Gloss

Overview

Gloss is the public knowledge-worker surface of the evidence-first stack: a local-first notebook/RAG desktop app for grounded chat over personal documents. It is built for privacy-sensitive research and personal knowledge work where source locality, resource control, and answer provenance matter.

Architecture

Gloss uses a Tauri 2 desktop shell with a Rust backend and React frontend. The current source basis points to per-notebook SQLite storage, local embeddings, Ollama integration, and local retrieval with source-linked answers.

Key implementation themes:

• notebook-level isolation
• SQLite-backed local persistence
• local LLM and embedding paths through Ollama and local model tooling
• hybrid retrieval with source evidence surfaced to the UI
• bounded local resource use so background work does not starve interactive chat

Current Focus

The current active Gloss pass is chat-runtime proof. The audit basis from May 18, 2026 focuses on making desktop chat always produce visible tokens, a visible error, or a durable ChatAttemptTraceV1 for every prompt.

The suspected failure surfaces are frontend event filtering by notebook ID, split provider URL state, missing durable attempt traces, preview retrieval blocking provider streaming, Ollama stream errors not being surfaced, and missing live desktop smoke proof.

Relationship To The Stack

Gloss is the local document-facing counterpart to Recall and semantic-memory. ClaimLedger is being shaped as the backend compiler/certifier for Gloss Claim Hygiene and Evidence Ledger features, while Gloss remains the UI and orchestrator.

Next Gates

• Prove prompt-to-token desktop chat behavior with live smoke evidence.
• Add durable chat attempt traces.
• Keep semantic-memory and TurboQuant promotion behind explicit readiness gates.
• Make claim/evidence bundles reviewable through the Gloss UI without moving claim authority into Gloss itself.