LocalFind

Windows-first, privacy-preserving desktop search: find local files and precise visual moments using everyday language, without uploading personal content.

Status: Phase 0 feasibility — not implemented. There is no working app in this repository yet. The planning documents below are specifications, not proof that the proposed stack performs adequately.

Consented local search

• You explicitly choose which folders are indexed; nothing is crawled silently (no whole-drive, Documents/Downloads, or network-share scans by default).
• Original files are read-only: indexing never modifies, moves, or deletes them. Index data, thumbnails, and caches live in separate app-owned directories and index cleanup never touches originals.
• Search works offline after the model is locally installed. Exact filename/path search must keep working even when the AI model is unavailable.
• Default logs and telemetry contain counts and error codes only — never file contents, queries, transcripts, vectors, or unredacted paths.

Proposed stack (provisional, pending Phase 0 evidence)

Area	Proposal	Reference
Desktop UI + host	Tauri 2 + Svelte/TypeScript, Rust core	PLAN.md ADR-001/ADR-002
Metadata + lexical index	SQLite (WAL) + FTS5	PLAN.md ADR-003
Vector baseline	float32 BLOBs + exact cosine behind a VectorIndex port	PLAN.md ADR-004
Media probing	Bundled FFmpeg/ffprobe candidate	PLAN.md ADR-005
File watching	Rust notify adapter + periodic reconciliation	PLAN.md ADR-006
Inference	Separate single Python worker over stdio JSON protocol	PLAN.md ADR-007
Model (candidate)	google/embeddinggemma-2 (768-dim native, bf16/fp32 only, never fp16)	PLAN.md §3
Distribution	User-initiated model download, no background telemetry	PLAN.md ADR-008

All ADRs are proposed until validated by Phase 0 measurements.

Documented unknowns (verified in Phase 0, not assumed)

• Compatible Python/PyTorch/sentence-transformers versions on Windows 11, and CPU vs GPU inference cost (PLAN.md Q-001, Q-002).
• Whether the Python sidecar can be packaged and started on a clean Windows VM, and its install footprint (Q-003).
• Truncation dimensions, segmentation policy, and vector-backend choice backed by retrieval-quality and latency benchmarks (Q-004, Q-005, Q-006).
• Model, library, FFmpeg, and codec redistribution/license obligations (Q-007).
• Phase 0 exit gate: Recall@5 >= 0.75 on a small curated suite of ≥30 visual queries, or an explicit PIVOT/STOP decision (PLAN.md §2).

Documents (authoritative)

• PLAN.md — product scope, provisional architecture, ADRs, risks.
• STEPS.md — task roadmap; P0-T006 is the current task.
• RULES.md — enforceable constraints (user-data safety first).
• AGENTS.md — one-task-at-a-time agent workflow.

If instructions conflict: user-data safety and explicit user instruction, then RULES.md, then accepted ADRs in PLAN.md, then the current task in STEPS.md.

Getting started

Nothing to build or run yet. Phase 0 feasibility prototypes (under benchmarks/, explicitly not production code) begin at P0-T002 once this foundation task (P0-T001) is accepted. See STEPS.md Project State for the current task — do not jump ahead to later phases.

git status --short

Contributing

• Work one STEPS.md task at a time; keep diffs focused (RULES.md R-25).
• Test media must be synthetic, consented, or redistributable — never commit real user files, private embeddings, model weights, secrets, or machine paths (RULES.md R-24, R-26).
• Report exact test commands and results; never claim an unexecuted check passed (RULES.md R-23, R-29).