Release notes
1.0.0
What is new
-
A node and its edges in one graph transaction. An edge can now point at a node created earlier in the same transaction, by that operation’s index, so a new node and its edges commit together. See Java SDK and Python SDK.
-
Edges to nodes by external key. An edge endpoint in a graph transaction can also name a node by its external key: Java
EdgeEndpoint.key(…), PythonEndpointKey,source_key/target_keyon the wire,sourceKey/targetKeyover REST and MCP. A caller that knows its nodes by its own keys connects them in one request without reading their ids first. A key no node holds fails the whole transaction as not found. Since resolving a key reads by key, such an edge needsGRAPHS_READas well asGRAPHS_CREATE. -
Nodes by external key, in one call.
GetNodesByKeysreads a named set of nodes by their external keys, asGetNodesdoes by id: a key no node holds is left out, duplicates collapse, and each node carries its key so the result maps back to the request. JavagetNodesByKeys, Pythonget_nodes_by_keys, RESTPOST /api/v1/graphs/{graphId}/nodes-by-key. It replaces oneGetNodeByKeyround trip per key.
Changes
-
The index of a long string no longer grows with its length. A
HIGH-cardinalitySTRINGmetadata field is indexed by a hash of its value instead of the value’s bits, so a 100 KB value costs the same index entries as a short one, where it used to cost about 40 000. Exact matches (=,IN,!=,NOT IN,IS [NOT] NULL) use the index; ranges andLIKE/CONTAINSon such a field were already evaluated on the stored values. Existing stores keep their current index on existing fields. The storage format is now version 15: a version 14 store still opens and is re-stamped, and an older build can no longer read it afterwards. See the metadata reference for the details.
Changes to the API
-
A graph transaction’s add-edge takes endpoints. Each endpoint is an existing node or an earlier operation of the batch: Java
AddEdge(EdgeEndpoint source, EdgeEndpoint target, …)withEdgeEndpoint.node(id)/EdgeEndpoint.operation(index)(theAddEdge(long, long, …)constructor remains for two existing nodes, but thesourceId()/targetId()accessors are nowsource()/target()); PythonAddEdgeOp(source, target, …), where a plainintstill names a node id but the fields are renamed fromsource_id/target_id. On the wireAddEdgeOp’s endpoints are `oneof`s, adding `source_op_index/target_op_index; over REST and in thecyrock_db_graph_transactionMCP tool,sourceOpIndex/targetOpIndex. -
An add-edge in a transaction without an endpoint now answers "exactly one of sourceId, sourceOpIndex or sourceKey is required" (likewise for the target) instead of "sourceId is required".
Performance
-
Graph reads no longer pay for the heap check. Reading edges with their metadata checks heap pressure once per batch, and that check asked every garbage collector for its last collection on each call, about 20 µs. Cellular keeps the value from the collectors' notifications instead, about 5 ns a call. Under parallel load the check had taken most of the graph-read threads' time and halved the throughput of traversal-heavy queries.
Upgrading
-
The packages moved to the releases repository. The Java SDK, its BOM and the Python SDK are published from cyrock-ai/releases. Point the Maven repository at
maven.pkg.github.com/cyrock-ai/releases, and give thesettings.xmlserver the idcyrock-packages(it wascyrock-early-access). The Python SDK installs from the tagdb-v1.0.0in that repository. Issues and discussions live there too. See Java SDK and Python SDK. -
Licence. Use of CYROCK.AI DB is governed by the CYROCK.AI DB End User License Agreement, which distinguishes an embedded, a server and a cluster variant.
-
Start with an empty storage directory. The storage engine, Cellular, writes entities in a compact format (smaller entity headers and 4 byte references) and cannot read the earlier one. The storage format is now version 15, and a store of version 13 or earlier is refused at startup with the banner described under 0.9.1. Take a backup, then start fresh - Operations covers volumes and backup.
0.9.1
Still one container image. What moved is reach - a Python SDK, connectors for the two JVM AI frameworks, an asynchronous Java API - and correctness, in a long run of changes to what the APIs answer when a request is wrong.
What is new
Clients
-
A Python SDK at parity with the Java client: collections, documents, graphs, CyQL, transactions and the community layer, as a blocking client and an asyncio one. See Python SDK.
