- Proof of concept · open source
- Public repository (source: github.com/dani-b-g/chatbot_angular_rag)
ITXample Shop
A customer support assistant with local AI that cites the documents it is based on. No paid APIs or keys.

- Who I built it for
- Teams that want to evaluate a support chatbot with an LLM without sending data to third parties or paying per use.
- What problem it tackles
- Trying a support assistant with AI usually requires keys and paid APIs, and its answers do not say where they come from.
- What I built
- A monorepo with the store in Angular 21 and the API in NestJS 11: local RAG over Markdown documents, generation with Ollama and a streamed answer with its sources. A single docker compose up starts everything.
- Where to try or review it
- On GitHub. It runs locally with Docker; the store, the catalog and its policies are fictional. View the repository (opens in a new tab)
From the question to an answer with sources
A customer asks about a return. This is the path her question takes inside the proof of concept.
Example of how it works
Fixed data to show the flow. It is not the model answering live. Translated from Spanish for this page. The proof of concept’s documents and answers are in Spanish.
Question: I bought headphones two weeks ago and I’m not happy with them. Can I return them?
Angular interface. The chat sends the question and a short conversation history to the NestJS API.
Technical detail
Local retrieval
On startup, the API splits the Markdown documents into fragments by section and computes their embeddings with embeddinggemma. Each question is compared by cosine similarity and the 3 closest fragments are used.
Generation and streaming
NestJS builds the prompt with the support instructions, a short history and the retrieved context, and sends it to qwen3:4b-instruct on Ollama. Tokens reach Angular through Server-Sent Events without exposing Ollama to the browser; nginx forwards /api without buffering.
What is missing for production
The repository says it plainly: authentication, rate limiting, a vector database and evaluation of answer faithfulness. The sources are retrieved documents, not verified claims.
Technologies in this project. Each one links to what I have done with it:
The store, the catalog and the policies are fictional.