# An AI chatbot that answers from your company’s documents

> An AI assistant that knows your manuals, procedures, contracts and history, answers with the source cited, and says “I don’t know” when the answer is not there. For customers or for your internal team, inside the software you already use.

- URL: https://mojtabaamini.com/services/ai-chatbot
- Language: en
- Author: Mojtaba Amini, AI Software Engineer (Verona, Italy)

## Why most chatbots disappoint

A generic chatbot gives plausible answers that are not in your documents. That is how many projects fail in production: plausible-but-wrong is worse than “I don’t know”.

A useful company chatbot answers only from approved sources, cites them, and refuses when it cannot find confirmation. That is an architecture problem, not a model problem.

## What I build

- **Customer support** that answers from technical manuals, catalogues and FAQs, with a page reference.
- **Internal assistants** for your team: procedures, contracts and past decisions, without interrupting a colleague.
- **Copilots inside your ERP and CRM** that answer on your operational data at the point where it is needed.

## How a RAG chatbot that does not make things up works

1. **Structure-aware indexing**: Documents are split while keeping their structure (manual, chapter, section), so one product’s procedure is never confused with another’s.
2. **Hybrid search**: Semantic and keyword search work together, which is essential for part numbers and technical terms, and a re-ranking model picks the passages that really matter.
3. **Answers with sources**: The model answers only from the retrieved passages and cites the source, for example the manual and page.
4. **Calibrated refusal**: If no passage supports the answer, the assistant says so and hands the question to a person.

## Real projects

For Extrema I built a chatbot that answers from about 1,400 pages of internal technical manuals: internal customer-service response times dropped by 60%.

- [Stairlift configurator with GNN cost & lead-time prediction](https://mojtabaamini.com/projects/extrema-en) (Extrema · Mar 2024 — Dec 2025): DNN + GNN turn customer specs into a complete bill of materials and an accurate delivery quote in minutes. LLM chatbot cuts internal response times by 60%.
- [AI co-pilot for ERP & production planning](https://mojtabaamini.com/projects/dekali-en) (Dekali (ex Virevo) · Jul 2023 — Sep 2026): LLM assistant + RL scheduler embedded in a manufacturing ERP. Cuts cross-team coordination from hours to seconds.

## Technology

LLM, RAG, Hybrid search, Re-ranking, Azure, Python, .NET, Angular

## Further reading

- [A RAG pipeline that actually answers from the manual](https://mojtabaamini.com/writing/rag-that-works-en) (Article · Oct 2025): Lessons from shipping a customer-support chatbot grounded in 1,400 pages of Extrema manuals — and reducing internal response times by 60%.
- [How I expose Python models behind a .NET API consumed by Angular](https://mojtabaamini.com/writing/angular-dotnet-ai-en) (Article · May 2025): The Dekali stack is .NET + Angular with Python models behind the curtain. Here is the wiring that keeps latency under 200 ms on the order list.

## Frequently asked questions

**Can the chatbot make up answers?**

That is the main risk with generic chatbots. Here the assistant answers only from approved sources, cites the document and, when it finds no confirmation, says it does not know and involves a person.

**Which documents can it work with?**

Technical manuals, catalogues, internal procedures, contracts, FAQs, tickets and email history, as PDF, Word, web pages or databases.

**Is it for the website or internal use?**

Both. It can be a customer assistant on your website or customer portal, or an internal tool built into your ERP, CRM or intranet.

**Do company documents stay confidential?**

Yes: we can use models on European cloud with contractual data guarantees, or models running on your own servers, and document access can follow the permissions you already have.

**How long does it take?**

A first assistant on a defined set of documents can be tried within a few weeks, measuring from day one how many answers are correct and how many are rightly refused.

## Let’s talk about your project

Tell me about the process you want to improve: I will reply with a first assessment of where AI can help and where to start.