logosdb 0.2.0

Fast semantic vector database (HNSW + mmap) with Python bindings

LogosDB is a fast semantic vector database written in C/C++ that provides approximate nearest-neighbor search over embedding vectors with associated text metadata.

Authors: Jose (@jose-compu)

Features

• Vectors and metadata are stored as flat binary files, memory-mapped for zero-copy reads.
• Approximate nearest-neighbor search via HNSW (hnswlib), O(log n) query time.
• Each vector row carries optional text and ISO 8601 timestamp metadata (JSONL sidecar).
• The basic operations are Put(embedding, text, timestamp) and Search(query, top_k).
• Bulk vector access for direct tensor construction (e.g. loading into GPU memory).
• Thread-safe writes via internal mutex; concurrent reads are lock-free.
• Crash recovery: HNSW index is automatically backfilled from the append-only vector store on open.
• Scales to millions of vectors.

Documentation

The public interface is in include/logosdb/logosdb.h. Callers should not include or rely on the details of any other header files in this package. Those internal APIs may be changed without warning.

Guide to header files:

• include/logosdb/logosdb.h: Main interface to the DB. Start here. Contains:
  • C API with opaque handles and errptr convention (RocksDB/LevelDB style)
  • C++ convenience wrapper (logosdb::DB) with RAII and exceptions
  • logosdb::Options for HNSW tuning parameters
  • logosdb::SearchHit result struct

Limitations

• This is not a general-purpose vector database. It is purpose-built for embedding-based memory retrieval in LLM inference (funes.cpp).
• Only a single process (possibly multi-threaded) can access a particular database at a time.
• There is no client-server support built into the library. An application that needs such support will have to wrap their own server around the library.
• Vectors must be L2-normalized before insertion (inner-product similarity is used).
• Embedding generation is external — the caller provides pre-computed float vectors.

Getting the Source

git clone --recurse-submodules <repository-url>
cd logosdb

Building

This project supports CMake out of the box.

Quick start:

mkdir -p build && cd build
cmake -DCMAKE_BUILD_TYPE=Release .. && cmake --build .

This builds:

Target	Description
logosdb	Static library (liblogosdb.a)
logosdb-cli	Command-line tool for put, search, info
logosdb-bench	Benchmark: HNSW vs brute-force, with ChromaDB comparison
logosdb-test	Unit tests

Usage (C API)

#include <logosdb/logosdb.h>

char *err = NULL;
logosdb_options_t *opts = logosdb_options_create();
logosdb_options_set_dim(opts, 2048);

logosdb_t *db = logosdb_open("/tmp/mydb", opts, &err);
logosdb_options_destroy(opts);

float vec[2048] = { /* ... */ };
logosdb_put(db, vec, 2048, "My commute is 42 minutes",
            "2025-06-25T10:00:00Z", &err);

logosdb_search_result_t *res = logosdb_search(db, query_vec, 2048, 5, &err);
for (int i = 0; i < logosdb_result_count(res); i++) {
    printf("#%d score=%.4f text=%s\n", i,
           logosdb_result_score(res, i),
           logosdb_result_text(res, i));
}
logosdb_result_free(res);
logosdb_close(db);

Usage (C++ wrapper)

#include <logosdb/logosdb.h>

logosdb::DB db("/tmp/mydb", {.dim = 2048});
db.put(embedding, "My commute is 42 minutes", "2025-06-25T10:00:00Z");

auto results = db.search(query, 5);
for (auto &r : results) {
    printf("id=%llu score=%.4f text=%s\n", r.id, r.score, r.text.c_str());
}

Python

LogosDB ships Python bindings built with pybind11 and scikit-build-core.

Install from PyPI (binary wheels provided for Linux x86_64/aarch64 and macOS x86_64/arm64 on CPython 3.9–3.13):

pip install logosdb

Or build from source in a clone:

pip install .

