For the complete documentation index, see llms.txt. Markdown versions of documentation pages are available by appending .md to page URLs.
Skip to content

PyGridGain 1.7.0 Release Notes

PyGridGain 1.7.0 supports Python 3.11 to 3.15, adds vector query filters and tuning parameters, and uses TLS 1.3 by default. Also covers langchain-gridgain and langgraph-checkpoint-gridgain 3.1.0.

New Features

PyGridGain 1.7.0 adds new vector search parameters, supports newer Python versions, and changes several defaults. This release also covers langchain-gridgain 3.1.0 and the new langgraph-checkpoint-gridgain 3.1.0 package.

Python 3.11 to 3.15

The client now requires Python 3.11 or later, and pre-built binaries are available for Python 3.11 to 3.15. Python 3.9 and 3.10 are no longer supported. To use the client with these Python versions, install version 1.6.0 or earlier.

Vector Query Filters

Cache.vector() and AioCache.vector() accept a field_filter parameter that restricts a query to entries with given field values. GridGain applies the filter while it searches the index, so the query returns k matching entries. The new oversample parameter controls re-scoring for indexes that store BINARY codes.

cursor = cache.vector(
    type_name=Article.type_name,
    field='vector_field',
    clause_vector=query_vector,
    k=10,
    threshold=0,
    field_filter={'tenant': 'acme'},
)

Vector Query and Index Parameters

Cache.vector() and AioCache.vector() accept the following new parameters:

  • ef_search sets the search beam width.

  • with_scores adds the similarity score to each result row.

  • no_content returns keys without values.

Vector indexes accept the new hnsw_m, hnsw_ef_construction, quantization (including INT8), max_segments, and query_threads keys.

The new vector parameters require GridGain 8.10 or later. If the cluster does not support a parameter that you set, the client raises NotSupportedByClusterError.

For more information, see Vector Queries.

LangChain and LangGraph Integration

langchain-gridgain 3.1.0 and the new langgraph-checkpoint-gridgain 3.1.0 package include the following changes, including the changes from langchain-gridgain 2.0.0 and 3.0.0:

  • New langgraph-checkpoint-gridgain package. It provides the following LangGraph components, which do not require a license for vector search:

    • GridGainCheckpointSaver, a checkpointer.

    • GridGainMemoryStore, a store for long-term memory with structured filters and expiry enforced by the server.

    • GridGainNodeCache, a node cache.

  • GridGainByteStore. A key-value store for binary values, such as embeddings cached with CacheBackedEmbeddings.

  • Entry expiry. GridGainStore, GridGainByteStore, GridGainCache, GridGainSemanticCache, and GridGainVectorStore accept a ttl argument.

  • Vector store metadata filtering. GridGainVectorStore filters searches by declared metadata keys on the server. Filtering requires GridGain 8.10 or later.

  • Vector store improvements. GridGainVectorStore supports IDs, deleting and getting documents by ID, and maximal marginal relevance search, and returns real cosine distances as scores. The new similarity_search_by_vector method searches by an embedding vector.

  • LangChain fixes. GridGainStore.yield_keys returns the keys. GridGainDocumentLoader applies filter_criteria and create_cache_if_not_exists. GridGainVectorStore searches work with score_threshold=None. add_documents stores each document of a batch in which only some documents have IDs. Components that share one client no longer corrupt the connection when used from several threads.

Changes in Behavior

  • TLS 1.3 by default. SSL connections now use TLS 1.3 or later by default. Before, the client used TLS 1.2. To connect to a cluster that supports only TLS 1.2, set ssl_version=ssl.TLSVersion.TLSv1_2. ssl_version now takes an ssl.TLSVersion value, which is the minimum version. The ssl.PROTOCOL_TLSv1_2, ssl.PROTOCOL_TLSv1_1, and ssl.PROTOCOL_TLSv1 constants still work; other ssl.PROTOCOL_* values, such as ssl.PROTOCOL_TLS_SERVER, raise ParameterError.

  • Vector query page size. Cache.vector() and AioCache.vector() use k as the default page_size instead of 1, so the whole result arrives in one page. To keep a smaller page, set page_size explicitly.

  • Vector query k check. The client raises ValueError for a k that is not positive, instead of sending the query to the server.

  • Vector query response shapes. An asyncio (key, value) row is now a tuple instead of a list. The low-level pygridgain.api.sql.vector() function returns rows as a list of (key, value) tuples instead of a dictionary.

