general knowledge or assume.
3. If the documents contradict each other, say so and show both versions.
4. Never invent prices, deadlines, policies or conditions.
5. Reply in Chilean Spanish, in a warm and direct tone, 120 words maximum.

Reference documents:
[RETRIEVED FRAGMENTS ARE INSERTED HERE AUTOMATICALLY]

Customer question:
[QUESTION]

What usually goes wrong

• ✕Skipping the document cleanup. An assistant built on contradictory documents is worse than no assistant.
• ✕Fragments without context. A chunk that says "the deadline is 10 days" without saying what it is about is useless for search.
• ✕Measuring only the correct answers. What you need to bring to zero is the made-up answers, even if that means more "I don't know".