One transaction boundary
SQL, structures, and search mutations commit under one catalog, WAL, MVCC root set, and commit sequence.
Software system 01 / Data
One local authority for data that must survive and explain itself.
Hyphae is a Rust data engine where SQL, native structures, lexical/vector search, transactions, recovery, and proofs share one owned storage and commit substrate.
Replace a stack of local databases and search services with one bounded, offline-capable authority whose state and releases can be verified.
01 / Benefits
Why teams use Hyphae
Local-first by construction
Embedded Rust, local sockets or named pipes, CLI, and optional loopback HTTP work without a required cloud account.
Search without a sidecar
Lexical, exact-vector, incremental ANN, filtered, faceted, metric, and hybrid search live in the same engine.
Durability is inspectable
Checkpoints, backups, restores, recovery, vacuum generations, doctor reports, proofs, and witnesses have bounded failure behavior.
Public contracts
OpenAPI, JSON Schema, typed Python and TypeScript SDKs, and versioned local protocols define the wire boundary.
Release evidence
Archives, checksums, SBOMs, signatures, provenance, and gate receipts bind releases to exact source.
02 / Architecture
How the system is arranged
All public surfaces converge on one Native product and transaction authority; storage, recovery, and evidence are not separate services.
- 01
Applications
Embedded Rust, CLI, Python/TypeScript SDKs, MCP, local protocol, optional HTTP /v2
- 02
Native product authority
Identity, access control, sessions, limits, cancellation, operations, and transaction scheduling
- 03
Three data surfaces
Relational SQL + native structures + lexical/vector/hybrid search
- 04
Durable substrate
Catalog, pages, blobs, WAL, MVCC, indexes, object IDs, checkpoints, backup, and proofs
- 05
Verifiable outputs
Witnesses, receipts, doctor reports, release provenance, and offline verification
03 / Surface
What Hyphae provides
Relational
Bounded DDL/DML, prepared queries, secondary indexes, joins, constraints, windows, and transactions.
Native structures
Strings, counters, hashes, lists, sets, sorted sets, streams, TTL, scans, algebra, and atomic batches.
Retrieval
BM25, exact vectors, HNSW ANN, filters, facets, metrics, deterministic fusion, and same-snapshot hybrid search.
Operations
Single-directory ownership, checkpoints, backup/restore, vacuum, import/promotion, access control, and diagnostics.
04 / Evidence
What has been demonstrated
The G7 operational-scale authority is strong evidence for its named VM and protocol, not a universal latency claim.
measured observations
33 surface/concurrency cells
logical commits recovered
zero missing or mismatched
ANN recall@10
in every measured cell
closed bounded gates
source-bound release evidence
These measurements establish bounded correctness, accounting, recall, concurrency, and recovery in the published environment. They do not certify portable latency, bare metal, or distributed operation.
What it does now
- Runs as one local Rust binary and one exclusively owned data directory.
- Unifies relational, structure, and retrieval workloads under one transaction authority.
- Supports embedded, local multi-client, CLI, and optional loopback HTTP use.
- Ships cross-platform archives and typed clients with release verification.
What it intends to do
- Reduce embedded-path contention and complete single-stage filtering.
- Add SIMD and optional heterogeneous acceleration only behind evidence gates.
- Advance attested embeddings and Proof of Retrieval.
- Treat hosted, distributed, and model-driven programs as separate contracts rather than silent scope growth.
What it deliberately is not
- Not a distributed cluster, hosted SaaS, billing/control plane, or multitenant database.
- No built-in TLS or at-rest encryption; operators own remote termination and filesystem policy.
- No embedded LLM or required embedding provider.
- Proofs verify documented commitments and rules; they do not prove that input facts are true.
Hyphae