Implement AI Assistant to our saas web & mobile apps to answer user queries based on their own data

Our product is a SaaS app where customers keep their own records and documents. They constantly ask us questions that are answerable from that data, and today a human answers every one. We want that answered inside the app. What it must do Answer a customer's question using only their own data: their records plus the documents we load. Show the source of every answer (document name + page, or the record it came from). Say clearly "that isn't in your documents" instead of guessing. Stream the answer into a panel inside our existing React app, behind our current login. Keep a log of every question, the answer, the sources used, and the token cost. Deliverables An ingestion command that loads 50 documents (PDF and web pages) into a vector store and can re-index them. A /chat endpoint: question + customer id in, streamed answer + sources out. The assistant panel in our app (React), with the sources toggle. Guardrails: refuses anything outside the loaded data, caps tokens per answer, throttles per user. An admin page to edit the system prompt and read the question log. A README with setup, .env.example, and the re-index steps, plus a 2-3 minute walkthrough. Acceptance criteria 10 written test questions: at least 8 answered correctly, and every correct answer cites at least one source. 3 out-of-scope questions are refused with the fixed message, not answered from general knowledge. A question whose answer is absent returns "not in the documents" - verified on one document we hold back. First token arrives within 2 seconds on our test data; no key or model name hard-coded. Everything runs from a fresh clone with the README. Timeline Day 1 ingestion + index, Day 2 retrieval + cited answers, Day 3 endpoint + streaming + auth, Day 4 the in-app panel, Day 5 guardrails + log + admin, Day 6 the 10-question test, docs and handover. What to include in your proposal A link to one similar build, the stack you would use for retrieval, and how you would measure whether the answers are grounded.