Air-Gapped & Private AI Reading: Yomitomo + Ollama Local LLM Workflow

As Large Language Models become integral to research, due diligence, and code review, a pressing paradox arises: the higher the sensitivity of the reading material (unreleased patents, proprietary M&A contracts, cutting-edge drafts), the less acceptable it is to upload it to commercial cloud APIs.

Furthermore, during international flights or inside classified network zones, cloud-dependent AI reading tools fail entirely.

Yomitomo natively supports local inference engines like Ollama and LM Studio, enabling an air-gapped, zero-data-leakage, fully offline analytical reading environment.

Local Private AI Reading Architecture

Pipeline Node	Input	System Processing	Output Deliverable	Security & Boundary
Model Hosting	Local Ollama service running `qwen2.5` or `deepseek-r1`	Local CPU/GPU hosts `http://localhost:11434` HTTP endpoint	High-speed offline inference ready on device	Zero external network calls; model weights reside on local disk
Endpoint Setup	Configure `http://localhost:11434/v1` in Yomitomo	Validates local endpoint and maps to companion personas	Local models assigned to `@ZhouYan`, `@HeMingheng`, etc.	No external API keys required; zero token subscription fees
Passage Debate	Select text, press `A`, mention companion in thread	Yomitomo extracts anchor context and POSTs prompt to `localhost`	Millisecond-level logic critiques and concept clarifications	Data packets circulate strictly within local loopback
Synthesis Card	Press `T` to open Distillation Studio and audit draft	Local LLM checks reasoning gaps and refines prose	High-density Markdown distillation cards in SQLite	Knowledge assets archived offline for permanent reuse

4 Steps to Configure Yomitomo + Ollama for Offline Reading

Step 1: Install and Launch Ollama

Visit Ollama’s official site to download the installer for macOS, Windows, or Linux.

In your terminal, pull your preferred analytical model:

# Recommended for bilingual non-fiction & concept parsing:
ollama run qwen2.5:14b

# Recommended for intense logical auditing & reasoning:
ollama run deepseek-r1:14b

Step 2: Configure Custom Provider in Yomitomo

In Yomitomo, navigate to Settings > AI Models.
Click Add Provider and set the following parameters:
- Provider Type: Custom / OpenAI Compatible;
- Base URL: http://localhost:11434/v1;
- API Key: Enter any placeholder string (e.g., ollama);
- Model Name: Enter your pulled model tag (e.g., qwen2.5:14b or deepseek-r1:14b).
Click Test Connection and save upon success.

Step 3: Map Local Models to Reading & Review Companions

Under Settings > AI Assistants, assign your local Ollama model as the default engine across all reading and review companions.

Step 4: Air-Gapped Deep Reading

Turn off Wi-Fi or switch to airplane mode, then open an EPUB or PDF paper:

Select a paragraph and press A;
In the thread, type: @ZhouYan Audit the necessary conditions for this causal deduction;
Your local GPU performs rapid inference, generating rigorous critiques right in your margin;
Press T to distill your conclusions into permanent knowledge assets.

Recommended Hardware & Model Sizing

Hardware Setup	Recommended Model	Strengths	Expected Inference Speed
Apple Silicon (16GB Unified RAM)	`qwen2.5:7b` / `deepseek-r1:8b`	Rapid concept explanations, standard logical audits	30–45 tokens/s
Apple Silicon (32GB+ Unified RAM)	`qwen2.5:14b` / `deepseek-r1:14b`	Deep academic critiques, rigorous evidence auditing	20–35 tokens/s
PC + RTX 4070/4080 (12–16GB VRAM)	`qwen2.5:14b-instruct`	Blazing-fast document analysis and debate	40–60 tokens/s

Target Audience & Usage Boundaries

Who This Is For

Researchers, legal analysts, and engineers handling sensitive data under strict non-disclosure terms;
Knowledge workers needing uninterrupted study workflows during flights or in offline research spaces;
Power users with modern hardware looking for zero-cost, unlimited local token generation.

What This Is Not For

Low-spec machines (e.g., 8GB RAM without dedicated GPU): Running 14B models locally may cause sluggishness; lighter 7B models or cloud APIs are advised;
Real-time live web search: Local models focus on logical deconstruction of the provided passages rather than real-time web querying.

Frequently Asked Questions (FAQ)

Q1: Does running local models drain my laptop battery rapidly?

Answer: Yomitomo only triggers local inference when you deliberately invoke an assistant with @ or request a distillation review (generating only a few hundred tokens per call). It does not run continuous background workloads, minimizing battery impact.

Q2: Do I need to configure firewall rules for Ollama?

Answer: No. Ollama listens on http://localhost:11434. As long as Yomitomo runs on the same machine, communication travels across local loopback without modifying firewall policies.