The sharpest eye for your local AI.

An embedded vector database that runs inside your app, needs no server,
and keeps your data safe through crashes. Rust core, Python first.

Documentation · Quickstart · Local RAG guide · Changelog · Discussions

---

SrvDB (spelled S-R-V-D-B) is an open-source embedded vector database for Python and Rust: a library for semantic search, RAG and AI memory that stores embeddings on local disk, with no server and no cloud.

Not SurrealDB. SrvDB is an independent project and is not related to SurrealDB, the multi-model database. Install SrvDB with pip install srvdb (PyPI) or cargo add srvdb (crates.io).

Why SrvDB

• 🦅 In-process, offline. A library, not a service: no server, no Docker, no network access. pip install srvdb and go.
• 🛡️ Crash-safe by design. After persist() returns, your data survives a crash or power loss, and the database reopens exactly as it was. Every write path has a crash test.
• 🎯 Honest results. True cosine scores in [-1, 1] in every mode. Bad input (wrong dimension, NaN, duplicate ids, invalid metadata) is rejected before anything is written.
• ⚖️ Your trade-off. Exact search, HNSW, IVF, or 4× / up to 32× smaller storage with SQ8 and PQ: one argument to switch.
• 🧵 Thread-safe. Share one database between threads; every call releases the GIL and searches run in parallel.
• 🦀 Rust inside. SIMD distance kernels, typed errors, no Python runtime dependencies. Usable from Rust too.

Use cases

• Local RAG over private documents with a local LLM
• AI agent memory that survives restarts: remember, recall, forget
• Semantic search inside desktop, CLI and notes apps
• Offline and air-gapped systems where data must stay on the machine
• Edge devices with 4× (SQ8) to 32× (PQ) smaller storage
• Code, image and multimodal search, deduplication, "more like this"
• RAG tests in CI without starting a database service

Details and when to choose something else: Use cases.

100 projects in 20 industries, five levels each, from beginner to frontier. Cybersecurity alone goes from a SOC Copilot to Phishing Hunter, the Wireshark-style PacketSense, a Threat Hunter Agent and an air-gapped OT Sentinel. Also AML, fiscal auditing, GRC, aerospace, mechanical and electrical engineering, quantum computing, energy and more: Industries. Live examples and cheat sheets are coming soon.

Install

pip install srvdb

Wheels for Python 3.9–3.14 on Linux, macOS and Windows. Building from source needs Rust 1.99+ (details).

60-second quickstart

import json
import srvdb

db = srvdb.Database("./my-db", dimension=384)    # any dimension 128–4096

db.add(
    ids=["intro", "pricing"],
    embeddings=[intro_vector, pricing_vector],   # lists of 384 floats
    metadatas=[json.dumps({"page": "intro"}), json.dumps({"page": "pricing"})],
)
db.persist()                                     # crash-safe from here on

for doc_id, score in db.search(question_vector, k=5):
    print(doc_id, round(score, 3), db.get(doc_id))

db.upsert(ids=["pricing"], embeddings=[new_vector], metadatas=['{"page": "pricing", "v": 2}'])
db.delete(["intro"])
db.persist()

Local RAG in a few lines

import json
import srvdb
from sentence_transformers import SentenceTransformer   # pip install sentence-transformers

model = SentenceTransformer("all-MiniLM-L6-v2")          # 384-dim, runs offline
db = srvdb.Database("./rag-db", dimension=384)

chunks = ["SrvDB stores vectors on local disk.", "persist() makes writes crash-safe."]
db.upsert(
    ids=[f"chunk-{i}" for i in range(len(chunks))],
    embeddings=model.encode(chunks).tolist(),
    metadatas=[json.dumps({"text": c}) for c in chunks],
)
db.persist()

question = model.encode("How do I make writes durable?").tolist()
context = [json.loads(db.get(i))["text"] for i, _ in db.search(question, k=2)]

Full walkthrough with chunking and a local LLM: Build a local RAG app.

Index modes

Mode	Search	Storage per vector	Use when
flat (default)	Exact	4 B × dim	You want exact results
hnsw	Graph, approximate	vectors + graph	Search must be faster on large data
ivf	Partitions + graph	vectors + index	Very large collections
sq8	Scan, approximate	1 B × dim (4×)	Storage matters
pq	Scan, approximate	dim ÷ 8 B (up to 32×)	Storage matters most
auto	Flat, then HNSW	vectors	You'd rather not choose

Approximate modes trade accuracy for speed or size, depending on your data: measure recall against flat before switching. How to choose.

The durability promise

When persist() returns, every earlier add, upsert and delete survives a process crash or power loss. After a crash the database reopens in exactly the state of the last persist(): later writes are absent, never half-present.

How it works: Durability · ADR 0003.

Performance

SrvDB publishes performance numbers only from a committed, reproducible benchmark (hardware, OS, dataset, commit and command listed). That suite is coming in 0.4; numbers from earlier versions were withdrawn because they measured code with bugs fixed in 0.3. Measure on your own machine:

pip install "srvdb[bench]"
python benchmarks/universal_benchmark.py

Use it from Rust

use srvdb::{types::DatabaseConfig, SrvDB, Vector, VectorEngine};

let mut db = SrvDB::new_with_config("./rust-db", DatabaseConfig::new(384)?)?;
db.add(&Vector::new(vec![0.1; 384]), r#"{"title": "first"}"#)?;
db.persist()?;
let hits = db.search(&Vector::new(vec![0.1; 384]), 5)?;

Rust guide

FAQ

What is SrvDB? An embedded vector database: a Python and Rust library that stores embeddings on local disk and finds the most similar ones to a query, inside your own process. Use it for local RAG, semantic search and AI agent memory.

Is SrvDB the same as SurrealDB? No. SrvDB is an embedded vector database library; SurrealDB is an unrelated multi-model database server. The names only look alike.

Does SrvDB need a server, Docker or internet access? No. It runs in-process, works fully offline and never opens a network connection.

How is SrvDB different from Chroma, FAISS or LanceDB? SrvDB focuses on a small, embeddable core with a documented crash-safety guarantee: after persist() returns, data survives a crash or power loss. FAISS is a similarity-search library that leaves metadata and transactional storage to you; Chroma and LanceDB are broader data platforms.

Which embedding models work? Any model with 128 to 4096 dimensions, for example sentence-transformers, OpenAI, Cohere or Nomic embeddings.

Roadmap

• 0.3 — Trustworthy storage ✅ crash safety, delete/upsert/compact, validation, typed errors, thread safety, Apache-2.0
• 0.4 — Retrieval: metadata filtering, hybrid search, LangChain and LlamaIndex integrations, reproducible benchmarks, research into training-free quantization

Full roadmap

Contributing

Bug reports, docs, tests, benchmarks from your hardware and code are all welcome. Start with CONTRIBUTING.md; the engineering rules every change follows are in AGENTS.md. Questions and ideas: GitHub Discussions. Please follow the Code of Conduct; report security issues privately as described in SECURITY.md.

License

SrvDB is licensed under the Apache License 2.0: free to use, modify and ship in open-source and commercial products, with an explicit patent grant. Releases up to and including 0.2.0 were published under AGPL-3.0. The SrvDB name and eagle logo are not covered by the licence (brand guide).

Commercial support and consulting: srinivasvarma764@gmail.com

Acknowledgments

Built on SimSIMD, Rayon, PyO3, redb and parking_lot.