logosdb 0.7.6

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)

Contributing: See CONTRIBUTING.md for build instructions, style guide, and PR workflow.

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).
• Timestamp range filtering: search within a time window (e.g., "last 24 hours").
• Multiple distance metrics: inner product, cosine similarity (auto-normalized), or L2 Euclidean.
• 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.
• Framework integrations: LangChain and LlamaIndex VectorStore adapters.
• MCP server: first-class Claude Code integration via logosdb-mcp-server.

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.

Planning a deployment? See docs/sizing.md for disk/RAM estimates based on N×dim, or run python -m logosdb.sizing --rows 1_000_000 --dim 768.

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 and distance metric selection
  • logosdb::SearchHit result struct
  • logosdb_search_ts_range() for timestamp-filtered search

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.
• For inner-product distance (LOGOSDB_DIST_IP, the default), vectors must be L2-normalized before insertion. Use LOGOSDB_DIST_COSINE for automatic normalization.
• 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] = { /* ... unnormalized vector ... */ };

// L2-normalize for inner-product distance (returns 0 on success, -1 if zero norm)
if (logosdb_l2_normalize(vec, 2048) == 0) {
    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);

Search with timestamp range filter

#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);

// Search for top-5 matches within the last 24 hours
logosdb_search_result_t *res = logosdb_search_ts_range(
    db, query_vec, 2048, 5,
    "2025-04-21T10:00:00Z",  // from (inclusive), NULL for no lower bound
    "2025-04-22T10:00:00Z",  // to (inclusive), NULL for no upper bound
    50,                      // candidate_k: internal fetch multiplier (10x top_k recommended)
    &err);

for (int i = 0; i < logosdb_result_count(res); i++) {
    printf("#%d score=%.4f ts=%s text=%s\n", i,
           logosdb_result_score(res, i),
           logosdb_result_timestamp(res, i),
           logosdb_result_text(res, i));
}
logosdb_result_free(res);
logosdb_close(db);

Using different distance metrics

#include <logosdb/logosdb.h>

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

// Use cosine similarity (automatically normalizes vectors)
logosdb_options_set_distance(opts, LOGOSDB_DIST_COSINE);

// Or use L2 Euclidean distance
// logosdb_options_set_distance(opts, LOGOSDB_DIST_L2);

// Default is LOGOSDB_DIST_IP (inner product on L2-normalized vectors)

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

// For cosine: vectors are automatically normalized on put/search
float vec[2048] = { /* ... unnormalized vector ... */ };
logosdb_put(db, vec, 2048, "entry", "2025-04-22T10:00:00Z", &err);

logosdb_close(db);

Usage (C++ wrapper)

#include <logosdb/logosdb.h>
#include <vector>

// Basic usage with default inner-product distance
logosdb::DB db("/tmp/mydb", {.dim = 2048});

// L2-normalize your vectors before insertion (required for inner-product distance)
std::vector<float> embedding = load_some_vector();  // unnormalized
if (logosdb::l2_normalize(embedding)) {
    db.put(embedding, "My commute is 42 minutes", "2025-06-25T10:00:00Z");
}

// Or use l2_normalized() to get a normalized copy
auto normalized = logosdb::l2_normalized(query);
auto results = db.search(normalized, 5);
for (auto &r : results) {
    printf("id=%llu score=%.4f text=%s\n", r.id, r.score, r.text.c_str());
}

C++: Search with timestamp filter

#include <logosdb/logosdb.h>

logosdb::DB db("/tmp/mydb", {.dim = 2048});

// Search within a time window
auto results = db.search_ts_range(
    query, 5,
    "2025-04-21T00:00:00Z",  // from timestamp
    "2025-04-22T00:00:00Z",  // to timestamp
    50);                     // candidate_k (optional, defaults to 10x top_k)

for (auto &r : results) {
    printf("id=%llu score=%.4f ts=%s\n", r.id, r.score, r.timestamp.c_str());
}

C++: Using cosine or L2 distance

#include <logosdb/logosdb.h>

// Cosine similarity - vectors are automatically normalized
logosdb::DB db("/tmp/mydb", {.dim = 2048, .distance = LOGOSDB_DIST_COSINE});

