Local Semantic Search Engine (Local-First Notes)

A fully local, privacy-focused semantic search engine for personal notes.

The goal is simple: search by intent, not by keywords, while keeping all data and inference on-device.

---

Current Architecture (Semantic-First)

This version is pure semantic retrieval (ANN over embeddings).
There is no lexical retrieval stage (no FTS5/BM25) in the current pipeline.

Components

• DuckDB as the embedded database
• VSS (Vector Similarity Search) extension for vector indexes
• HNSW (ANN) index for fast nearest-neighbor retrieval
• Embedding model → converts text into vectors (e.g. FLOAT[1024])
• Reranker model (optional but recommended) → re-scores query–chunk pairs for higher precision
• (Optional) Tagger / Topics → enriches chunks with tags/topics for better UX and filtering

---

Models

This repo assumes local models stored under something like:

core-rs/models/
  bge-m3.gguf
  bge-reranker-v2-m3-Q8_0.gguf
  paraphrase-multilingual-mpnet-base-v2-onnx/

1) Embeddings (Semantic Retrieval)

Default (GGUF via llama.cpp)

• Model: bge-m3.gguf
• Output: 1024-dim embedding (stored as FLOAT[1024])

How it’s used:

• At import time: each chunk → embedding
• At search time: query string → embedding
• DuckDB VSS + HNSW retrieves Top-K nearest embeddings

2) Reranking (Precision Boost)

Default (GGUF via llama.cpp)

• Model: bge-reranker-v2-m3-Q8_0.gguf
• Output: a score for each (query, chunk_text) pair

How it’s used:

• After ANN retrieval (Top-K candidates), reranker re-scores and reorders them.
• This usually improves quality on ambiguous queries or short queries.

3) ONNX (Optional Runtime)

• Folder: paraphrase-multilingual-mpnet-base-v2-onnx/
• Intended use: run a model via ONNX Runtime inside the application runtime (no external binary).
• Status: optional / future-facing (helps remove dependency on llama.cpp for some pipelines).

---

Database

DuckDB database file (example): data/remin.duckdb

Tables (high-level)

• notes — note metadata (title, timestamps, etc.)
• chunks — chunked note text
• chunk_embeddings — embedding vectors per chunk (FLOAT[1024])
• chunk_tags / chunk_topics — optional enrichment layer (if enabled)

Vector Index

VSS creates an HNSW index on chunk_embeddings.embedding so search does not full-scan all vectors.

---

Setup

Requirements

• Rust (stable)
• DuckDB (embedded, shipped via dependency)
• DuckDB VSS extension available/enabled at runtime
• Local models (GGUF and/or ONNX)

Configuration

Suggested environment variables:

export REMIN_DB_PATH="data/remin.duckdb"
export REMIN_MODELS_DIR="core-rs/models"
export REMIN_EMBED_MODEL="bge-m3.gguf"
export REMIN_RERANK_MODEL="bge-reranker-v2-m3-Q8_0.gguf"   # optional
# export REMIN_ONNX_DIR="core-rs/models/paraphrase-multilingual-mpnet-base-v2-onnx"  # optional

---

Initialize the Database

Create the DB file, schema and vector index:

cargo run -- init-db

This should:

• Create DuckDB database file (if missing)
• Enable VSS extension
• Create tables
• Create HNSW vector index

---

Import Notes

Import a JSONL dataset:

cargo run -- import notes.jsonl

Import pipeline (semantic-first):

1. Parse notes from JSONL
2. Split each note into chunks
3. Generate an embedding per chunk (BGE-M3 → FLOAT[1024])
4. Insert notes, chunks, chunk_embeddings
5. (Optional) generate tags/topics and store them

---

Search

Run semantic search:

cargo run -- search "Andalusian cold soup"

Search pipeline:

1. Generate query embedding (BGE-M3)
2. ANN search via HNSW (DuckDB VSS) → Top-K candidates
3. (Optional) rerank candidates (BGE reranker) for precision
4. Return final ranked results (with note + chunk context)

---

Roadmap

• ONNX Runtime integration to reduce/replace llama.cpp where appropriate
• Mobile packaging (local-first)
• Multimodal notes (OCR + image embeddings)
• Better tagging/topic extraction and UI-level filtering

---

Philosophy

• Local-first
• Privacy by default
• Intent over keywords
• Engineering-driven AI systems

---

License

This project is licensed under the GNU General Public License v3.0 (GPLv3).

Why GPLv3?

Remin is built around the idea of local-first, privacy-focused AI systems. Choosing GPLv3 ensures that:

• Any distributed derivative work must remain open-source.
• Improvements to the core engine benefit the community.
• The project cannot be closed off into proprietary forks.

If you plan to use this project in a commercial or closed-source context, please review the implications of GPLv3 carefully.

For full license terms, see the LICENSE file or visit: https://www.gnu.org/licenses/gpl-3.0.html