v1.18.1

Compare Source

Bug Fixes

• bump dependencies to latest (#​1283) (f29b639)

Documentation

• use lambda over getconn func (#​1251) (6ecf894)

v1.18.0

Compare Source

Features

• add domain name validation (#​1246) (15934bd)
• reset connection when the DNS record changes (#​1241) (1405f56)

v1.89.0

Compare Source

Features

• Add support for env_vars parameter when creating or updating Agent Engine. (691c1f6)
• Add UpdateRagEngineConfig rpc (64386e9)
• Implement preview get_rag_engine_config in rag_data.py (cda064e)
• Implement preview update_rag_engine_config in rag_data.py (cda064e)
• Update win rate calculation in GenAI Evaluation for rubric based evaluation (d442c42)

Bug Fixes

• Include the env variables of the AdkApp when cloning (9814768)

v1.88.0

Compare Source

Features

• Add a module-level function to create a Gemini template config for single-turn Gemini examples without having to explicitly construct the Gemini example. (126d10c)
• Add autoscaling_target_request_count_per_minute option in Preview model deployment on Endpoint & Model classes. (0f1f10a)
• Add Gen AI logging public preview API (1589f66)
• Add new template for AgentEngine (0478f10)
• Add the possibility to create multimodal datasets without explicitly specifying a bigquery dataset/table. (f5043a6)
• Allow EvalTask to take dataset as a Google sheet - simplifying rubric revision CUJ (15df1f6)
• Allow table targets in multi-region datasets when creating multimodal datasets (2d7bc32)
• Check if rubrics column is present before converting list of rubrics to string (6c1569b)
• Convert list of rubrics to a string before sending the API request. This allows users to use default parse_rubrics function that returns a list of rubrics with customized prompts, and bring their own rubrics as a list. (9f21b73)
• GenAI Evaluation: Release GenAI Evaluation SDK autorater metric configuration utils to vertexai.preview module. (f816d5a)
• GenAI Evaluation: Release GenAI Evaluation SDK parsing rubric generation response with additional fields to vertexai.preview module. (79ca86a)
• GenAI Evaluation: Release GenAI Evaluation SDK rubric based evaluation to vertexai.preview module. (bb07581)
• Model optimizer SDK support (f257298)
• Track the output path for metrics_table in experiments metadata, if output bucket is specified but no file name is specified, we will generate a unique file name (be2c99f)

Bug Fixes

• Preprend question tag to rubric critiquing response after editing the predefined prompts to end with <question> to avoid parsing errors (0abd6ad)

v1.87.0

Compare Source

Features

• Add page spans in retrieved contexts from Vertex RAG Engine in aiplatform v1 (30f0fcf)
• Add page spans in retrieved contexts from Vertex RAG Engine in aiplatform v1beta1 (30f0fcf)
• Add support for Vertex AI Search engine (30f0fcf)
• Add validation of the BigQuery location when creating a MultimodalDataset (98459aa)
• Adding Vertex AI Search Config for RAG corpuses to SDK (a7453da)
• Allow using multimodal datasets in the SDK. (d951b74)
• ExampleStore SDK in public preview (50fbdee)
• GenAI Evaluation: Release GenAI Evaluation SDK multimodal evaluation to vertexai.preview module. (f090ca1)
• Make it optional to pass in an instance of an agent when creating a new ReasoningEngine instance (f3addc9)
• Support creating multimodal datasets from a JSONL file containing Gemini requests (a323679)
• Vertex AI Model Garden open model export SDK Public Preview (898109d)

Bug Fixes

• Propagating import result sink correctly in the vertexai sdk. (184cca5)
• Remove the xprof dependency from Vertex Tensorboard uploader. (1336b85)

Documentation

• A comment for field rag_files_count in message .google.cloud.aiplatform.v1beta1.RagCorpus is changed (30f0fcf)

