Samyama-ai

We build Samyama Graph — an open-source graph-vector database in Rust. OpenCypher, RESP, HTTP, vector search, graph algorithms (read the paper, arXiv:2603.08036). Alongside it we run a daily DBMS systems-research program — small, reproducible, pre-registered papers on the database problems we hit while building it.

What is measured, and re-measured on a schedule (v1.8.0, 2026-09):

Every number above comes from a suite that writes a result envelope, and a regression gate blocks the release tag when a verdict is red or stale. Where something is unmeasured, the scorecard says so rather than leaving a gap.

We loaded the entire PubMed corpus (every article since 1966), ClinicalTrials.gov, Reactome pathways, and DrugBank into one graph. Then asked:

"What drugs are most tested in cancer clinical trials?"

Answer: Pembrolizumab (137 trials), followed by Carboplatin (106) and Paclitaxel (106). 5.2 seconds. One query spanning four databases.

96 of 100 benchmark queries pass →

Eleven arXiv preprints — the industrial asset-operations paper is accepted at the Agents+Graph workshop @ VLDB 2026 — reproducible, pre-registered, code + data with each. Honest baselines, not SOTA claims.

Also: dbms_research — a catalog of 1000+ open DBMS problems with formal statements, bounds, and 6,700+ link-checked references. Preprints in progress: clinical-llm-graphrag, bandwidth-bound-scan.

Most ship a prebuilt .sgsnap snapshot on the samyama-graph releases page — import one and start querying in under a second, no ETL.

Data licences are recorded, not assumed. Each of these repositories is Apache 2.0 for its loader; the data it reads carries its own terms. A DATA-LICENSES.md names every upstream source, what its own terms page says, and the date we read it — and states where that leaves the derived graph. Some of those answers are inconvenient and are written down anyway: the WHO-based graphs are non-commercial and share-alike, and druginteractions-kg combines share-alike, non-commercial and possibly non-redistributable sources, so we ship the loader rather than the graph. Where a source could not be verified, the row says so instead of guessing.

Newest — edge-ai-kg: the question it answers is "which operators in my model have no kernel on this accelerator, and silently fall back to the CPU?" Built on three real public sources — the ONNX operator catalog, ONNX Runtime kernel registrations across CPU/CUDA/DirectML, and 73 measured MLPerf Tiny v1.2 submissions — plus a generated hardware fleet for scale. Every node is stamped real or synthetic, so the two are never conflated.

git clone https://github.com/samyama-ai/samyama-graph && cd samyama-graph
cargo build --release
./target/release/samyama
# RESP on :6379, HTTP on :8080

Website: samyama.dev · Book: samyama-ai.github.io/samyama-graph-book · Contact: samyama.dev/contact