  • Feature bits of vector index parameters. The client uses the same feature bits as GridGain 8.10 for Hierarchical Navigable Small World (HNSW) build parameters and vector quantization. It no longer uses these features with GridGain 8.9.38 or 8.9.39, which advertise the old bits but do not support the fields.

  • Dependencies. attrs and tzlocal are no longer pinned to exact versions, and contextvars is no longer required.

  • Packaging. The package metadata moved to pyproject.toml. The installed package no longer contains a top-level tests package. Building from a source distribution requires setuptools 77 or later. The package includes a py.typed marker, so type checkers use its type hints. The empty pygridgain.stream.aio_cluster module was removed; pygridgain.aio_cluster is not affected.

  • License. The client is licensed under the MariaDB Software License Terms.

  • LangChain and LangGraph requirements. langchain-gridgain 3.1.0 and langgraph-checkpoint-gridgain 3.1.0 require Python 3.11 or later and PyGridGain 1.7.0 or later (below 2.0). langgraph-checkpoint-gridgain 3.1.0 requires langchain-gridgain 3.1.0.

  • LangGraph imports. In langchain-gridgain 3.0.0, GridGainCheckpointSaver and GridGainMemoryStore moved to langgraph-checkpoint-gridgain. Import them from langgraph.checkpoint.gridgain and langgraph.store.gridgain.

  • LangChain dependencies and cache keys. langchain-gridgain depends on langchain-core 1.4.7 or later instead of langchain-community, and GridGainCache keys now include the large language model (LLM). Clear LLM caches written by version 1.0.x once after the upgrade.

Improvements and Fixed Issues

Issue ID
Description

GG-38200

A partition-aware client now registers the binary type of a complex key before it sends the key. Before, SQL and other clients could not read such keys. AioClient was affected by default.

GG-45279

langchain-gridgain: GridGainVectorStore.clear() and from_texts() now work correctly.

GG-49286

Added the ef_search, with_scores, and no_content vector query parameters.

GG-49287

Primitive arrays, such as vectors in query results, are decoded in one pass instead of element by element.

GG-49543

Added the hnsw_m and hnsw_ef_construction vector index parameters.

GG-49786

langchain-gridgain: added GridGainByteStore and TTL support; fixed the LLM cache key, similarity scores, and clearing of the semantic cache.

GG-49787, GG-49788, GG-49789

Added the LangGraph GridGainCheckpointSaver and GridGainMemoryStore.

GG-49932

Vector queries use k as the default page size and decode responses faster. Strings that contain a NUL character (U+0000) are no longer truncated.

GG-50683

Primitive arrays are encoded in one pass. A NumPy array whose dtype matches the wire format, such as '<f4' for floats, is sent without conversion.

GG-50943

Added the quantization (including INT8), max_segments, and query_threads vector index parameters.

GG-51543

The client now sends its time zone ID, such as Europe/Berlin, in the handshake. In versions 1.5.0 and 1.6.0, the time zone was not sent for zones that use daylight saving time, so the server shifted timestamps.

GG-51569

The LangGraph components moved to the new langgraph-checkpoint-gridgain package.

GG-51572

Added the LangGraph GridGainNodeCache.

GG-51636

Modernized the packaging and dependencies. The installation instructions now use pip install "pygridgain<9", because a plain pip install pygridgain installs the GridGain 9 client.

GG-51637

Added support for Python 3.11 to 3.15 and removed support for Python 3.9 and 3.10. SSL connections use TLS 1.3 by default. The client no longer emits DeprecationWarning on Python 3.14 and later.

GG-51945

Added the field_filter and oversample vector query parameters.

GG-51946

langchain-gridgain: GridGainVectorStore filters searches by metadata on the server.

GG-52701

Fixed reading and writing vector index parameters with GridGain 8.9.38 and 8.9.39 clusters.

GG-52848

The client is licensed under the MariaDB Software License Terms.

Installation and Upgrade Information

To upgrade an existing package, use the following command:

To install the latest version of a package:

To install a specific version:

To install the LangChain and LangGraph packages:

Previous Releases

We Value Your Feedback

Your comments and suggestions are always welcome. You can reach us here: docs@gridgain.com

Please visit the documentation for more information.

This page is: Copyright © 2026 MariaDB. All rights reserved.

Last updated

Was this helpful?