v1.86.0

Compare Source

Features

• Add a ExportPublisherModel API (#​5055) (4c69301)
• Add AssessData and AssembleData RPCs to DatasetService (4c69301)
• Add batch prediction assessments to multimodal dataset RPCs (a6d975e)
• Add env variables and agent framework to ReasoningEngineSpec (4c69301)
• Add env variables and agent framework to ReasoningEngineSpec in v1beta1 (4c69301)
• Add example, example_store, and example_store_service protos (a6d975e)
• Add import result bq sink to the import files API (4c69301)
• Add import result bq sink to the import files API (4c69301)
• Add import result gcs sink to the import files API (4c69301)
• Add import result gcs sink to the import files API (4c69301)
• Add model_config field for model selection preference (a6d975e)
• Add multihost_gpu_node_count to Vertex SDK. (03d9bf7)
• Add ray 2.33 and 2.42 unit tests as required checks (51dbe94)
• Add session.proto and session_service.proto (a6d975e)
• Add support for version 2.42 for RoV Bigquery read/write (6090205)
• Add the initial version of the LlamaIndex agent prebuilt template. (f1f2ea5)
• Add the initial version of the LlamaIndex agent prebuilt template. (8cbea81)
• Add VertexAISearch.engine option (4c69301)
• Change format string assembly logic for prompt templates. (43ebe97)
• Copy output uploading features from GA Eval SDK to preview. (d6f99cb)
• Enable force deletion in ReasoningEngine (a6d975e)
• Enable force deletion in ReasoningEngine v1beta1 (a6d975e)
• Modify v1 sdk to support rerankers (ececfb3)
• Update GenAI Evaluation QPS to 10. (0995ce2)
• Update GenAI Evaluation QPS to 10. (a59b0a6)
• Update multimodal evaluation (content_map_instance), rubric generation (rubric_based_instance, etc) and raw_output(raw_output, custom_output, etc) proto change in online eval API (a6d975e)
• Update Ray system tests to be compatible with new RoV 2.42 changes (fbb9560)
• Update setup.py for including Ray v2.42 (827b484)

Bug Fixes

• Format of the input to GetPublisherModelRequest for Hugging Face models. (8f7e8b9)
• Format of the input to GetPublisherModelRequest for Hugging Face models. (96d2ecb)
• GenAI Eval: fix a prompt template variable parsing issue when multiline json structures are present (5da362f)
• GenAI Eval: fix a prompt template variable parsing issue when multiline json structures are present (fd12cb8)
• GenAI Evaluation: Fix system instruction not being added to PairwiseMetric. (5c60d95)
• GenAI Evaluation: Fix system instruction not being added to PairwiseMetric. (c65879d)
• Update ipython import according to the latest ipython docs (533012a)

Documentation

• Add deployment_spec and agent_framework field to ReasoningEngineSpec. (4c69301)
• Add deployment_spec and agent_framework field to ReasoningEngineSpec. (4c69301)
• Update comment for package_spec from required to optional in ReasoningEngineSpec. (4c69301)
• Update comment for package_spec from required to optional in ReasoningEngineSpec. (4c69301)

v1.85.0

Compare Source

Features

• Add function_call.id and function_response.id (3f6c824)
• Add Python version 3.12 to Agent Engine (1232132)
• Add Ray 2.42 support to SDK Client Builder (2a67076)
• Add reranker config to RAG v1 API (3f6c824)
• Add support for version 2.42 for RoV Bigquery read/write (f4ce684)
• Implement GA versions of agent engine prebuilt templates (37b72c2)
• Modify v1 sdk to support rerankers (a6b7de5)

Bug Fixes

• Allow Protobuf 6.x (3f6c824)
• Preserve system_instruction in LangchainAgent clone() (63fc33e)
• Resolve issue where pre-release versions of dependencies are installed (de83fe3)

Documentation

• Update SDK Job Submission with Ray v2.42 updated comments (f281113)

v1.84.0

Compare Source

Features

• Add Layout Parser to RAG v1 API (ba9a314)
• Add multihost_gpu_node_count to Vertex SDK for multihost GPU support (ba9a314)
• Add request/response logging to PSC endpoints (b36a43a)
• Add support for Document AI Layout Parser in RAG v1 (1837390)
• Add the system tests for AG2 prebuilt template (80cfc2f)
• Add the system tests for Langgraph prebuilt template (833c1d2)
• Allowing users to choose whether to use the hf model cache (ba9a314)
• Allowing users to choose whether to use the hf model cache (ba9a314)
• Allowing users to specify the version id of the Model Garden model (ba9a314)
• Allowing users to specify the version id of the Model Garden model (ba9a314)
• GA Context Cache Python SDK (3aa0c6d)
• Support custom predictor Docker image builds on non-x86 architectures (#​2115) (87dd5c0)
• Vertex AI Model Garden deploy SDK Support for container specifications and equivalent Model Garden models for Hugging Face (e425ded)

v1.83.0

Compare Source

Features

• Add EnterpriseWebSearch tool option (816a607)
• Add notebook helper functions to eval SDK to display and visualize evaluation results in an IPython environment (acf8a80)
• Add VertexAISearch.engine option (816a607)
• Eval SDK: Modify Gemini scraping to log errors during response generation and return a consistent error message (cc2d36e)
• Vertex AI Model Garden open model deploy SDK Public Preview (4c8c277)

Bug Fixes

• Add request metadata to rag methods. (b2cc069)
• Migration to python 3.12. called_once was deleted in favor of assert_called_once, since it was largely misused (b4ec000)
• Resolve issue where pre-release versions of dependencies are installed (35cf246)

v1.82.0

Compare Source

Features

• Add Model Garden deploy API (e356f5a)
• Add the initial version of the AG2 agent prebuilt template. (abf08da)
• Implement preview deployment with RolloutOptions. (8cefabb)
• Vertex AI Agent Engine GA SDK (632730c)