-
Framework connectors for the JVM AI frameworks: a Spring AI
VectorStorewith Spring Boot auto-configuration and a health indicator, and a LangChain4jEmbeddingStore, plus retrievers that reach the graph traversal and community summaries a plain vector store cannot express. See Framework connectors. -
An asynchronous Java API.
client.async()returns the same operations returningCompletableFuture, sharing the synchronous client’s connection, key exchange and lifecycle, so a caller can hold many calls in flight without a thread waiting on each. See Java SDK. -
Renaming a collection is now available in the SDK, alongside the console and REST.
Data model
-
A collection or a graph need not declare a vector field. Without one you have a document or node store with typed metadata filtering and full-text search, and only similarity search is unavailable. A vector field added later takes effect immediately, so vectors can arrive after the data does.
New reads
-
GetNodesreads a named set of nodes in one call instead of one at a time. -
An induced subgraph read: given a set of nodes, get the edges among them without paying for the frontier around them.
-
Bounded community reads. Reading a community’s members, the node-to-community assignments and the stored summaries are all windowed now, and each says whether more was held back, so reading a large graph’s community layer no longer means holding all of it at once.
-
Updating an edge’s metadata in place, rather than by replacing the edge.
Agents
-
One new MCP tool,
cyrock_db_graph_get_community_summaries, which reads the stored summaries at a hierarchy level by name rather than by similarity to a query - the right tool when you already know which communities you want. The catalogue is now 40 tools. See MCP for AI agents.
Console
-
The CyQL editor has syntax highlighting, completion and live validation. Completion offers your own fields alongside the language’s keywords and functions, and syntax is checked as you type rather than on submit.
-
A context-sensitive help panel that describes the screen you are on, with tooltips on the controls that earn them.
-
The console’s strings are externalized for translation, which is the groundwork for shipping it in more than one language.
-
A pass over the whole console so it reads as one designed product: consistent spacing, headings, empty states and actions across every tab.
-
The console says which organization and project you are working in, which starts to matter the moment you have more than one.
Platform
-
The storage format is versioned. Storage this build cannot read is detected when it is opened at startup and the engine refuses to serve it, naming both format versions and telling you to start with an empty storage directory, rather than failing later inside a request. See Operations.
-
The tenancy model is enforced across both planes. Organization and project scoping now applies on the platform and data planes alike, so a request cannot reach across a boundary the console does not show you.
Correctness
-
A large pass over the API contract, around thirty changes, in three themes. A caller error now answers with an honest status code instead of a 500: a wrong vector dimension, a degenerate or empty vector, a blank or omitted required field and a number written into a
STRINGfield are all 400s, and a missing graph node or edge is a 404. Input is validated at the boundary rather than deep inside a write. And result sets are bounded, so a list or a search cannot be asked for an unbounded window, and a zero or negative count means an empty result rather than everything. -
Graph branching fixes. An upsert on a branch no longer loses data or leaks into the graph it was forked from, a vector field removed on a branch is removed at merge, and
UNIQUEand eventual indexing survive in a branch overlay instead of being silently dropped.
Upgrading from 0.9.0
-
Start with an empty storage directory. 0.9.1 will not read 0.9.0’s storage: the format is versioned now, and a 0.9.0 directory carries no marker, so it is refused at startup with the banner described above. Take a backup, then start fresh - Operations covers volumes and backup.
-
Some failures now answer with a different status. Requests that used to come back as 500 come back as 400 or 404 where that is the honest answer, and blank or missing required values are refused at the client boundary rather than sent. Code that branched on a 500, or that relied on a blank id being accepted, is worth a look. Branch on the SDK’s exception subtypes rather than on the number - see Java SDK.
-
Unbounded result windows are refused. A list or a search that asked for everything now has to name a window, and a count of zero or less means an empty result.
Known limitations
These are current boundaries rather than defects. Several are simply the next things we are building.
-
One container only. The multi-service topology - separate platform and data servers behind a single-port gateway, and horizontally scaled data servers - is not part of this release.
-
Evaluation authentication. The image seeds three in-memory logins at fixed roles. Identity provider integration over OIDC belongs to the deployment topology arriving in a future release.
