Architecture

Insight is built as a Tauri 2.0 desktop application with a Rust backend and Svelte frontend.

Core Stack

Agent Architecture

User Query
    ↓
Local LLM (via mistralrs)
    ↓
Tool Calling Loop
    ↓
Synthesized Answer (with citations)

The agent has tools for searching and reading documents. It iteratively gathers evidence to answer questions, citing sources along the way. There is no direct user-facing search—all document retrieval happens through the agent.

Data Model

Collections as Namespaces

Each collection is an iroh-docs namespace. Sharing a collection means sharing namespace access.

Namespace: 7f3a8b2c... ("Climate Research")
│
├── files/abc123     → blob with metadata JSON
├── files/def456     → blob with metadata JSON
└── _collection      → blob with collection settings

Document Metadata

Document metadata is stored as a blob and referenced by an entry in the namespace:

{
	"name": "paper.pdf",
	"pdf_hash": "blake3-hash-of-pdf",
	"text_hash": "blake3-hash-of-extracted-text",
	"tags": ["research", "climate"],
	"created_at": "2024-01-15T10:30:00Z"
}

Content-Addressed Storage

All file content (PDFs, extracted text) is stored in iroh-blobs using content-addressing:

• Files are identified by their BLAKE3 hash
• Duplicate files are automatically deduplicated
• Content can be verified for integrity

Embedding Sync

Embeddings are stored in iroh-docs under embeddings/{doc_id}/{model_id}. This design:

• Avoids redundant computation — generating embeddings is expensive, so peers share them
• Preserves model flexibility — different peers can use different embedding models
• Enables offline use — embeddings sync with documents, ready for immediate use

When a peer receives a document, it checks for existing embeddings matching its configured model. If found, they’re used directly. If not (different model or new document), embeddings are generated locally and stored for other peers to use.

Data Flow

Local Import

1. User adds PDF to collection
2. Extract text via lopdf
3. Store PDF blob → get pdf_hash
4. Store text blob → get text_hash
5. Create metadata entry in iroh-docs
6. Index text + generate embeddings in milli

On Sync

When a new metadata entry arrives from a peer:

1. Fetch text blob using text_hash
2. Index text + generate embeddings in milli
3. PDF blob fetched on-demand (when user opens document)

What Syncs vs What’s Local

Local Storage

~/.local/share/insight/
├── iroh/               # iroh data (blobs, docs)
└── search/             # milli index

~/.cache/huggingface/hub/
└── models--*/          # Downloaded models (LLM + embedding)

On Windows, app data is under %LOCALAPPDATA%\insight\ and models under %USERPROFILE%\.cache\huggingface\hub\.