RAG Assistant That Answers Staff Questions from Your Own Company Documents

Construction and remodeling companies keep their operations manuals, sales process, and training material spread across separate documents, so field crews and office staff either dig through files themselves or stop a manager mid task to ask a question that is already written down somewhere. Putting a general purpose AI chatbot on top of that scattered material makes it worse rather than better, because the chatbot fills gaps with invented answers, and a confident wrong answer about pricing, scope, or warranty terms costs more than getting no answer at all.

1. The company's operations manuals, sales process documents, and training material are ingested and split into passages, then converted into vector embeddings using OpenAI and stored in Supabase Vector, which is Supabase's vector database built on the pgvector extension for Postgres.

2. A staff member asks a question through a Telegram bot, either as a typed message or as a voice note, and voice notes are converted to text before the search runs.

3. The workflow runs a similarity search against the Supabase vector database and pulls back only the passages that most closely match that specific question, instead of sending the entire document library to the model.

4. Those retrieved passages are passed to an OpenAI model as the only permitted source for the answer, which is what makes this Retrieval-Augmented Generation rather than a general chatbot, because the model is restricted to company material and blocked from filling gaps with its own general knowledge.

5. When the retrieved passages do not cover the question, the workflow falls back to a web search and labels that answer clearly, so the reader can see it came from the open web and not from an approved company document.

6. Suggested corrections and new information go into a human approval step, where a nominated reviewer signs off before anything is allowed to change what the system treats as company truth.

Every step above is a node in n8n, an open source workflow automation platform, so the ingestion, retrieval, answering, and approval paths can each be inspected, edited, or swapped without rebuilding the system.

Any staff member can message the assistant in their own language and get an answer drawn from your actual company documents in seconds, with a visible label on any answer that came from the web instead. Because a human reviewer gates every change to the knowledge base, the assistant cannot quietly drift away from approved company process over time.

• This runs on n8n, an open source workflow automation platform, which orchestrates document ingestion, similarity search, answer generation, and the approval step as connected nodes.
• A Supabase project is required with vector storage enabled, and this is where the document passages and their embeddings are stored and searched.
• An OpenAI API key is required, and it is used both to generate the embeddings at ingestion time and to write the final answer from the retrieved passages.
• A Telegram bot has to be created and its token supplied to n8n, because Telegram is the interface staff use to ask questions.
• A web search provider has to be connected to serve the fallback path for questions the company documents do not cover.
• The company documents have to be gathered into one place first, with a designated owner who decides which version of each document counts as current.
• At least one internal person has to be nominated as the reviewer who approves corrections, because the learning loop will not apply a change without a human signing off.

• The assistant can only answer accurately on subjects covered by the documents that were ingested, so a gap in the source material surfaces as a web fallback or no answer rather than as a fabricated one.
• Answer quality tracks document quality directly, which means outdated or contradictory source documents will produce outdated or contradictory answers.
• Staff reach the assistant through Telegram only, so it is not available as a website chat widget, an email responder, or a phone line without additional build work.
• Corrections are not applied immediately, because the approval loop deliberately places a human reviewer between a suggested change and the knowledge base.
• The assistant answers questions and does not take actions, so it cannot update a CRM record, schedule a job, or send anything on a user's behalf.
• Web fallback answers are labelled but not verified, so they carry the normal reliability limits of a web search rather than the reliability of an approved company document.