// Put unnormalized vectors - they will be normalized automatically
db.put(unnormalized_embedding, "entry", "2025-04-22T10:00:00Z");

auto results = db.search(query, 5);
// scores are cosine similarities in [0, 1]

// L2 Euclidean distance
// logosdb::DB db("/tmp/mydb", {.dim = 2048, .distance = LOGOSDB_DIST_L2});

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 logosdb
from sentence_transformers import SentenceTransformer

# Local Hugging Face embeddings model (runs on your machine)
model = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
dim = model.get_sentence_embedding_dimension()

# Use cosine distance so LogosDB auto-normalizes vectors
db = logosdb.DB("/tmp/agent_memory", dim=dim, distance=logosdb.DIST_COSINE)

# Three learnings captured by an AI agent
learnings = [
    ("Retrying API calls with exponential backoff reduced transient failures by 42%.", "2026-05-06T09:00:00Z"),
    ("Splitting long tasks into smaller batches improved throughput and lowered memory spikes.", "2026-05-06T09:05:00Z"),
    ("Adding idempotency keys prevented duplicate writes during network retries.", "2026-05-06T09:10:00Z"),
]

for text, ts in learnings:
    emb = model.encode(text).astype("float32")
    db.put(emb, text=text, timestamp=ts)

# Ask a natural-language question
question = "How can we avoid duplicate writes when retries happen?"
q_emb = model.encode(question).astype("float32")

hits = db.search(q_emb, top_k=3)
for h in hits:
    print(f"{h.score:.4f}  {h.text}")

Python: Using cosine distance (no manual normalization needed)

import numpy as np
import logosdb

# With cosine distance, vectors are automatically normalized
db = logosdb.DB("/tmp/mydb", dim=128, distance=logosdb.DIST_COSINE)

# No need to normalize - just put raw vectors
v = np.random.randn(128).astype(np.float32)
rid = db.put(v, text="unnormalized vector", timestamp="2025-04-22T10:00:00Z")

# Search also works with unnormalized queries
query = np.random.randn(128).astype(np.float32)
hits = db.search(query, top_k=5)

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

Memory-Efficient On-Prem RAG

LogosDB is designed for memory-efficient retrieval-augmented generation (RAG) that runs entirely on your hardware.

RAM Model

LogosDB uses mmap() for zero-copy access. Your RAM usage scales with query patterns, not dataset size:

Dataset	Dim	Disk	Typical Query RAM
100K	384	153 MB	<20 MB
1M	384	1.5 GB	<100 MB
10M	384	15 GB	<200 MB

The OS caches hot index pages; cold data stays on disk. No explicit loading/unloading needed.

Quick RAG Example

import numpy as np
import logosdb
from sentence_transformers import SentenceTransformer

# 1. Load model (runs locally)
model = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
dim = model.get_sentence_embedding_dimension()

# 2. Create DB with cosine distance (auto-normalizes)
db = logosdb.DB("/data/knowledge", dim=dim, distance=logosdb.DIST_COSINE)

# 3. Index documents
for text in documents:
    emb = model.encode(text)
    db.put(emb, text=text)  # Auto-normalized with cosine distance

# 4. Query (only touched pages load into RAM)
query_emb = model.encode("What is HNSW?")
for hit in db.search(query_emb, top_k=3):
    print(f"{hit.score:.4f}  {hit.text}")

See docs/rag-on-prem.md for complete guide including:

• Time-sharding for infinite retention
• External quantization patterns
• Architecture patterns for production

See docs/sizing.md for detailed disk/RAM formulas and the python -m logosdb.sizing calculator.

