An AI assistant that answers from your own documents, on web and mobile.
People ask. Apps stay silent.
Most apps hold the answer somewhere: in a help page, a PDF or a database. Users still cannot find it. They type a question and get a list of links.
A plain chatbot does not fix this. A large language model (LLM) on its own guesses. It has never read your documents.
One API. Two apps.
I built one back-end API and put the same assistant in a website and in a mobile app. The API reads your documents, cuts them into small chunks and turns each chunk into an embedding: a list of numbers that stores its meaning.
When a user asks a question, the API finds the chunks closest in meaning and hands them to the LLM with the question. This pattern is called retrieval-augmented generation (RAG). The model answers from your text and shows its sources.
- 01Question
- 02Embedding
- 03Vector search
- 04Top chunks
- 05LLM
- 06Answer and sources
Five ideas do the work
- Embeddings
- Each chunk of text becomes a vector. Close vectors mean close meaning.
- Vector database
- Stores the vectors and finds the nearest ones in milliseconds.
- RAG
- The model reads the retrieved chunks before it answers, so it does not guess.
- Streaming
- Words appear as they are generated. Users do not wait for the full reply.
- One API, two clients
- The web app (React) and the mobile app (React Native) call the same endpoint.
Try the retrieval step
This runs in your browser. Type a question about me. The page scores every note by cosine similarity and shows the closest ones. No model is called: this is only the “R” in RAG.
Vectors: TF-IDF. Score: cosine similarity. Runs in your browser.
The design in numbers
- 1 API
- serves both the web app and the mobile app
- 2 clients
- React on the web, React Native on mobile
- < 2 s
- target time to the first streamed word
What comes next
Tool calling, so the assistant can act: book, refund, update. And an evaluation set, so every change is tested against known questions before it ships.