-
Credentials come from the startup banner. There is no flow yet for rotating the seeded API key from the console; create an additional key in the administration area if you need a second one.
-
A vector field’s dimension is fixed once the field exists. A collection or graph can now be created without a vector field at all, and a new field can be added later and takes effect immediately - but an existing field’s dimension cannot change, so moving to an embedding model of a different size still means a new field and a re-embed rather than a migration in place.
-
Single writer per collection. Consistency for HNSW indices depends on one writer at a time. This is invisible in a single container and becomes relevant only in the scaled topology.
-
Visualization suits hundreds of nodes, not tens of thousands. Start from a query result rather than the whole graph.
-
Natural language search needs a language model. Without one configured the feature is unavailable, and the console now says so plainly instead of leaving you to work it out. You write CyQL directly, which the new editor makes considerably easier. Everything else works with no provider at all.
-
The console is English only for now. Its strings are externalized for translation, but no second language ships in this release.
-
The framework connectors do not create schema. They read and write a collection or graph that already exists; creating it is still your step.
0.9.0
The first release available outside CYROCK.AI. It ships as a single container image holding the whole engine, reachable from the web console, REST, the Java SDK and MCP.
What is in it
Retrieval
-
Vector search over HNSW indices, with cosine, dot-product and euclidean similarity.
-
Hybrid retrieval fusing BM25 full-text scoring with vector similarity.
-
Typed, indexed metadata filters applied before ranking, so a narrow filter still returns a full result set.
-
Multiple vector fields per record, searchable independently - the basis for multi-modal search.
-
Optional cross-encoder reranking through Cohere.
Knowledge graphs
-
Typed, weighted, directed graphs with multi-label nodes.
-
CyQL, a Cypher-inspired query language covering matching, traversal, vector similarity, full-text search, writes and schema changes.
-
Bounded variable-length traversal, one-hop neighbours and context-window construction.
-
Graph branching: fork a graph, change the copy, merge or discard.
-
Graph RAG - community detection with hierarchical summaries and global search.
Agentic memory
-
An observe/recall/decay lifecycle with entity linking, for agents that remember across sessions.
Interfaces
-
A web console for administration, querying and 2D/3D graph visualization.
-
REST APIs on two ports, each with an interactive Swagger UI.
-
A Java SDK over gRPC that needs no port argument against this image.
-
An MCP endpoint exposing 39 tools, with per-request authentication.
Platform
-
Multi-tenant, role-based access control applied uniformly across console, REST, SDK and MCP.
-
Built-in durable storage: every statement is one atomic commit, with multi-operation transactions and rollback.
-
An in-process ONNX embedder, so vector search works with no provider account and no network.
-
Prometheus metrics, OTLP tracing and unauthenticated health endpoints.
-
Multi-architecture image -
linux/amd64andlinux/arm64both native.
Known limitations
These are current boundaries rather than defects. Several are simply the next things we are building.
-
One container only. The multi-service topology - separate platform and data servers behind a single-port gateway, and horizontally scaled data servers - is not part of this release.
-
Evaluation authentication. The image seeds three in-memory logins at fixed roles. Identity provider integration over OIDC belongs to the deployment topology arriving in a future release.
-
Credentials come from the startup banner. There is no flow yet for rotating the seeded API key from the console; create an additional key in the administration area if you need a second one.
-
Vector dimensions are fixed at creation. Changing the embedding provider or model usually changes the dimension, which currently means creating a new field and re-embedding rather than migrating in place.
-
Single writer per collection. Consistency for HNSW indices depends on one writer at a time. This is invisible in a single container and becomes relevant only in the scaled topology.
-
Visualization suits hundreds of nodes, not tens of thousands. Start from a query result rather than the whole graph.
-
Natural language search needs a language model. Without one configured, the feature is unavailable and you write CyQL directly. Everything else works with no provider at all.
Support expectations
This is pre-release software, supported on a best-effort basis through issues and discussions in our GitHub repository. We read everything.
Most useful to us, roughly in order: something that does not work, something that was confusing, and something missing that blocks a real use case.
Storage may not carry over between releases. When a release changes the on-disk format, the engine now detects it at startup and refuses to serve the old storage with a clear message telling you to start with an empty storage directory, rather than failing later with an obscure error. Take a backup before upgrading, and see Operations.