Bug Fixes

• Avoid hard-coding the LinearRegression class name in tests (7121770)
• Clean up unused EmbeddingModelConfig. (4998c1a)
• Fix rag_embedding_model_config logic indentation (63736ef)
• PrivateEndpoint predict method should set parameters in request payload. (c7526a1)

v1.14.2: google-cloud-resource-manager: v1.14.2

Compare Source

Bug Fixes

• [Many APIs] Allow Protobuf 6.x (784a3ca)
• remove setup.cfg configuration for creating universal wheels (#​13659) (59bfd42)

v0.13.0

Compare Source

Features

• langgraph: Add Langgraph Checkpointer (#​284) (14a4240)

Bug Fixes

• deps: Update dependency numpy to v2 (#​251) (a164aa2)
• engine: Loop error on close (#​285) (e8bd4ae)

v0.3.31

Compare Source

Changes since 0.3.30

• 0.3.31
• Add delete_thread method to Checkpointer class (#​4328)
• langgraph: update min bound for prebuilt (#​4319)
• docs: add more prominent workflows link (#​4297)
• Same in get/update state/history
• Fix
• Stringify thread_id when of a different type

v0.3.30

Compare Source

Changes since 0.3.29

• langgraph: release 0.3.30 (#​4267)
• langgraph: support streaming messages from Command.update (#​4250)
• langgraph: handle pydantic updates consistently in Command (#​4255)

v0.3.29

Compare Source

Changes since 0.3.28

• 0.3.29
• Don't run branch reader in bg thread
• Reduce perf impact of set_context (#​4256)
• Use tuple entry for control branch (#​4248)
• Simplify path for control branch attached to every node (#​4247)
• Avoid creating contexts for control branches
• Avoid validating node input more than once per superstep (#​4242)
• Simplify path for control branch attached to every node
• Use cache in attach_branch

v0.3.28

Compare Source

Changes since 0.3.27

• langgraph: release 0.3.28 (#​4243)
• langgraph: allow passing a list of retry policies (#​4240)

v0.3.27

Compare Source

Changes since 0.3.26

• 0.3.27
• Add checkpoint_during arg (#​4169)
• Implement simpler filtering of config keys in RemoteGraph
• Make compatible with langchain-core 0.1 by conditionally importing _StreamingCallbackHandler (#​4203)
• Save Sends unconditionally
• Subgraphs inherit checkpoint mode
• Fix assignment of pending writes

v0.3.26

Compare Source

Changes since 0.3.25

• langgraph: release 0.3.26 (#​4204)
• langgraph: raise GraphInterrupt only if used as a subgraph
• Add benchmark script for typed dict version of existing wide state benchmark (#​4174)

v0.3.25

Compare Source

Changes since 0.3.24

• release(langgraph): 0.3.25
• Update naming
• Move to langgraph/graph
• Update jinja2 dev dep

v0.3.24

Compare Source

Changes since 0.3.23

• 0.3.24
• Disable tests in 3.9
• Fix
• Update
• Migrate quadratic channels too
• Lint
• Implement checkpoint migration
• WIP: Add migrate_checkpoint

v0.3.23

Compare Source

Changes since 0.3.22

• langgraph: release 0.3.23 (#​4141)
• Maintain checkpoint LATEST_VERSION constant in langgraph lib
• langgraph: support removing all messages with RemoveMessage (#​4117)

v0.3.22

Compare Source

Changes since 0.3.21

• 0.3.22
• Lint
• Avoid creating checkpoint unless we're saving it
• Avoid raise-catch strategy in BaseChannel.checkpoint()
• Lint
• Lint
• Lint
• Update
• Add fast path to serialize None values
• Lazily create atomic counters in pregel scratchpad
• Reduce perf impact of pregel scratchpad creation
• Reduce the number of channels created for each node by 50%
• In Python 3.12 or above, use asyncio eager task factory (#​4055)
• Fix
• Remove internal frames from stack traces (#​4054)
• In Python 3.12 or above, use asyncio eager task factory
• Remove internal frames from stack traces
• langgraph: use correct type for node destination annotations (#​4053)

v0.3.21

Compare Source

Changes since 0.3.20

• langgraph: release 0.3.21 (#​4050)
• langgraph: add tests for remote graph interrupts (#​4048)
• patch: fix return type of Topic.update (#​4029)
• benchmark: remove some benchmarks (#​4039)
• langgraph: fix interrupt deserialization in RemoteGraph (#​4040)
• x
• docs: add version admonitions for Interrupt and RetryPolicy (#​3988)
• Warn when get_graph tries to draw edge that doesn't exist

