AI, OpenRouter and Pricing
Leveraging AI
A Precision-Engineered Pipeline for Your Personal Knowledge
pinakea does not just “use AI.” It orchestrates a sequence of specialized stages and model choices so raw items become structured knowledge you can search, summarize, and chat with.
This page explains the pipeline architecture and why each stage exists.
Two AI Modes: Online and Mixed
pinakea supports two operating modes. Online Mode is recommended; Mixed Mode remains available for smaller Sets when local automatic item processing matters more than speed and output quality.
Online Mode (Recommended)
In Online mode, the full pipeline runs through cloud models via OpenRouter.
- Uses top-tier cloud models for embeddings and pipeline generation.
- Supports parallel processing for large imports.
- Delivers the highest overall quality.
- Requires internet access and a valid OpenRouter API key.
Mixed Mode
In Mixed mode, high-volume pipeline stages run locally on Apple Silicon, while chat and deep summaries stay cloud-based. Mixed Mode is for smaller Sets when avoiding OpenRouter cost for automatic item processing matters more than speed and output quality.
- Local embeddings + micro summaries + titles + tags.
- Sequential processing tuned for local GPU/memory constraints.
- No OpenRouter cost for automatic item processing.
- Core pipeline stages can continue offline.
- Can take hours or days on larger Sets and keep your Mac busy.
pinakea uses around 1,000 items per Set as the recommended Mixed Mode guidance point. See Use Online Mode.
Seamless Switching
Switching modes switches the full pipeline profile, not just one model.
- Embedding dimensions differ (cloud/local), so vector spaces are not interchangeable.
- Online -> Mixed replaces Online embeddings generated with OpenRouter credit; those embeddings cannot be reused for Mixed Mode. Existing summaries, titles, and tags are kept.
- Mixed -> Online clears local embeddings, AI-generated summaries, AI-generated titles, and LLM tags, then regenerates them through OpenRouter credit.
- Starred items and explicit/user tags are preserved.
For practical mode behavior details, see Online and Mixed Modes. For the current recommendation, see Use Online Mode.
The AI Pipeline: From Raw Content to Refined Knowledge
When you add sources (folders, clips, mail, notes), each item flows through a staged pipeline.
Stage 1: Content Extraction
pinakea first normalizes content into clean text:
- Markdown/text: direct parse.
- PDFs: native extraction, with OCR fallback for scanned pages.
- Images/screenshots: OCR via Apple Vision.
- YouTube clips: transcript retrieval so spoken content becomes searchable.
- Web pages: readability extraction to remove noise and keep article content.
Stage 2: Chunking
Long content is split into overlapping chunks so semantic retrieval remains accurate across boundaries.
- Typical chunk size is tuned for retrieval efficiency.
- Overlap preserves context between adjacent chunks.
Stage 3: Embedding
Each chunk is converted into a semantic vector.
| Mode | Model | Details |
|---|---|---|
| Online | Qwen3 Embedding 8B | Cloud, high-dimensional semantic retrieval |
| Mixed | mxbai-embed-large-v1 | Local Apple Silicon embedding model |
This is what allows meaning-based search (“delivery schedule” matching “project deadlines”).
Stage 4: Concise Summary
Each item receives a short summary for fast scanning.
| Mode | Model | Details |
|---|---|---|
| Online | Gemini 2.5 Flash Lite | Cloud, high throughput |
| Mixed | Qwen 2.5 7B Instruct | Local via llama.cpp |
Stage 5: Intelligent Title
pinakea generates better descriptive titles from content, so your timeline remains readable at scale.
Stage 6: Automatic Tags
pinakea generates conceptual tags (not simple keyword extraction) to improve browse and retrieval flows.
Why These Model Choices
Embeddings: Retrieval Quality First
Embedding quality determines search and chat grounding quality. pinakea prioritizes:
- semantic depth over lexical matching,
- robust retrieval at large library sizes,
- consistency within each mode’s vector space.
Pipeline Generation: Speed + Consistency
For summaries/titles/tags, pinakea optimizes for:
- throughput during ingest,
- stable output quality,
- predictable cost behavior across long-running pipelines.
Chat: Conversations with Your Knowledge
Search gets you to relevant items. Chat synthesizes across them.
Retrieval Flow
- Your question is embedded into the same semantic space as your indexed chunks.
- Retrieval selects the most relevant chunks by semantic proximity.
- Multi-turn context keeps seed + incremental evidence aligned.
- Citations tie generated claims back to source items.
Citation Guarantees
Answers include clickable references so you can jump directly to supporting source content.
Summaries at Two Levels
Concise Summaries (Per Item)
- Precomputed during pipeline processing.
- Optimized for quick scanning in timeline workflows.
Full Summaries (On Demand)
- Generated for single items, days, or dayparts.
- Uses chat-grade models for deeper synthesis.
- Includes source-linked citations.
Privacy and Data Flow
pinakea is local-first by design.
- Source content stays on your Mac.
- In Mixed mode, core pipeline stages can run locally.
- Cloud-dependent operations (for example chat/deep summaries) send only required content over encrypted connections to model providers via OpenRouter.
Cost Control
Cloud usage is BYOK through OpenRouter:
- you control spend limits,
- you see spend in OpenRouter + in-app status reporting,
- no pinakea markup on model usage.
Detailed cost examples: LLM Cost (OpenRouter).