Architecture

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

Core Stack

Component	Library	Purpose
App framework	Tauri 2.0	Desktop app (Rust backend, web frontend)
UI	Svelte 5	Frontend
Styling	Tailwind 4	Utility-first CSS
Chat inference	rig	LLM framework: 25+ provider families, agent runtime, tools, memory
P2P / Sync	iroh	Connections, NAT traversal, sync
Content storage	iroh-blobs	Content-addressed file storage
Metadata sync	iroh-docs	CRDT key-value store for metadata
Real-time	iroh-gossip	Pub/sub for live updates
Search	milli	Full-text keyword search (BM25)
PDF text	lopdf	Digital text extraction

Agent Architecture

User Query
    ↓
rig Agent (per request: preamble + tools + memory + max turns)
    ↓
Registry-built Model (any supported rig provider family)
    ↓
Tool Calling Loop (multi-turn)
    ↓
Synthesized Answer (with citations)

Inference is remote-only: the chat provider is configured in Settings (family + model + optional API key / base URL). Insight supports every chat-completion provider family offered by rig — OpenAI, Anthropic, Gemini, Groq, OpenRouter, and local servers like Ollama and llama.cpp — through a data-driven registry. Only the conversation—queries and retrieved excerpts—goes to the provider. Documents are stored, searched, and synced entirely on your machine.

The runtime (crates/insight-core/src/runtime.rs) builds a fresh rig agent per request: a stable preamble, the document tools, file-backed conversation memory, and a scope hook that injects the conversation’s active collections as context on every turn. generate_title and predict_next_message are bare-model completions over the same provider settings. Conversation history persists as rig messages plus a metadata sidecar under conversations/{id}/. Agent progress streams to the frontend as rig MultiTurnStreamItem JSON events.

The agent has four typed tools (search, read_chunk, list_documents, get_collection_terms) built on milli search and iroh storage. 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/meta     → document metadata (JSON)
├── files/abc123/text     → extracted text
├── files/abc123/source   → original PDF bytes
├── files/def456/meta     → document metadata (JSON)
├── files/def456/text     → extracted text
├── files/def456/source   → original PDF bytes
├── _hash_index/{hash}    → duplicate detection index
└── _collection           → collection settings

Document Metadata

{
	"id": "abc123",
	"name": "paper.pdf",
	"file_type": "application/pdf",
	"page_count": 42,
	"tags": ["research", "climate"],
	"created_at": "2024-01-15T10:30:00Z",
	"page_boundaries": [0, 1500, 3200]
}

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

Data Flow

Local Import

User adds PDF to collection
Store: original PDF bytes written to files/{id}/source (iroh)
Extract: per-page text extracted with lopdf (digital text only; page boundaries recorded), written to files/{id}/text alongside files/{id}/meta
Index: the watcher triggers the index worker, which chunks text (~1500 chars, ~200 overlap, page-mapped) and indexes chunks in milli for full-text search

No OCR: extraction uses lopdf only. Scanned PDFs (images of text) import fine but yield little or no extracted text and are not text-searchable.

On Sync

When document entries arrive from a peer, iroh-docs automatically syncs the entry content blobs. The SyncWatcher listens for files/*/text entries and triggers indexing:

Text is already available at files/{id}/text (synced by iroh)
Chunks are indexed in milli for search
Source file at files/{id}/source is available immediately

What Syncs vs What’s Local

Data	Syncs	Stored in
PDF files	Yes	iroh-blobs
Extracted text	Yes	iroh-blobs
File metadata	Yes	iroh-docs
Collection info	Yes	iroh-docs
Search index	No (built from synced text)	milli

Search is local keyword indexing over synced text: each peer builds its own milli index. Nothing beyond the raw document data is derived or replicated.

Local Storage

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

Chat provider configuration (settings v2: active provider plus per-family api_keys) lives in ~/.local/share/insight/settings.json. Conversations are stored under conversations/{id}/ as messages.json (rig message history) and meta.json (title, linked collections). On Windows, app data is under %LOCALAPPDATA%\insight\.

Keyboard shortcuts

Insight