We run Apache Answer as the help centre for two client products. Instead of waiting for native AI chat to mature, we built a RAG layer alongside it and have been running it in production since April. Sharing numbers because I haven't seen many real-world results posted here.
Setup: Answer stays the source of truth (articles + Q&A). A separate pipeline ingests those docs plus PDFs/tables/images (Docling), embeds into Qdrant, and serves a chat that answers only from the docs, with citations — every claim links to its exact source chunk. Login reuses Answer's auth tokens.
5-month results (real production, ~8-person support team):
- ~50% of incoming questions handled by AI without a human
- ~$5,200 saved to date, ~$530/month at current run rate
- Satisfaction on AI answers: 100% of rated responses positive so far
- Bonus we didn't expect: unanswered questions cluster into a weekly "doc gaps" digest — it now drives what we write next
What Answer alone couldn't do for us: scanned PDFs and complex tables, per-answer citations, and knowing which docs are missing.
Happy to share architecture details. Also curious: for those following the native AI features (semantic search, Copilot chat) — would you use a companion service like this on top of your instance, or wait for core? Trying to gauge if this is worth packaging for others.