› can a chatbot answer from our own documents
AI & LLM Integration
Models grounded in your data, not in their own imagination.
Retrieval-augmented assistants, custom conversational models and evaluation harnesses that tell you when quality drops. Including Nova — the assistant in the corner of this page, which answers from this site’s own text and says so when it can’t.
What we hold ourselves to
100%
Grounded answers cited
2
Answer lanes, kept apart
4
Model providers, with failover
These are the budgets we build against and measure on handover — not a warranty. Where a target can’t be met we tell you which one and why, before you sign.
Evidence
One internal build, shipped: Nova, the assistant in the corner of this page. No client work yet.
What this actually includes.
Custom Chatbots
Assistants that know your product and admit when they don’t.
- Tone, scope and refusal behaviour defined as a spec, not a vibe
- Tool calling into your real systems
- Conversation logging with PII redaction
- Fallback to a human with full context handoff
RAG Systems
Answers with citations, or no answer at all.
- Chunking and embedding strategy tuned to your corpus
- Hybrid vector + keyword retrieval with reranking
- Every claim traceable to a source document
- Freshness pipeline so the index never silently goes stale
Guardrails & Failover
An assistant that cannot invent your prices, and does not go down.
- Isolated answer lanes — the lane that could invent a price cannot discuss prices
- A retrieval score floor, below which it says "I don’t know" rather than guessing
- Provider failover across separate accounts, so one outage is not an outage
- Conversation memory the server signs, so a client cannot forge what was said
Capabilities
- RAG
- Tool calling
- Evals
- Fine-tuning
- Failover
Typical stack
- NVIDIA NIM
- Llama 3.3
- BM25 retrieval
- AWS Lambda
- Evals
Live on this page
Meet Nova. She’s already running.
Nova is in the corner of every page on this site. Ask her anything — she will answer a general question like any assistant, and an Unova question strictly from what is written on these pages. Ask her what something costs and she will tell you she won’t invent a number, because she structurally cannot.
That split is the engineering. Her backend crawls this site, ranks the passages that actually bear on your question, and hands only those to the model — and if nothing scores above the floor, a second lane answers generally and is forbidden from saying anything about Unova at all. The lane that could invent a price is the lane not allowed to discuss prices. Retrieval, a scoring floor, two isolated lanes, provider failover: that is what we ship when a client says their chatbot must not make things up.
Opens the panel in the corner. It is on every page of this site, not just this one.