Usage:

import numpy as np
import logosdb

db = logosdb.DB("/tmp/mydb", dim=128)

v = np.random.randn(128).astype(np.float32)
v /= np.linalg.norm(v)

rid = db.put(v, text="hello", timestamp="2025-06-25T10:00:00Z")
# rid is the row id for this vector (often 0 for the first insert).
hits = db.search(v, top_k=5)
print(hits[0].text, hits[0].score)

# Replace an existing row: first arg is that row's id (must still be "live").
# update() tombstones the old row and appends a new one; it returns the NEW id.
v2 = np.random.randn(128).astype(np.float32)
v2 /= np.linalg.norm(v2)
new_id = db.update(rid, v2, text="replaced")

# Tombstone a row by id (excluded from search). count() includes deleted rows;
# count_live() does not.
db.delete(new_id)

# After a delete, that id cannot be updated again — insert a fresh row with put().
# zero-copy bulk view over mmap-backed storage (shape: (count, dim), read-only)
vectors = db.raw_vectors()

print(db.count(), db.count_live(), db.dim)

Run the Python tests and examples:

pip install ".[test]"
pytest tests/python/

python examples/python/basic_usage.py

# sentence-transformers demo (optional heavy dep)
pip install ".[examples]"
python examples/python/sentence_transformers_demo.py

CLI

# Database info
logosdb-cli info /tmp/mydb

# Search with a binary query vector file
logosdb-cli search /tmp/mydb --query-file q.bin --top-k 5

Performance

Here is a performance report from the included logosdb-bench program. The results are somewhat noisy, but should be enough to get a ballpark performance estimate.

Setup

We use databases with 1K, 10K, and 100K vectors. Each vector has 2048 dimensions (matching typical LLM embedding sizes). Vectors are L2-normalized random unit vectors.

LogosDB:    version 0.2.0
CPU:        Apple M-series (ARM64)
Dim:        2048
HNSW M:     16, ef_construction: 200, ef_search: 50

Write performance

put (1K vectors):    ~50 µs/op   (~20,000 inserts/sec)
put (10K vectors):   ~80 µs/op   (~12,500 inserts/sec)
put (100K vectors):  ~120 µs/op  (~8,300 inserts/sec)

Each "op" above corresponds to a write of a single vector + metadata + HNSW index update.

Search performance

HNSW top-5 (1K):     ~0.1 ms/query
HNSW top-5 (10K):    ~0.3 ms/query
HNSW top-5 (100K):   ~1.2 ms/query

Brute-force top-5 (1K):    ~0.3 ms/query
Brute-force top-5 (10K):   ~2.5 ms/query
Brute-force top-5 (100K):  ~25 ms/query

HNSW maintains sub-linear scaling while brute-force grows linearly with database size. At 100K vectors, HNSW is roughly 20x faster.

Benchmark vs ChromaDB

logosdb-bench --dim 2048 --counts 1000,10000,100000

Metric	ChromaDB	LogosDB
Language	Python + C (hnswlib)	Pure C/C++
Search algorithm	HNSW	HNSW (same hnswlib)
Storage	SQLite + Parquet	Binary mmap + JSONL
Startup overhead	Python runtime + deps	Zero (linked library)
Embedding generation	Built-in (Sentence Transformers)	External (caller provides vectors)
Target use case	General-purpose vector store	Embedded LLM inference memory
Search latency (100K, dim=2048)	~5-10 ms	~1-3 ms
Memory footprint (100K, dim=2048)	~1.5 GB (Python + SQLite)	~800 MB (mmap)
Cold start	~2-5 s (Python imports)	<10 ms
Dependencies	Python, NumPy, SQLite, hnswlib	hnswlib (header-only, vendored)

LogosDB uses the same HNSW implementation as ChromaDB (hnswlib) but eliminates Python overhead, SQLite serialization, and Sentence Transformer coupling. The result is a leaner library optimized for the single use case of embedded semantic memory for LLM inference.

Repository contents

include/logosdb/logosdb.h     Public C/C++ API (start here)
src/logosdb.cpp               Core engine: wires storage + index + metadata
src/storage.h / storage.cpp   Fixed-stride binary vector file with mmap
src/metadata.h / metadata.cpp Append-only JSONL text + timestamp store
src/hnsw_index.h / .cpp       Thin wrapper around hnswlib
tools/logosdb-cli.cpp         Command-line interface
tools/logosdb-bench.cpp       Benchmark tool
tests/test_basic.cpp          C++ unit tests
tests/python/test_smoke.py    Python smoke tests (pytest)
python/src/bindings.cpp       pybind11 Python bindings
python/logosdb/               Python package (logosdb._core + stubs)
examples/python/              Python usage examples
pyproject.toml                Python build/config (scikit-build-core)
third_party/hnswlib/          Vendored hnswlib (header-only)
CHANGELOG                     Release history
LICENSE                       MIT license text

License

MIT — see LICENSE for the full text.