Run the memory-efficient RAG example:

pip install ".[examples]"
python examples/python/memory_efficient_rag.py

Python: LlamaIndex VectorStore

pip install 'logosdb[llama-index]'

from logosdb import LogosDBIndex
from llama_index.core import Document
from llama_index.core.schema import TextNode
from llama_index.core.vector_stores import VectorStoreQuery
import numpy as np

# Create the vector store
db = LogosDBIndex(uri="/tmp/mydb", dim=128)

# Add nodes with pre-computed embeddings
node = TextNode(
    text="My commute is 42 minutes",
    embedding=np.random.randn(128).astype(np.float32).tolist(),
    metadata={"timestamp": "2025-04-28T10:00:00Z"}
)
db.add([node])

# Query
query_emb = np.random.randn(128).astype(np.float32).tolist()
query = VectorStoreQuery(query_embedding=query_emb, similarity_top_k=5)
results = db.query(query)

for node, score in zip(results.nodes, results.similarities):
    print(f"Score: {score:.4f}, Text: {node.text}")

# Timestamp range filtering
results = db.query(query, ts_from="2025-04-01T00:00:00Z", ts_to="2025-04-30T23:59:59Z")

The LogosDBIndex class implements LlamaIndex's VectorStore interface, supporting:

• add(nodes) - Add nodes with embeddings
• delete(node_id) - Delete by node ID
• query(VectorStoreQuery) - Similarity search by vector
• count() / len(store) - Number of live documents
• Timestamp filtering via ts_from and ts_to kwargs

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

Node.js

MCP Server — Claude Code integration

logosdb-mcp-server is a Model Context Protocol server that exposes LogosDB to Claude Code (and any other MCP client) over stdio. It lets Claude index files, persist knowledge across sessions, and do semantic search without leaving the conversation.

Quick start

1. Add to .claude/mcp.json in your project (or to ~/.claude.json for global use):

{
  "mcpServers": {
    "logosdb": {
      "command": "npx",
      "args": ["-y", "logosdb-mcp-server"],
      "env": {
        "LOGOSDB_PATH": "./.logosdb"
      }
    }
  }
}

By default the MCP server uses local Transformers.js embeddings (no API keys). Add EMBEDDING_PROVIDER / keys only if you want cloud or Ollama — see mcp/README.md.

Google Antigravity: same stdio + npx setup; step-by-step is in mcp/README.md — Google Antigravity.

2. Start Claude Code — the server is launched automatically on first tool call.

3. Use it in conversation:

> Index the src/ directory into a "code" namespace
> Find where JWT tokens are validated
> Remember that we decided to use UUIDs for all primary keys

Environment variables

Variable	Default	Description
LOGOSDB_PATH	./.logosdb	Root directory for all namespace databases
EMBEDDING_PROVIDER	(local)	Omit for Transformers.js on-device; or ollama, openai, voyage
TRANSFORMERS_MODEL	Xenova/all-MiniLM-L6-v2	Local embedding model (bundled MCP path)
OLLAMA_*	—	See mcp/README.md when using Ollama
OPENAI_API_KEY	—	Required when EMBEDDING_PROVIDER=openai
VOYAGE_API_KEY	—	Required when EMBEDDING_PROVIDER=voyage
LOGOSDB_CHUNK_SIZE	800	Target characters per chunk for file indexing

Voyage AI (voyage-3, dim=1024) is Anthropic's recommended cloud embedding model:

"env": {
  "LOGOSDB_PATH": "./.logosdb",
  "EMBEDDING_PROVIDER": "voyage",
  "VOYAGE_API_KEY": "<your-voyage-api-key>"
}

Available tools

Tool	Description
logosdb_index	Embed and store a text snippet in a namespace
logosdb_index_file	Chunk, embed, and store an entire file
logosdb_search	Semantic search across a namespace
logosdb_list	List all namespaces
logosdb_info	Stats for a namespace (count, dimension, path)
logosdb_delete	Delete an entry by row ID

Installing locally (without npx)

npm install -g logosdb-mcp-server

Then replace the command/args in mcp.json with:

"command": "logosdb-mcp-server",
"args": []

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 0.5.0
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)
mcp/                          MCP server (logosdb-mcp-server npm package)
.claude/mcp.json              Example Claude Code MCP configuration
CHANGELOG                     Release history
LICENSE                       MIT license text

Contributing

We welcome contributions. See CONTRIBUTING.md for:

• Building from source
• Running tests and benchmarks
• Code style and PR workflow

Please review our Code of Conduct and Security Policy.

License

MIT — see LICENSE for the full text.