v0.3.20

Compare Source

Changes since 0.3.19

• 0.3.20
• Use 128bit hash for task ids
• Log and continue for subgraph schema issues (#​3997)
• Switch serialization lib to ormsgpack (#​3953)

v0.3.19

Compare Source

Changes since 0.3.18

• 0.3.19
• Switch task ids to use xxhash3 (#​3954)
• prebuilt: release 0.1.4 (#​3977)
• Use incremental storage in memory checkpointer (#​3960)
• docs: use hosted logo (#​3959)
• Include enum in check for pydantic state (#​3955)
• benchmark: Add compilation only (#​3932)
• Include enum in check for pydantic state
• Use fast path for prepare_next_tasks on input (#​3931)
• More small perf improvements
• Use fast path for prepare_next_tasks on input
• langgraph: incorporate information about previously updated channels to identify which tasks to execute next (#​3916)
• Add refcount test (#​3910)

v0.3.18

Compare Source

Changes since 0.3.17

• Release 0.3.18
• Update snapshots more
• Update snapshots
• Reference to PregelExecutableTask
• Unset config context after function end

v0.3.17

Compare Source

Changes since 0.3.16

• release(langgraph): 0.3.17 (#​3918)
• Store all triggers in task (#​3912)
• Add pydantic state benchmark case (#​3872)

v0.3.16

Compare Source

Changes since 0.3.15

• 0.3.16 (#​3908)
• Schema coercer should never throw (#​3871)
• libs: pregel: find_subgraph_pregel: Do not cache introspection (#​3894)
• Enable larger cases of sequential bench graph (#​3905)
• Revert "libs: pregel: find_subgraph_pregel: Do not search via introspection"
• libs: pregel: find_subgraph_pregel: Do not search via introspection

v0.3.15

Compare Source

Changes since 0.3.14

• Speed up prepare_single_task
• Improve prepare_single_task trigger checks to linear complexity (#​3891)
• Speed up task triggers check (#​3890)
• benchmarks: add sequential graph of a few hundred nodes (#​3882)

v0.3.14

Compare Source

Changes since 0.3.13

• When using global resume value, ensure subgraphs consume it
• Fix concurrency issue in PregelScratchpad.consume_null_resume

v0.3.13

Compare Source

Changes since 0.3.12

• 0.3.13
• Fix missing interrupts in stream (#​3886)
• Enable xray for remote graphs (#​3879)
• Enable xray for remote graphs

v0.3.12

Compare Source

Changes since 0.3.11

• 0.3.12
• Update state.py

v0.3.11

Compare Source

Changes since 0.3.10

• 0.3.11
• Add optional encryption of checkpointer payloads

v0.3.10

Changes since 0.3.8

• 0.3.10
• Update
• Fix types
• Update handling of updates/inputs passed in as pydantic models
• 0.3.9
• Coerce nested pydantic (#​3806)
• Move files
• weakref
• docs: more concise readme (#​3815)
• docs: add logo to readme (#​3816)
• Cache
• Rename env var
• Add tests
• Make default recursion_limit configurable by env var
• docs: update README (#​3799)

v0.3.8

Compare Source

Changes since 0.3.7

• langgraph: release 0.3.8 (#​3803)
• langgraph: handle pydantic state updates better for fields w/ defaults (#​3783)
• chore(langgraph): add functional api test for multiple task interrupts (#​3790)

v0.3.7

Compare Source

Changes since 0.3.6

• 0.3.7
• Avoid validating pydantic state models when we can (#​3782)
• Fix kafka
• Fix py 3.9
• Lint
• Avoid validating pydantic state models when we can
• langgraph: support subgraphs with a single node (#​3780)
• docs: Fix version badge by linking it to PyPi instead of shield (#​3766)

v0.3.6

Compare Source

Changes since 0.3.5

• langgraph: release 0.3.6 (#​3775)
• langgraph: use input schema from conditional edge (#​2516)
• Make pydntic input test stricter (#​3703)
• Don't enforce stream order
• Make pydntic input test stricter
• chore(langgraph): fix typing of task decorator (#​3670)
• chore(langgraph): fix typing of task decorator

v0.3.5

Compare Source

Changes since 0.3.4

• langgraph: release 0.3.5 (#​3690)
• langgraph: do not pass subgraph state on Command.parent updates (#​3686)

v0.3.4

Compare Source

Changes since 0.3.3

• 0.3.4
• When rehydrating a pydantic module, fallback to returning the kwargs dict (#​3685)
• Pregel.config_schema should use config_type directly when present (#​3641)

v0.3.3

Compare Source

Changes since 0.3.2

• 0.3.3
• Retry condit