Infrastructure layer
Two providers run specific technical functions in the background. OpenAI processes desktop embeddings so your vault can be searched. Deepgram transcribes audio for Meeting Copilot and chat mic.
Models Hub
AI models inside ZSearch, without the lock-in.
Groq, Claude, Perplexity, DeepSeek, Kimi, Gemini, or Ollama. Pick the provider that fits the work, then switch when it changes.
How it works
Two layers, one clear boundary
ZSearch separates background infrastructure from the AI provider that answers your questions, so you always know what is running and why.
AI query layer
This is where model choice lives. Groq, Claude, Perplexity, DeepSeek, Kimi, Gemini, and Ollama can each power AI chat, document questions, and AI Lens search.
Always on
Infrastructure providers
These run automatically when their feature is used. No provider selection is needed for this layer.
Embeddings
OpenAI
- What it powers
- Makes your Knowledge Vault searchable so ZSearch can find the right passages when you ask a question.
- Why it is here
- OpenAI embeddings support the desktop search layer. They run when documents are added, not every time you ask a question.
- Best for
- All ZSearch Desktop users. Enterprise can run embedding compute inside your own infrastructure.
Voice transcription
Deepgram
- What it powers
- Transcribes microphone and system audio for Meeting Copilot and chat mic input.
- Why it is here
- Deepgram is built for accurate, real-time transcription, including multiple speakers.
- Best for
- Anyone using Meeting Copilot or voice input. If you do not use voice features, Deepgram is not called.
AI query providers
Route each question to the model that fits the work.
These providers power AI chat, document questions, and AI Lens search. Choose for the outcome you need, then switch providers when the task changes.
Speed
Groq keeps rapid research sessions moving.
Judgment
Claude is better when answer quality carries risk.
Sovereignty
Ollama keeps AI queries on your machine.
Lowest latency
Groq
Fast inference
AI chat and document questions with very low latency, so rapid follow-up questions feel natural.
Use when
Fast Q&A, high-volume document queries, and everyday research.
Highest confidence
Anthropic Claude
Careful reasoning
Longer, more structured responses for complex document analysis and high-stakes questions.
Use when
Legal, research, finance, compliance, and analytical document work.
Current citations
Perplexity
Live web search
Combines private document answers with current web results and source citations.
Use when
Journalists, analysts, consultants, and researchers who need cited current answers.
Local-only path
Ollama
Local AI queries
Connects ZSearch to a locally running LLM with no API key and no external AI query provider.
Use when
Maximum privacy, air-gapped workflows, security-sensitive teams, and local model control. AI queries stay on your machine. The only remaining external call is Deepgram, used only if you use Meeting Copilot or voice input.
Volume friendly
DeepSeek
Efficient reasoning
Fast chat and step-by-step reasoning for large document workloads.
Use when
High-volume processing, lower-cost deployments, and transparent reasoning workflows.
Long document memory
Moonshot Kimi
Large context
Long document analysis with very large context windows, built for contracts, books, and research archives.
Use when
Long documents, complex reasoning, and high-quality answers without flagship pricing.
Visual reasoning
Google Gemini
Vision and large context
Reads and reasons over images, charts, diagrams, scans, and very large collections.
Use when
Healthcare, architecture, finance, design, and visually rich document workflows.
Which model should you choose?
Do not overthink the first choice. Start with the provider closest to your workflow and switch later if your needs change.
I just want to get started.
Choose Groq.
Fast, simple, and strong enough for most everyday document questions.
I am doing serious analytical work.
Choose Anthropic Claude.
Best for contracts, compliance, research, finance, and answers with real consequences.
I run a small team handling sensitive documents.
Start with Claude.
Use Claude when quality matters, or choose a lower-cost option if budget matters more.
I need current web information.
Choose Perplexity.
It is the provider that combines your private vault with live web search.
I want great quality at lower cost.
Choose DeepSeek or Kimi.
Start with DeepSeek. Move to Kimi when very long documents are common.
My documents include images or scans.
Choose Google Gemini.
Gemini can reason about visual material, not just extracted text.
I want AI queries fully local.
Choose Ollama.
Your AI questions stay on your machine. Voice still uses Deepgram only when voice features are used.
Enterprise control
Sovereignty is a deployment choice, not a provider promise.
ZSearch Enterprise keeps model choice flexible while moving the sensitive compute layer inside your infrastructure. Use managed providers where they make sense, and local models where nothing should leave the perimeter.
Discuss deploymentEmbedding compute
Runs on your own infrastructure in Enterprise.
Local AI queries
Use Ollama when answers must stay on the machine.
Provider optionality
Route work to Groq, Claude, Perplexity, Gemini, and more.
Perimeter control
Keep sensitive workflows aligned to your network boundary.
No lock-in
Switch providers as requirements change.
No hidden layer
Understand which compute runs where.
No forced cloud
Enterprise can keep every layer inside your boundary.
Your AI. Your models. Your call.
Start with one provider, switch when the work changes, and keep enterprise deployment open for full infrastructure control.