[P1] Implement local embedding-provider support and encrypted semantic index #22

New issue

Closed

opened 2026-07-18 00:01:20 +00:00 by lost-rob0t · 0 comments

lost-rob0t commented

2026-07-18 00:01:20 +00:00

(Migrated from github.com)

Goal

Enable efficient semantic retrieval without sending captured data off-device.

Requirements

Local embedding-provider implementation, initially through Ollama or an in-process model.
Batch and cancellation support with bounded resource use.
Embed only redacted text and approved metadata.
Store vectors and document references using an encrypted or threat-model-approved index design.
Record model identity and embedding dimension for compatibility.
Support index rebuild and model migration from encrypted source records.

Acceptance criteria

Time and semantic search work over synthetic records without network access.
Index entries cannot expose raw OCR text or window titles through filesystem inspection.
Dimension/model mismatches are detected before corrupting the index.
Reindexing can be interrupted and resumed safely.

Depends on

## Goal Enable efficient semantic retrieval without sending captured data off-device. ## Requirements - Local embedding-provider implementation, initially through Ollama or an in-process model. - Batch and cancellation support with bounded resource use. - Embed only redacted text and approved metadata. - Store vectors and document references using an encrypted or threat-model-approved index design. - Record model identity and embedding dimension for compatibility. - Support index rebuild and model migration from encrypted source records. ## Acceptance criteria - Time and semantic search work over synthetic records without network access. - Index entries cannot expose raw OCR text or window titles through filesystem inspection. - Dimension/model mismatches are detected before corrupting the index. - Reindexing can be interrupted and resumed safely. ## Depends on - #11 - #21