AI chatbot development
A chatbot that answers from your material, not the internet.
Acasa Labs does AI chatbot development: a custom ChatGPT-style assistant for your business that answers from your documents, data, and rules, with sources shown. NoteTube's chat streams answers grounded in each student's own uploads, with the model chosen per conversation across OpenAI, Gemini, Claude, Llama, and Mistral.
The problem
A generic chatbot is confident and wrong about your business.
Point a general model at your customers and it will answer questions about your prices, policies, and products from whatever it learned on the internet. Sometimes that is right. Often it is out of date or invented, and it says both in the same confident tone. A business chatbot needs to answer from your material, say when it does not know, keep each user's data separate, and hand off to a person when the question is not one it should answer.
What we build
- 01
Chat grounded in your content
Answers drawn from your documents, help articles, or product data, with the sources shown so a user can check them.
- 02
Streaming that feels fast
Replies that start appearing at once, with timeouts so stalled sessions do not hold resources on the server.
- 03
Scoped to each user
Customer data, uploads, and history kept separate per user or per account, enforced on the server.
- 04
Guardrails and handoff
Topics it should decline, a plain "I do not know", and a path to a person for anything that needs one.
From production
From the chat we run in production.
Grounded in the user's own sources
NoteTubeNoteTube parses videos, PDFs, DOCX, images, articles, and audio, embeds them in Qdrant with per-user, per-session scoping, and streams chat answers grounded in that material, not the open web.
Streaming that cleans up after itself
NoteTubeNoteTube's chat streams over server-sent events with inactivity timeouts, so a dead connection does not hold resources on the server.
Pick the model per conversation
NoteTubeEach NoteTube conversation can run on OpenAI, Gemini, Claude, Llama, or Mistral, and AI keys live only in a separate FastAPI service that the web and mobile clients never see.
Answers that cite the page
SigmaCPMSigmaCPM requires a quote for every value its model extracts and checks that the quote is on the cited page. The same rule is what makes a chatbot's answers checkable.
Honest trade-offs
When you do not need a chatbot
If your customers ask the same twenty questions, a well-organized FAQ and a good search box may serve them better, and cost nothing per query. If the answers live across thousands of documents and accuracy is the whole point, the work is really retrieval, which our RAG development page covers in depth. Build a chatbot when people need to ask in their own words and get an answer about their own situation.
What we reach for
- OpenAI, Anthropic, Gemini, Llama, Mistral
- FastAPI
- Server-sent events
- Qdrant, pgvector
- Next.js, Expo
- Redis
- Supabase Postgres
Proof
Shipped, not pitched.
01Our product
NoteTube
A lecture becomes a study system: chat, flashcards, quizzes, all grounded in the source.
An AI learning platform that turns PDFs, YouTube lectures, recordings, and notes into a structured notebook with grounded chat, spaced-repetition flashcards, and graded quizzes.
Web · iOS · Android · Chrome extension · MCP server
02Our product
SigmaCPM
Construction project management where AI reads the tender and shows its work.
One project record for civil contractors, from tender to closeout: bids, contracts, RFIs, vendor quotes, and drawings. The AI pulls line items and bond terms out of tender PDFs and cites the page for every value.
Web · API · Chrome extension
Questions
AI chatbot development questions we get asked.
Can you build a custom ChatGPT for our company?
Yes. It uses the same kind of model, but it answers from your documents and data, follows your rules about what it may say, and runs in your product or on your website under your brand.
Will the chatbot make up answers?
We reduce that by answering only from retrieved sources, showing those sources, and instructing the model to say when it does not know. We also test it against a set of real questions before launch.
Where does our data go?
Your documents are stored in your own database and vector index. Model calls go through a server you control, and we pick providers and settings that match your data rules.
Can it work on our website and in our app?
Yes. NoteTube's chat runs in a Next.js web app and in its iOS and Android apps, all through one backend.
Can it hand off to a person?
Yes. We can route a conversation to email, your helpdesk, or your CRM when the user asks, or when the question is outside what the bot should answer.
How is this different from your RAG development service?
RAG is the retrieval engine: ingestion, search, reranking, and evaluation over a large body of documents. A chatbot is the product people talk to. Many chatbots need RAG underneath. Some only need your product data and a few rules.
AI chatbot development
What would your customers ask it first?
Bring ten real questions and the material that answers them. We will show you how a chatbot would handle them and where it would need a person.