Munind

Munind is a local-only, general-purpose vector database in Rust.

It stores fixed-dimension embeddings with JSON payloads and returns matching JSON documents for vector, hybrid, and filtered retrieval.

Munind is application-agnostic at the core. It is not tied to a single memory or RAG workflow.

Why Munind

• Local-only by design (desktop/mobile-oriented, no cloud scaling goal)
• Database-style workflow: insert embedding + JSON, search returns JSON
• Fixed embedding dimension set at database creation
• Hybrid retrieval support (vector + BM25F)
• Optional second-stage reranking
• Indexed JSON filter execution for common metadata fields

Why Munind (no, really)

• I built Munind because my AI agents needed a memory database, and tools like Milvus felt like overkill for a local-first setup

Current Capabilities

• Create/open DB with immutable embedding dimension
• CRUD over vectors with JSON payloads (insert, get, update, delete)
• Vector ANN search with optional ef_search and radius
• Hybrid search (text_query + vector fusion)
• Payload-filtered search (Eq, And) with indexed execution paths
• Storage optimize/compaction to checkpoint state and truncate WAL
• Ingestion/search pipeline crate with pluggable embedding and reranker providers
• Benchmarking for both speed and quality (recall@k, filtered recall, MRR, nDCG)

Example Use Cases

• Semantic search over app documents/messages
• Metadata-filtered vector retrieval for local products
• Personal memory and note retrieval
• RAG context retrieval for local assistants

Quick Start (CLI)

1) Build

cargo build --workspace

2) Create DB (512 dim)

cargo run -p munind-cli -- --db ./munind_data create --embedding-dim 512

3) Ingest a text file

cargo run -p munind-cli -- --db ./munind_data ingest --file ./test_doc.txt --doc-id demo-1

4) Search

cargo run -p munind-cli -- --db ./munind_data search "quick brown fox" -k 5

4b) CRUD via JSON (Developer-Friendly)

Insert:

cargo run -p munind-cli -- --db ./munind_data \
  insert \
  --embedding-json '[0.1,0.2,0.3]' \
  --document-json '{"doc_id":"x1","text":"hello"}'

Get:

cargo run -p munind-cli -- --db ./munind_data get --id 1 --include-embedding

Update:

cargo run -p munind-cli -- --db ./munind_data \
  update --id 1 \
  --embedding-json '[0.2,0.1,0.0]' \
  --document-json '{"doc_id":"x1","text":"updated"}'

Delete:

cargo run -p munind-cli -- --db ./munind_data delete --id 1

5) Health check

cargo run -p munind-cli -- --db ./munind_data check-health

6) Optimize (Compaction / WAL Truncation)

cargo run -p munind-cli -- --db ./munind_data optimize

Benchmark Quick Start (Real Embeddings)

Defaults are already in scripts:

• endpoint: http://localhost:8082/v1/embeddings
• model: nomic-embed-text-v1.5

Set EMBEDDING_API_KEY only if your endpoint requires auth.

Run the 3 phases:

bash benchmark/01_build_dataset.sh
bash benchmark/02_create_database.sh
bash benchmark/03_run_benchmark.sh

Phase 2 runs post-prepare compaction by default (checkpoint + WAL truncation).
Set COMPACT_AFTER_PREPARE=0 to skip it.

03_run_benchmark.sh is latency-only by default. For exact quality metrics too:

BENCH_WITH_QUALITY=1 bash benchmark/03_run_benchmark.sh

Quality mode uses vectors already stored in the DB for exact baseline (it does not re-embed all docs from input).

Or one command:

bash benchmark/run_benchmark.sh

GloVe ANN Benchmark (Full Dataset)

Build full DB once:

TRAIN_LIMIT=0 EXPORT_RAW=0 CHECKPOINT_WAL_AFTER_PREPARE=1 COMPACT_AFTER_PREPARE=0 bash benchmark/02_create_glove_database.sh

Run ANN search benchmark on existing full DB:

EF_SEARCH=400 QUERIES=10000 bash benchmark/03_run_glove_benchmark.sh

Latest snapshot (2026-03-08, cosine ANN, top_k=10, train_vectors=1183514):

• Search: 238.38 ops/s, p50 4.255 ms, p95 5.624 ms, p99 6.406 ms, mean 4.195 ms
• Recall@10: mean 0.9100, p50 1.0000, p95 1.0000

Documentation

• Documentation Index
• Architecture
• Search and Ranking
• CLI Guide
• Language Integration
• Rust API Guide
• Benchmarking
• Operations

Repository Layout

• crates/munind-core
• crates/munind-storage
• crates/munind-index
• crates/munind-api
• crates/munind-rag
• crates/munind-cli
• crates/munind-bench
• benchmark/
• docs/

License

MIT. See LICENSE.