Name: SwarnDB
Author: SarthiAI

Question 1

What is SwarnDB?

Accepted Answer

SwarnDB is a production-grade vector database written in Rust where a vector and a graph node are the same object. The id of a vector is the id of its graph node, so there is only one thing to store and reason about. By default it is a fast, accurate vector store, and when you need it, it becomes a real typed graph. It is one engine, not two systems bolted together.

Question 2

How is SwarnDB different from a vector database plus a separate graph database?

Accepted Answer

You get one identity (the vector id is the node id), one store, and one crash-recovery path. There is no foreign key to maintain, no mirror table, no sync job, and no copy of the data that can drift out of step. When you run a vector store and a graph store side by side, you spend real effort keeping them in agreement; here that problem does not exist because there is a single source of truth.

Question 3

Is the graph automatic, or do I create the edges?

Accepted Answer

You author the edges yourself, either one at a time with put_edge or in bulk through CSV and JSONL import. You can also have an LLM extract edges for you, with provenance recorded for each one. This is not a graph inferred from similarity. An optional auto-similarity mode (mode="auto_similarity") does exist, but it is a secondary feature and it is off by default.

Question 4

Does the graph make search faster or more accurate?

Accepted Answer

To be honest, not universally. On broad question-answering, the graph does not beat plain vector search, so we will not claim it does. The graph is the right tool where structure genuinely matters: knowledge graphs, agent memory, recommendations, and fraud and threat graphs. Use it when relationships carry meaning, not as a blanket speed-up.

Question 5

Do I have to use the graph?

Accepted Answer

No. Collections are vector-only by default, and you opt into the graph per collection by setting mode="hybrid". If you never need relationships, you can use SwarnDB purely as a vector store and ignore the graph entirely.

Question 6

How does a hybrid query work?

Accepted Answer

A hybrid query is one composable plan that scopes by structure, then ranks by meaning. You seed candidates by vector similarity, walk typed edges across the graph (single-hop, k-hop, or shortest path), and then rank the surviving frontier exactly by meaning with vector_rank. It all runs in one engine over one copy of the data, so there is no round trip to a separate graph store.

Question 7

What does it cost, and what is the license?

Accepted Answer

SwarnDB is released under the Elastic License 2.0 and is source-available. You can read and use the source under the terms of that license.

Question 8

How do I install and run it?

Accepted Answer

Install the client with pip install swarndb, or pip install swarndb[async] if you want the async extra. To run the server in a container, use docker run sarthiai/swarndb. The service listens on gRPC port 50051 and REST port 8080.

Question 9

How does SwarnDB compare to Pinecone, Weaviate, Qdrant, or Neo4j?

Accepted Answer

To be honest, the differentiator is the integrated typed-graph-plus-vector hybrid in one engine with one identity, where the vector id is the node id. It is not about managed-convenience polish. If your priority is a fully hosted, hands-off experience, the established managed services have more of that today; if you want one engine that is both a vector store and a typed graph over the same objects, that is what SwarnDB offers.

Question 10

What are the benchmark numbers?

Accepted Answer

These figures are vector-only and dated 2026-05-20, and they were not re-run for v1.1.0. On DBpedia 1M (1536-dim float32, cosine) we measured peak throughput of about 2,398 QPS and recall@10 of 0.9974 at ef=800. Latency at that setting was p50 3.16 ms and p95 4.91 ms.

Question 11

Is it production-ready?

Accepted Answer

It is Rust-native with WAL crash recovery, and SQ8 scalar quantization cuts index memory by about 75% while keeping an exact final ranking and a fast restart. We will be honest about a documented limitation: during a large file-based bulk load, health probes can be unresponsive, which is tracked as known issue F3. Plan your load windows accordingly.

Question 12

What is the default retrieval mode?

Accepted Answer

For vector collections the default is vector_only, and for graph-augmented (hybrid) queries the default is vector_rank. RRF (hybrid_rrf) and graph_only also exist, but they are not recommended because RRF regressed in our validation. Stick with the defaults unless you have a specific reason to change them.

Questions, answered honestly.