The AI stack
Models, tools, memory, and how the assistant answers questions about your own health.
The AI companion looks simple from the outside — you ask, it answers. Underneath, several systems work together to make those answers grounded, useful, and specific to you.
Models, routed centrally
AverCare doesn't hard-wire itself to a single AI model. Requests flow through a gateway that lets us pick the best model for each job — a fast, cheap model for quick suggestions; a stronger one for careful analysis — and switch providers without rewriting the app.
Tools: the assistant can do things
A plain chatbot can only talk. AverCare's assistant is equipped with tools — capabilities it can choose to use mid-conversation. When you say "log my lunch," it doesn't just acknowledge you; it calls a nutrition tool that actually writes the entry. Tools cover logging meals and workouts, reading your vitals, managing your calendar, searching your documents, saving memories, starting a scan, and more.
Different modes, different tools
In voice and video conversations, AverCare swaps in a tailored, smaller set of tools — dropping things that don't make sense to speak aloud (like image generation) and keeping the ones that do. Same brain, context-appropriate hands.
RAG: answering questions about your documents
When you ask about something specific — your lab results, an uploaded care plan — the assistant uses retrieval-augmented generation (RAG):
Find by meaning
Your documents are split into chunks and converted into vectors (mathematical fingerprints of meaning). Your question becomes a vector too, and we find the closest chunks.
Re-rank for relevance
A second model re-orders the candidates so the most relevant passages rise to the top — not just the ones that happened to share words.
Answer with citations
The assistant writes its answer using those passages and shows you where the information came from.
Fun fact
"Find by meaning" is why you can ask about "my cholesterol" and get the right answer even if your lab report says "lipid panel." Vector search matches concepts, not exact words.
Memory: high-signal, not a transcript
The assistant keeps a compact set of core memories — the handful of facts worth injecting into every conversation — and a larger pool of notes it can search on demand. It's deliberately small and curated, so it sharpens responses without drowning the model in noise. You can list and delete memories anytime.
Driving a video avatar with our own model
For AI professionals, the video provider handles the face, voice, and real-time streaming — but the thinking is still ours. The avatar's words come from the same AverCare brain, so it has your memory and your context, just delivered through a different medium.