The product is answers with receipts
Staff do not need a chatbot personality. They need answers grounded in policy PDFs, with citations they can open.
Architecture we like for client work
- Ingest documents into object storage
- Extract text, chunk with overlap, store embeddings
- Postgres +
pgvector(or a managed vector store) filtered by tenant permissions - Laravel API orchestrates retrieval + model call
- UI shows answer + source snippets
Permissions are the hard part
If user A cannot download a file, they must not retrieve its chunks. Metadata filters on every query are non-negotiable.
Evaluation before launch
Maintain a question set with expected doc references. Measure retrieval hit rate and answer usefulness with a human spot-check — not only vibe tests.
Cost and latency knobs
Cache repeated questions, cap top-k, and refuse when retrieval confidence is low. “I don’t know” beats a confident hallucination.
Rayaxis builds this class of practical AI into systems clients already trust — usually beside Laravel, not as a separate science project. (Note: an earlier draft on a similar theme may exist; this batch post focuses on the delivery checklist.)