From half-day ticket hunts to answers in seconds.
An IT services company running Alvao ITSM with 50,000+ historical tickets.
With 50,000+ historical tickets containing solutions, configurations, troubleshooting procedures, and client-specific knowledge accumulated over years of delivery.
THE PROBLEM
Alvao's native search relied on basic keyword matching — semantically relevant tickets weren't surfacing. The knowledge was there; it was just unreachable at the speed the team needed.
Concrete pain points:
- Sales teams spent 2–3 hours searching old tickets when preparing new offers.
- Technical staff couldn't quickly find proven solutions from past cases.
- Valuable institutional knowledge was buried in ticket histories.
- New employees had no efficient way to learn from past experience.
OUR APPROACH
We built a custom RAG (Retrieval-Augmented Generation) assistant that connects directly to the Alvao database. Ticket content is indexed into a vector store, enabling semantic search that understands meaning rather than just keyword matches.
Users can ask questions in natural language ("How did we solve the VPN issues for manufacturing clients?"), get synthesised answers with citations back to source tickets, and jump straight to the original ticket to verify. The system reindexes new and updated tickets on schedule so answers stay current.
KEY TECHNICAL DECISIONS
- Qdrant vector store, on-premise. The client required data sovereignty — tickets contain sensitive client configurations. Qdrant runs inside their infrastructure; no data leaves their network.
- Ticket-aware chunking. Standard text splitters broke ticket structure; a custom chunker preserves description, comments, and resolution as separate chunks with shared metadata — retrieval quality jumped.
- Hybrid search + metadata filters. Semantic similarity combined with metadata filters (client, date range, category) enables both broad exploration and scoped searches.
- Incremental sync. A scheduled job indexes new and updated tickets without a full reindex — the vector store stays fresh at low cost.
- Citation tracking. Every generated answer includes source ticket IDs so users can verify — critical for a knowledge-worker audience that mistrusts unattributed AI.
RESULTS
- Offer preparation time: from ~2–3 hours to under 1 hour on typical proposals.
- Sub-second query response times on the vector index.
- Onboarding acceleration for new engineers. New joiners can query institutional knowledge instead of asking the same senior engineers repeatedly.
Have a knowledge base nobody can find anything in?
— A few seconds to fill in a form and we'll do the rest!
Discovery call, 45 minutes. Bring your data sources + a few of the questions your team wastes time answering. We'll tell you what a RAG assistant would actually change.